blackteay/GPT-SoVITS
1
0
Fork 0
mirror of https://github.com/RVC-Boss/GPT-SoVITS.git synced 2025-08-25 15:04:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge 1e59f757a29e7e9eff73fb65b5740af098fe064e into ed89a023378dabba9d4b6580235bb9742245816d

Browse Source
This commit is contained in:
XXXXRT666 2025-06-11 15:19:08 +00:00 committed by GitHub
commit bd8022916e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
47 changed files with 1591 additions and 1254 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
.github/build_windows_packages.ps1 vendored
View File

@ -162,7 +162,7 @@ Copy-Item -Path $curr -Destination $pkgName -Recurse
$7zPath = "$pkgName.7z"
$start = Get-Date
Write-Host "Compress Starting at $start"
& "C:\Program Files\7-Zip\7z.exe" a -t7z "$7zPath" "$pkgName" -m0=lzma2 -mx=9 -md=1g -ms=1g -mmc=500 -mfb=273 -mlc=0 -mlp=4 -mpb=4 -mc=8g -mmt=on -bsp1
& "C:\Program Files\7-Zip\7z.exe" a -t7z "$7zPath" "$pkgName" -m0=lzma2 -mx=9 -mmt=on -bsp1
$end = Get-Date
Write-Host "Elapsed time: $($end - $start)"
Get-ChildItem .

3
GPT_SoVITS/AR/models/t2s_lightning_module.py
View File

@ -28,7 +28,8 @@ class Text2SemanticLightningModule(LightningModule):
self.load_state_dict(
torch.load(
pretrained_s1,
map_location="cpu", weights_only=False,
map_location="cpu",
weights_only=False,
)["weight"],
)
)

74
GPT_SoVITS/TTS_infer_pack/TTS.py
View File

@ -32,19 +32,21 @@ from transformers import AutoModelForMaskedLM, AutoTokenizer
from tools.audio_sr import AP_BWE
from tools.i18n.i18n import I18nAuto, scan_language_list
from tools.my_utils import load_audio
from TTS_infer_pack.text_segmentation_method import splits
from TTS_infer_pack.TextPreprocessor import TextPreprocessor
from sv import SV
resample_transform_dict={}
def resample(audio_tensor, sr0,sr1,device):
resample_transform_dict = {}
def resample(audio_tensor, sr0, sr1, device):
global resample_transform_dict
key="%s-%s-%s"%(sr0,sr1,str(device))
key = "%s-%s-%s" % (sr0, sr1, str(device))
if key not in resample_transform_dict:
resample_transform_dict[key] = torchaudio.transforms.Resample(
sr0, sr1
).to(device)
resample_transform_dict[key] = torchaudio.transforms.Resample(sr0, sr1).to(device)
return resample_transform_dict[key](audio_tensor)
language = os.environ.get("language", "Auto")
language = sys.argv[-1] if sys.argv[-1] in scan_language_list() else language
i18n = I18nAuto(language=language)
@ -111,6 +113,7 @@ def speed_change(input_audio: np.ndarray, speed: float, sr: int):
return processed_audio
class DictToAttrRecursive(dict):
def __init__(self, input_dict):
super().__init__(input_dict)
@ -479,7 +482,7 @@ class TTS:
def init_vits_weights(self, weights_path: str):
self.configs.vits_weights_path = weights_path
version, model_version, if_lora_v3 = get_sovits_version_from_path_fast(weights_path)
if "Pro"in model_version:
if "Pro" in model_version:
self.init_sv_model()
path_sovits = self.configs.default_configs[model_version]["vits_weights_path"]
@ -498,9 +501,9 @@ class TTS:
else:
hps["model"]["version"] = "v2"
version = hps["model"]["version"]
v3v4set={"v3", "v4"}
v3v4set = {"v3", "v4"}
if model_version not in v3v4set:
if "Pro"not in model_version:
if "Pro" not in model_version:
model_version = version
else:
hps["model"]["version"] = model_version
@ -542,7 +545,7 @@ class TTS:
if "pretrained" not in weights_path and hasattr(vits_model, "enc_q"):
del vits_model.enc_q
self.is_v2pro=model_version in {"v2Pro","v2ProPlus"}
self.is_v2pro = model_version in {"v2Pro", "v2ProPlus"}
if if_lora_v3 == False:
print(
@ -632,7 +635,9 @@ class TTS:
)
self.vocoder.remove_weight_norm()
state_dict_g = torch.load(
"%s/GPT_SoVITS/pretrained_models/gsv-v4-pretrained/vocoder.pth" % (now_dir,), map_location="cpu", weights_only=False
"%s/GPT_SoVITS/pretrained_models/gsv-v4-pretrained/vocoder.pth" % (now_dir,),
map_location="cpu",
weights_only=False,
)
print("loading vocoder", self.vocoder.load_state_dict(state_dict_g))
@ -752,11 +757,13 @@ class TTS:
if raw_sr != self.configs.sampling_rate:
audio = raw_audio.to(self.configs.device)
if (audio.shape[0] == 2): audio = audio.mean(0).unsqueeze(0)
if audio.shape[0] == 2:
audio = audio.mean(0).unsqueeze(0)
audio = resample(audio, raw_sr, self.configs.sampling_rate, self.configs.device)
else:
audio = raw_audio.to(self.configs.device)
if (audio.shape[0] == 2): audio = audio.mean(0).unsqueeze(0)
if audio.shape[0] == 2:
audio = audio.mean(0).unsqueeze(0)
maxx = audio.abs().max()
if maxx > 1:
@ -775,8 +782,9 @@ class TTS:
audio = resample(audio, self.configs.sampling_rate, 16000, self.configs.device)
if self.configs.is_half:
audio = audio.half()
else:audio=None
return spec,audio
else:
audio = None
return spec, audio
def _set_prompt_semantic(self, ref_wav_path: str):
zero_wav = np.zeros(
@ -1073,7 +1081,10 @@ class TTS:
###### setting reference audio and prompt text preprocessing ########
t0 = time.perf_counter()
if (ref_audio_path is not None) and (ref_audio_path != self.prompt_cache["ref_audio_path"] or (self.is_v2pro and self.prompt_cache["refer_spec"][0][1] is None)):
if (ref_audio_path is not None) and (
ref_audio_path != self.prompt_cache["ref_audio_path"]
or (self.is_v2pro and self.prompt_cache["refer_spec"][0][1] is None)
):
if not os.path.exists(ref_audio_path):
raise ValueError(f"{ref_audio_path} not exists")
self.set_ref_audio(ref_audio_path)
@ -1212,9 +1223,10 @@ class TTS:
t_34 += t4 - t3
refer_audio_spec = []
if self.is_v2pro:sv_emb=[]
for spec,audio_tensor in self.prompt_cache["refer_spec"]:
spec=spec.to(dtype=self.precision, device=self.configs.device)
if self.is_v2pro:
sv_emb = []
for spec, audio_tensor in self.prompt_cache["refer_spec"]:
spec = spec.to(dtype=self.precision, device=self.configs.device)
refer_audio_spec.append(spec)
if self.is_v2pro:
sv_emb.append(self.sv_model.compute_embedding3(audio_tensor))
@ -1249,10 +1261,14 @@ class TTS:
torch.cat(pred_semantic_list).unsqueeze(0).unsqueeze(0).to(self.configs.device)
)
_batch_phones = torch.cat(batch_phones).unsqueeze(0).to(self.configs.device)
if self.is_v2pro!=True:
_batch_audio_fragment = self.vits_model.decode(all_pred_semantic, _batch_phones, refer_audio_spec, speed=speed_factor).detach()[0, 0, :]
if self.is_v2pro != True:
_batch_audio_fragment = self.vits_model.decode(
all_pred_semantic, _batch_phones, refer_audio_spec, speed=speed_factor
).detach()[0, 0, :]
else:
_batch_audio_fragment = self.vits_model.decode(all_pred_semantic, _batch_phones, refer_audio_spec, speed=speed_factor,sv_emb=sv_emb).detach()[0, 0, :]
_batch_audio_fragment = self.vits_model.decode(
all_pred_semantic, _batch_phones, refer_audio_spec, speed=speed_factor, sv_emb=sv_emb
).detach()[0, 0, :]
audio_frag_end_idx.insert(0, 0)
batch_audio_fragment = [
_batch_audio_fragment[audio_frag_end_idx[i - 1] : audio_frag_end_idx[i]]
@ -1266,9 +1282,13 @@ class TTS:
pred_semantic_list[i][-idx:].unsqueeze(0).unsqueeze(0)
) # .unsqueeze(0)#mq要多unsqueeze一次
if self.is_v2pro != True:
audio_fragment = self.vits_model.decode(_pred_semantic, phones, refer_audio_spec, speed=speed_factor).detach()[0, 0, :]
audio_fragment = self.vits_model.decode(
_pred_semantic, phones, refer_audio_spec, speed=speed_factor
).detach()[0, 0, :]
else:
audio_fragment = self.vits_model.decode(_pred_semantic, phones, refer_audio_spec, speed=speed_factor,sv_emb=sv_emb).detach()[0, 0, :]
audio_fragment = self.vits_model.decode(
_pred_semantic, phones, refer_audio_spec, speed=speed_factor, sv_emb=sv_emb
).detach()[0, 0, :]
batch_audio_fragment.append(audio_fragment) ###试试重建不带上prompt部分
else:
if parallel_infer:
@ -1410,7 +1430,7 @@ class TTS:
raw_entry = self.prompt_cache["refer_spec"][0]
if isinstance(raw_entry, tuple):
raw_entry = raw_entry[0]
refer_audio_spec = raw_entry.to(dtype=self.precision,device=self.configs.device)
refer_audio_spec = raw_entry.to(dtype=self.precision, device=self.configs.device)
fea_ref, ge = self.vits_model.decode_encp(prompt_semantic_tokens, prompt_phones, refer_audio_spec)
ref_audio: torch.Tensor = self.prompt_cache["raw_audio"]
@ -1480,7 +1500,7 @@ class TTS:
raw_entry = self.prompt_cache["refer_spec"][0]
if isinstance(raw_entry, tuple):
raw_entry = raw_entry[0]
refer_audio_spec = raw_entry.to(dtype=self.precision,device=self.configs.device)
refer_audio_spec = raw_entry.to(dtype=self.precision, device=self.configs.device)
fea_ref, ge = self.vits_model.decode_encp(prompt_semantic_tokens, prompt_phones, refer_audio_spec)
ref_audio: torch.Tensor = self.prompt_cache["raw_audio"]

4
GPT_SoVITS/TTS_infer_pack/TextPreprocessor.py
View File

@ -160,7 +160,9 @@ class TextPreprocessor:
else:
for tmp in LangSegmenter.getTexts(text):
if langlist:
if (tmp["lang"] == "en" and langlist[-1] == "en") or (tmp["lang"] != "en" and langlist[-1] != "en"):
if (tmp["lang"] == "en" and langlist[-1] == "en") or (
tmp["lang"] != "en" and langlist[-1] != "en"
):
textlist[-1] += tmp["text"]
continue
if tmp["lang"] == "en":

152
GPT_SoVITS/eres2net/ERes2Net.py
View File

@ -2,13 +2,12 @@
# Licensed under the Apache License, Version 2.0 (http://www.apache.org/licenses/LICENSE-2.0)
"""
Res2Net implementation is adapted from https://github.com/wenet-e2e/wespeaker.
ERes2Net incorporates both local and global feature fusion techniques to improve the performance.
The local feature fusion (LFF) fuses the features within one single residual block to extract the local signal.
The global feature fusion (GFF) takes acoustic features of different scales as input to aggregate global signal.
Res2Net implementation is adapted from https://github.com/wenet-e2e/wespeaker.
ERes2Net incorporates both local and global feature fusion techniques to improve the performance.
The local feature fusion (LFF) fuses the features within one single residual block to extract the local signal.
The global feature fusion (GFF) takes acoustic features of different scales as input to aggregate global signal.
"""
import torch
import math
import torch.nn as nn
@ -16,15 +15,14 @@ import torch.nn.functional as F
import pooling_layers as pooling_layers
from fusion import AFF
class ReLU(nn.Hardtanh):
class ReLU(nn.Hardtanh):
def __init__(self, inplace=False):
super(ReLU, self).__init__(0, 20, inplace)
def __repr__(self):
inplace_str = 'inplace' if self.inplace else ''
return self.__class__.__name__ + ' (' \
+ inplace_str + ')'
inplace_str = "inplace" if self.inplace else ""
return self.__class__.__name__ + " (" + inplace_str + ")"
class BasicBlockERes2Net(nn.Module):
@ -32,28 +30,28 @@ class BasicBlockERes2Net(nn.Module):
def __init__(self, in_planes, planes, stride=1, baseWidth=32, scale=2):
super(BasicBlockERes2Net, self).__init__()
width = int(math.floor(planes*(baseWidth/64.0)))
self.conv1 = nn.Conv2d(in_planes, width*scale, kernel_size=1, stride=stride, bias=False)
self.bn1 = nn.BatchNorm2d(width*scale)
width = int(math.floor(planes * (baseWidth / 64.0)))
self.conv1 = nn.Conv2d(in_planes, width * scale, kernel_size=1, stride=stride, bias=False)
self.bn1 = nn.BatchNorm2d(width * scale)
self.nums = scale
convs=[]
bns=[]
convs = []
bns = []
for i in range(self.nums):
convs.append(nn.Conv2d(width, width, kernel_size=3, padding=1, bias=False))
bns.append(nn.BatchNorm2d(width))
convs.append(nn.Conv2d(width, width, kernel_size=3, padding=1, bias=False))
bns.append(nn.BatchNorm2d(width))
self.convs = nn.ModuleList(convs)
self.bns = nn.ModuleList(bns)
self.relu = ReLU(inplace=True)
self.conv3 = nn.Conv2d(width*scale, planes*self.expansion, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(planes*self.expansion)
self.conv3 = nn.Conv2d(width * scale, planes * self.expansion, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(planes * self.expansion)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion * planes:
self.shortcut = nn.Sequential(
nn.Conv2d(in_planes, self.expansion * planes, kernel_size=1,
stride=stride, bias=False),
nn.BatchNorm2d(self.expansion * planes))
nn.Conv2d(in_planes, self.expansion * planes, kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(self.expansion * planes),
)
self.stride = stride
self.width = width
self.scale = scale
@ -64,18 +62,18 @@ class BasicBlockERes2Net(nn.Module):
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
spx = torch.split(out,self.width,1)
spx = torch.split(out, self.width, 1)
for i in range(self.nums):
if i==0:
sp = spx[i]
else:
sp = sp + spx[i]
sp = self.convs[i](sp)
sp = self.relu(self.bns[i](sp))
if i==0:
out = sp
else:
out = torch.cat((out,sp),1)
if i == 0:
sp = spx[i]
else:
sp = sp + spx[i]
sp = self.convs[i](sp)
sp = self.relu(self.bns[i](sp))
if i == 0:
out = sp
else:
out = torch.cat((out, sp), 1)
out = self.conv3(out)
out = self.bn3(out)
@ -86,22 +84,23 @@ class BasicBlockERes2Net(nn.Module):
return out
class BasicBlockERes2Net_diff_AFF(nn.Module):
expansion = 2
def __init__(self, in_planes, planes, stride=1, baseWidth=32, scale=2):
super(BasicBlockERes2Net_diff_AFF, self).__init__()
width = int(math.floor(planes*(baseWidth/64.0)))
self.conv1 = nn.Conv2d(in_planes, width*scale, kernel_size=1, stride=stride, bias=False)
self.bn1 = nn.BatchNorm2d(width*scale)
width = int(math.floor(planes * (baseWidth / 64.0)))
self.conv1 = nn.Conv2d(in_planes, width * scale, kernel_size=1, stride=stride, bias=False)
self.bn1 = nn.BatchNorm2d(width * scale)
self.nums = scale
convs=[]
fuse_models=[]
bns=[]
convs = []
fuse_models = []
bns = []
for i in range(self.nums):
convs.append(nn.Conv2d(width, width, kernel_size=3, padding=1, bias=False))
bns.append(nn.BatchNorm2d(width))
convs.append(nn.Conv2d(width, width, kernel_size=3, padding=1, bias=False))
bns.append(nn.BatchNorm2d(width))
for j in range(self.nums - 1):
fuse_models.append(AFF(channels=width))
@ -110,14 +109,14 @@ class BasicBlockERes2Net_diff_AFF(nn.Module):
self.fuse_models = nn.ModuleList(fuse_models)
self.relu = ReLU(inplace=True)
self.conv3 = nn.Conv2d(width*scale, planes*self.expansion, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(planes*self.expansion)
self.conv3 = nn.Conv2d(width * scale, planes * self.expansion, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(planes * self.expansion)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion * planes:
self.shortcut = nn.Sequential(
nn.Conv2d(in_planes, self.expansion * planes, kernel_size=1,
stride=stride, bias=False),
nn.BatchNorm2d(self.expansion * planes))
nn.Conv2d(in_planes, self.expansion * planes, kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(self.expansion * planes),
)
self.stride = stride
self.width = width
self.scale = scale
@ -128,19 +127,19 @@ class BasicBlockERes2Net_diff_AFF(nn.Module):
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
spx = torch.split(out,self.width,1)
spx = torch.split(out, self.width, 1)
for i in range(self.nums):
if i==0:
if i == 0:
sp = spx[i]
else:
sp = self.fuse_models[i-1](sp, spx[i])
sp = self.fuse_models[i - 1](sp, spx[i])
sp = self.convs[i](sp)
sp = self.relu(self.bns[i](sp))
if i==0:
if i == 0:
out = sp
else:
out = torch.cat((out,sp),1)
out = torch.cat((out, sp), 1)
out = self.conv3(out)
out = self.bn3(out)
@ -151,16 +150,19 @@ class BasicBlockERes2Net_diff_AFF(nn.Module):
return out
class ERes2Net(nn.Module):
def __init__(self,
block=BasicBlockERes2Net,
block_fuse=BasicBlockERes2Net_diff_AFF,
num_blocks=[3, 4, 6, 3],
m_channels=32,
feat_dim=80,
embedding_size=192,
pooling_func='TSTP',
two_emb_layer=False):
def __init__(
self,
block=BasicBlockERes2Net,
block_fuse=BasicBlockERes2Net_diff_AFF,
num_blocks=[3, 4, 6, 3],
m_channels=32,
feat_dim=80,
embedding_size=192,
pooling_func="TSTP",
two_emb_layer=False,
):
super(ERes2Net, self).__init__()
self.in_planes = m_channels
self.feat_dim = feat_dim
@ -176,20 +178,24 @@ class ERes2Net(nn.Module):
self.layer4 = self._make_layer(block_fuse, m_channels * 8, num_blocks[3], stride=2)
# Downsampling module for each layer
self.layer1_downsample = nn.Conv2d(m_channels * 2, m_channels * 4, kernel_size=3, stride=2, padding=1, bias=False)
self.layer2_downsample = nn.Conv2d(m_channels * 4, m_channels * 8, kernel_size=3, padding=1, stride=2, bias=False)
self.layer3_downsample = nn.Conv2d(m_channels * 8, m_channels * 16, kernel_size=3, padding=1, stride=2, bias=False)
self.layer1_downsample = nn.Conv2d(
m_channels * 2, m_channels * 4, kernel_size=3, stride=2, padding=1, bias=False
)
self.layer2_downsample = nn.Conv2d(
m_channels * 4, m_channels * 8, kernel_size=3, padding=1, stride=2, bias=False
)
self.layer3_downsample = nn.Conv2d(
m_channels * 8, m_channels * 16, kernel_size=3, padding=1, stride=2, bias=False
)
# Bottom-up fusion module
self.fuse_mode12 = AFF(channels=m_channels * 4)
self.fuse_mode123 = AFF(channels=m_channels * 8)
self.fuse_mode1234 = AFF(channels=m_channels * 16)
self.n_stats = 1 if pooling_func == 'TAP' or pooling_func == "TSDP" else 2
self.pool = getattr(pooling_layers, pooling_func)(
in_dim=self.stats_dim * block.expansion)
self.seg_1 = nn.Linear(self.stats_dim * block.expansion * self.n_stats,
embedding_size)
self.n_stats = 1 if pooling_func == "TAP" or pooling_func == "TSDP" else 2
self.pool = getattr(pooling_layers, pooling_func)(in_dim=self.stats_dim * block.expansion)
self.seg_1 = nn.Linear(self.stats_dim * block.expansion * self.n_stats, embedding_size)
if self.two_emb_layer:
self.seg_bn_1 = nn.BatchNorm1d(embedding_size, affine=False)
self.seg_2 = nn.Linear(embedding_size, embedding_size)
@ -243,18 +249,16 @@ class ERes2Net(nn.Module):
fuse_out123 = self.fuse_mode123(out3, fuse_out12_downsample)
out4 = self.layer4(out3)
fuse_out123_downsample = self.layer3_downsample(fuse_out123)
fuse_out1234 = self.fuse_mode1234(out4, fuse_out123_downsample).flatten(start_dim=1,end_dim=2).mean(-1)
fuse_out1234 = self.fuse_mode1234(out4, fuse_out123_downsample).flatten(start_dim=1, end_dim=2).mean(-1)
return fuse_out1234
if __name__ == '__main__':
if __name__ == "__main__":
x = torch.zeros(10, 300, 80)
model = ERes2Net(feat_dim=80, embedding_size=192, pooling_func='TSTP')
model = ERes2Net(feat_dim=80, embedding_size=192, pooling_func="TSTP")
model.eval()
out = model(x)
print(out.shape) # torch.Size([10, 192])
print(out.shape) # torch.Size([10, 192])
num_params = sum(param.numel() for param in model.parameters())
print("{} M".format(num_params / 1e6)) # 6.61M
print("{} M".format(num_params / 1e6)) # 6.61M

200
GPT_SoVITS/eres2net/ERes2NetV2.py
View File

@ -2,14 +2,12 @@
# Licensed under the Apache License, Version 2.0 (http://www.apache.org/licenses/LICENSE-2.0)
"""
To further improve the short-duration feature extraction capability of ERes2Net, we expand the channel dimension
within each stage. However, this modification also increases the number of model parameters and computational complexity.
To alleviate this problem, we propose an improved ERes2NetV2 by pruning redundant structures, ultimately reducing
both the model parameters and its computational cost.
To further improve the short-duration feature extraction capability of ERes2Net, we expand the channel dimension
within each stage. However, this modification also increases the number of model parameters and computational complexity.
To alleviate this problem, we propose an improved ERes2NetV2 by pruning redundant structures, ultimately reducing
both the model parameters and its computational cost.
"""
import torch
import math
import torch.nn as nn
@ -17,47 +15,42 @@ import torch.nn.functional as F
import pooling_layers as pooling_layers
from fusion import AFF
class ReLU(nn.Hardtanh):
class ReLU(nn.Hardtanh):
def __init__(self, inplace=False):
super(ReLU, self).__init__(0, 20, inplace)
def __repr__(self):
inplace_str = 'inplace' if self.inplace else ''
return self.__class__.__name__ + ' (' \
+ inplace_str + ')'
inplace_str = "inplace" if self.inplace else ""
return self.__class__.__name__ + " (" + inplace_str + ")"
class BasicBlockERes2NetV2(nn.Module):
def __init__(self, in_planes, planes, stride=1, baseWidth=26, scale=2, expansion=2):
super(BasicBlockERes2NetV2, self).__init__()
width = int(math.floor(planes*(baseWidth/64.0)))
self.conv1 = nn.Conv2d(in_planes, width*scale, kernel_size=1, stride=stride, bias=False)
self.bn1 = nn.BatchNorm2d(width*scale)
width = int(math.floor(planes * (baseWidth / 64.0)))
self.conv1 = nn.Conv2d(in_planes, width * scale, kernel_size=1, stride=stride, bias=False)
self.bn1 = nn.BatchNorm2d(width * scale)
self.nums = scale
self.expansion = expansion
convs=[]
bns=[]
convs = []
bns = []
for i in range(self.nums):
convs.append(nn.Conv2d(width, width, kernel_size=3, padding=1, bias=False))
bns.append(nn.BatchNorm2d(width))
convs.append(nn.Conv2d(width, width, kernel_size=3, padding=1, bias=False))
bns.append(nn.BatchNorm2d(width))
self.convs = nn.ModuleList(convs)
self.bns = nn.ModuleList(bns)
self.relu = ReLU(inplace=True)
self.conv3 = nn.Conv2d(width*scale, planes*self.expansion, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(planes*self.expansion)
self.conv3 = nn.Conv2d(width * scale, planes * self.expansion, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(planes * self.expansion)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion * planes:
self.shortcut = nn.Sequential(
nn.Conv2d(in_planes,
self.expansion * planes,
kernel_size=1,
stride=stride,
bias=False),
nn.BatchNorm2d(self.expansion * planes))
nn.Conv2d(in_planes, self.expansion * planes, kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(self.expansion * planes),
)
self.stride = stride
self.width = width
self.scale = scale
@ -68,18 +61,18 @@ class BasicBlockERes2NetV2(nn.Module):
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
spx = torch.split(out,self.width,1)
spx = torch.split(out, self.width, 1)
for i in range(self.nums):
if i==0:
sp = spx[i]
else:
sp = sp + spx[i]
sp = self.convs[i](sp)
sp = self.relu(self.bns[i](sp))
if i==0:
out = sp
else:
out = torch.cat((out,sp),1)
if i == 0:
sp = spx[i]
else:
sp = sp + spx[i]
sp = self.convs[i](sp)
sp = self.relu(self.bns[i](sp))
if i == 0:
out = sp
else:
out = torch.cat((out, sp), 1)
out = self.conv3(out)
out = self.bn3(out)
@ -90,22 +83,22 @@ class BasicBlockERes2NetV2(nn.Module):
return out
class BasicBlockERes2NetV2AFF(nn.Module):
class BasicBlockERes2NetV2AFF(nn.Module):
def __init__(self, in_planes, planes, stride=1, baseWidth=26, scale=2, expansion=2):
super(BasicBlockERes2NetV2AFF, self).__init__()
width = int(math.floor(planes*(baseWidth/64.0)))
self.conv1 = nn.Conv2d(in_planes, width*scale, kernel_size=1, stride=stride, bias=False)
self.bn1 = nn.BatchNorm2d(width*scale)
width = int(math.floor(planes * (baseWidth / 64.0)))
self.conv1 = nn.Conv2d(in_planes, width * scale, kernel_size=1, stride=stride, bias=False)
self.bn1 = nn.BatchNorm2d(width * scale)
self.nums = scale
self.expansion = expansion
convs=[]
fuse_models=[]
bns=[]
convs = []
fuse_models = []
bns = []
for i in range(self.nums):
convs.append(nn.Conv2d(width, width, kernel_size=3, padding=1, bias=False))
bns.append(nn.BatchNorm2d(width))
convs.append(nn.Conv2d(width, width, kernel_size=3, padding=1, bias=False))
bns.append(nn.BatchNorm2d(width))
for j in range(self.nums - 1):
fuse_models.append(AFF(channels=width, r=4))
@ -114,17 +107,14 @@ class BasicBlockERes2NetV2AFF(nn.Module):
self.fuse_models = nn.ModuleList(fuse_models)
self.relu = ReLU(inplace=True)
self.conv3 = nn.Conv2d(width*scale, planes*self.expansion, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(planes*self.expansion)
self.conv3 = nn.Conv2d(width * scale, planes * self.expansion, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(planes * self.expansion)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion * planes:
self.shortcut = nn.Sequential(
nn.Conv2d(in_planes,
self.expansion * planes,
kernel_size=1,
stride=stride,
bias=False),
nn.BatchNorm2d(self.expansion * planes))
nn.Conv2d(in_planes, self.expansion * planes, kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(self.expansion * planes),
)
self.stride = stride
self.width = width
self.scale = scale
@ -135,19 +125,19 @@ class BasicBlockERes2NetV2AFF(nn.Module):
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
spx = torch.split(out,self.width,1)
spx = torch.split(out, self.width, 1)
for i in range(self.nums):
if i==0:
if i == 0:
sp = spx[i]
else:
sp = self.fuse_models[i-1](sp, spx[i])
sp = self.fuse_models[i - 1](sp, spx[i])
sp = self.convs[i](sp)
sp = self.relu(self.bns[i](sp))
if i==0:
if i == 0:
out = sp
else:
out = torch.cat((out,sp),1)
out = torch.cat((out, sp), 1)
out = self.conv3(out)
out = self.bn3(out)
@ -158,19 +148,22 @@ class BasicBlockERes2NetV2AFF(nn.Module):
return out
class ERes2NetV2(nn.Module):
def __init__(self,
block=BasicBlockERes2NetV2,
block_fuse=BasicBlockERes2NetV2AFF,
num_blocks=[3, 4, 6, 3],
m_channels=64,
feat_dim=80,
embedding_size=192,
baseWidth=26,
scale=2,
expansion=2,
pooling_func='TSTP',
two_emb_layer=False):
def __init__(
self,
block=BasicBlockERes2NetV2,
block_fuse=BasicBlockERes2NetV2AFF,
num_blocks=[3, 4, 6, 3],
m_channels=64,
feat_dim=80,
embedding_size=192,
baseWidth=26,
scale=2,
expansion=2,
pooling_func="TSTP",
two_emb_layer=False,
):
super(ERes2NetV2, self).__init__()
self.in_planes = m_channels
self.feat_dim = feat_dim
@ -181,42 +174,29 @@ class ERes2NetV2(nn.Module):
self.scale = scale
self.expansion = expansion
self.conv1 = nn.Conv2d(1,
m_channels,
kernel_size=3,
stride=1,
padding=1,
bias=False)
self.conv1 = nn.Conv2d(1, m_channels, kernel_size=3, stride=1, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(m_channels)
self.layer1 = self._make_layer(block,
m_channels,
num_blocks[0],
stride=1)
self.layer2 = self._make_layer(block,
m_channels * 2,
num_blocks[1],
stride=2)
self.layer3 = self._make_layer(block_fuse,
m_channels * 4,
num_blocks[2],
stride=2)
self.layer4 = self._make_layer(block_fuse,
m_channels * 8,
num_blocks[3],
stride=2)
self.layer1 = self._make_layer(block, m_channels, num_blocks[0], stride=1)
self.layer2 = self._make_layer(block, m_channels * 2, num_blocks[1], stride=2)
self.layer3 = self._make_layer(block_fuse, m_channels * 4, num_blocks[2], stride=2)
self.layer4 = self._make_layer(block_fuse, m_channels * 8, num_blocks[3], stride=2)
# Downsampling module
self.layer3_ds = nn.Conv2d(m_channels * 4 * self.expansion, m_channels * 8 * self.expansion, kernel_size=3, \
padding=1, stride=2, bias=False)
self.layer3_ds = nn.Conv2d(
m_channels * 4 * self.expansion,
m_channels * 8 * self.expansion,
kernel_size=3,
padding=1,
stride=2,
bias=False,
)
# Bottom-up fusion module
self.fuse34 = AFF(channels=m_channels * 8 * self.expansion, r=4)
self.n_stats = 1 if pooling_func == 'TAP' or pooling_func == "TSDP" else 2
self.pool = getattr(pooling_layers, pooling_func)(
in_dim=self.stats_dim * self.expansion)
self.seg_1 = nn.Linear(self.stats_dim * self.expansion * self.n_stats,
embedding_size)
self.n_stats = 1 if pooling_func == "TAP" or pooling_func == "TSDP" else 2
self.pool = getattr(pooling_layers, pooling_func)(in_dim=self.stats_dim * self.expansion)
self.seg_1 = nn.Linear(self.stats_dim * self.expansion * self.n_stats, embedding_size)
if self.two_emb_layer:
self.seg_bn_1 = nn.BatchNorm1d(embedding_size, affine=False)
self.seg_2 = nn.Linear(embedding_size, embedding_size)
@ -228,7 +208,11 @@ class ERes2NetV2(nn.Module):
strides = [stride] + [1] * (num_blocks - 1)
layers = []
for stride in strides:
layers.append(block(self.in_planes, planes, stride, baseWidth=self.baseWidth, scale=self.scale, expansion=self.expansion))
layers.append(
block(
self.in_planes, planes, stride, baseWidth=self.baseWidth, scale=self.scale, expansion=self.expansion
)
)
self.in_planes = planes * self.expansion
return nn.Sequential(*layers)
@ -264,7 +248,7 @@ class ERes2NetV2(nn.Module):
out3_ds = self.layer3_ds(out3)
fuse_out34 = self.fuse34(out4, out3_ds)
# print(111111111,fuse_out34.shape)#111111111 torch.Size([16, 2048, 10, 72])
return fuse_out34.flatten(start_dim=1,end_dim=2).mean(-1)
return fuse_out34.flatten(start_dim=1, end_dim=2).mean(-1)
# stats = self.pool(fuse_out34)
#
# embed_a = self.seg_1(stats)
@ -276,17 +260,13 @@ class ERes2NetV2(nn.Module):
# else:
# return embed_a
if __name__ == '__main__':
if __name__ == "__main__":
x = torch.randn(1, 300, 80)
model = ERes2NetV2(feat_dim=80, embedding_size=192, m_channels=64, baseWidth=26, scale=2, expansion=2)
model.eval()
y = model(x)
print(y.size())
macs, num_params = profile(model, inputs=(x, ))
print("Params: {} M".format(num_params / 1e6)) # 17.86 M
print("MACs: {} G".format(macs / 1e9)) # 12.69 G
macs, num_params = profile(model, inputs=(x,))
print("Params: {} M".format(num_params / 1e6)) # 17.86 M
print("MACs: {} G".format(macs / 1e9)) # 12.69 G

157
GPT_SoVITS/eres2net/ERes2Net_huge.py
View File

@ -1,14 +1,13 @@
# Copyright 3D-Speaker (https://github.com/alibaba-damo-academy/3D-Speaker). All Rights Reserved.
# Licensed under the Apache License, Version 2.0 (http://www.apache.org/licenses/LICENSE-2.0)
""" Res2Net implementation is adapted from https://github.com/wenet-e2e/wespeaker.
ERes2Net incorporates both local and global feature fusion techniques to improve the performance.
The local feature fusion (LFF) fuses the features within one single residual block to extract the local signal.
The global feature fusion (GFF) takes acoustic features of different scales as input to aggregate global signal.
ERes2Net-huge is an upgraded version of ERes2Net that uses a larger number of parameters to achieve better
recognition performance. Parameters expansion, baseWidth, and scale can be modified to obtain optimal performance.
"""Res2Net implementation is adapted from https://github.com/wenet-e2e/wespeaker.
ERes2Net incorporates both local and global feature fusion techniques to improve the performance.
The local feature fusion (LFF) fuses the features within one single residual block to extract the local signal.
The global feature fusion (GFF) takes acoustic features of different scales as input to aggregate global signal.
ERes2Net-huge is an upgraded version of ERes2Net that uses a larger number of parameters to achieve better
recognition performance. Parameters expansion, baseWidth, and scale can be modified to obtain optimal performance.
"""
import pdb
import torch
import math
@ -17,15 +16,14 @@ import torch.nn.functional as F
import pooling_layers as pooling_layers
from fusion import AFF
class ReLU(nn.Hardtanh):
class ReLU(nn.Hardtanh):
def __init__(self, inplace=False):
super(ReLU, self).__init__(0, 20, inplace)
def __repr__(self):
inplace_str = 'inplace' if self.inplace else ''
return self.__class__.__name__ + ' (' \
+ inplace_str + ')'
inplace_str = "inplace" if self.inplace else ""
return self.__class__.__name__ + " (" + inplace_str + ")"
class BasicBlockERes2Net(nn.Module):
@ -33,27 +31,28 @@ class BasicBlockERes2Net(nn.Module):
def __init__(self, in_planes, planes, stride=1, baseWidth=24, scale=3):
super(BasicBlockERes2Net, self).__init__()
width = int(math.floor(planes*(baseWidth/64.0)))
self.conv1 = nn.Conv2d(in_planes, width*scale, kernel_size=1, stride=stride, bias=False)
self.bn1 = nn.BatchNorm2d(width*scale)
width = int(math.floor(planes * (baseWidth / 64.0)))
self.conv1 = nn.Conv2d(in_planes, width * scale, kernel_size=1, stride=stride, bias=False)
self.bn1 = nn.BatchNorm2d(width * scale)
self.nums = scale
convs=[]
bns=[]
convs = []
bns = []
for i in range(self.nums):
convs.append(nn.Conv2d(width, width, kernel_size=3, padding=1, bias=False))
bns.append(nn.BatchNorm2d(width))
convs.append(nn.Conv2d(width, width, kernel_size=3, padding=1, bias=False))
bns.append(nn.BatchNorm2d(width))
self.convs = nn.ModuleList(convs)
self.bns = nn.ModuleList(bns)
self.relu = ReLU(inplace=True)
self.conv3 = nn.Conv2d(width*scale, planes*self.expansion, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(planes*self.expansion)
self.conv3 = nn.Conv2d(width * scale, planes * self.expansion, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(planes * self.expansion)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion * planes:
self.shortcut = nn.Sequential(
nn.Conv2d(in_planes, self.expansion * planes, kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(self.expansion * planes))
nn.BatchNorm2d(self.expansion * planes),
)
self.stride = stride
self.width = width
self.scale = scale
@ -64,18 +63,18 @@ class BasicBlockERes2Net(nn.Module):
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
spx = torch.split(out,self.width,1)
spx = torch.split(out, self.width, 1)
for i in range(self.nums):
if i==0:
sp = spx[i]
else:
sp = sp + spx[i]
sp = self.convs[i](sp)
sp = self.relu(self.bns[i](sp))
if i==0:
out = sp
else:
out = torch.cat((out,sp),1)
if i == 0:
sp = spx[i]
else:
sp = sp + spx[i]
sp = self.convs[i](sp)
sp = self.relu(self.bns[i](sp))
if i == 0:
out = sp
else:
out = torch.cat((out, sp), 1)
out = self.conv3(out)
out = self.bn3(out)
@ -86,22 +85,23 @@ class BasicBlockERes2Net(nn.Module):
return out
class BasicBlockERes2Net_diff_AFF(nn.Module):
expansion = 4
def __init__(self, in_planes, planes, stride=1, baseWidth=24, scale=3):
super(BasicBlockERes2Net_diff_AFF, self).__init__()
width = int(math.floor(planes*(baseWidth/64.0)))
self.conv1 = nn.Conv2d(in_planes, width*scale, kernel_size=1, stride=stride, bias=False)
self.bn1 = nn.BatchNorm2d(width*scale)
width = int(math.floor(planes * (baseWidth / 64.0)))
self.conv1 = nn.Conv2d(in_planes, width * scale, kernel_size=1, stride=stride, bias=False)
self.bn1 = nn.BatchNorm2d(width * scale)
self.nums = scale
convs=[]
fuse_models=[]
bns=[]
convs = []
fuse_models = []
bns = []
for i in range(self.nums):
convs.append(nn.Conv2d(width, width, kernel_size=3, padding=1, bias=False))
bns.append(nn.BatchNorm2d(width))
convs.append(nn.Conv2d(width, width, kernel_size=3, padding=1, bias=False))
bns.append(nn.BatchNorm2d(width))
for j in range(self.nums - 1):
fuse_models.append(AFF(channels=width))
@ -110,13 +110,14 @@ class BasicBlockERes2Net_diff_AFF(nn.Module):
self.fuse_models = nn.ModuleList(fuse_models)
self.relu = ReLU(inplace=True)
self.conv3 = nn.Conv2d(width*scale, planes*self.expansion, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(planes*self.expansion)
self.conv3 = nn.Conv2d(width * scale, planes * self.expansion, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(planes * self.expansion)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion * planes:
self.shortcut = nn.Sequential(
nn.Conv2d(in_planes, self.expansion * planes, kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(self.expansion * planes))
nn.BatchNorm2d(self.expansion * planes),
)
self.stride = stride
self.width = width
self.scale = scale
@ -127,20 +128,19 @@ class BasicBlockERes2Net_diff_AFF(nn.Module):
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
spx = torch.split(out,self.width,1)
spx = torch.split(out, self.width, 1)
for i in range(self.nums):
if i==0:
if i == 0:
sp = spx[i]
else:
sp = self.fuse_models[i-1](sp, spx[i])
sp = self.fuse_models[i - 1](sp, spx[i])
sp = self.convs[i](sp)
sp = self.relu(self.bns[i](sp))
if i==0:
if i == 0:
out = sp
else:
out = torch.cat((out,sp),1)
out = torch.cat((out, sp), 1)
out = self.conv3(out)
out = self.bn3(out)
@ -151,16 +151,19 @@ class BasicBlockERes2Net_diff_AFF(nn.Module):
return out
class ERes2Net(nn.Module):
def __init__(self,
block=BasicBlockERes2Net,
block_fuse=BasicBlockERes2Net_diff_AFF,
num_blocks=[3, 4, 6, 3],
m_channels=64,
feat_dim=80,
embedding_size=192,
pooling_func='TSTP',
two_emb_layer=False):
def __init__(
self,
block=BasicBlockERes2Net,
block_fuse=BasicBlockERes2Net_diff_AFF,
num_blocks=[3, 4, 6, 3],
m_channels=64,
feat_dim=80,
embedding_size=192,
pooling_func="TSTP",
two_emb_layer=False,
):
super(ERes2Net, self).__init__()
self.in_planes = m_channels
self.feat_dim = feat_dim
@ -176,17 +179,22 @@ class ERes2Net(nn.Module):
self.layer3 = self._make_layer(block_fuse, m_channels * 4, num_blocks[2], stride=2)
self.layer4 = self._make_layer(block_fuse, m_channels * 8, num_blocks[3], stride=2)
self.layer1_downsample = nn.Conv2d(m_channels * 4, m_channels * 8, kernel_size=3, padding=1, stride=2, bias=False)
self.layer2_downsample = nn.Conv2d(m_channels * 8, m_channels * 16, kernel_size=3, padding=1, stride=2, bias=False)
self.layer3_downsample = nn.Conv2d(m_channels * 16, m_channels * 32, kernel_size=3, padding=1, stride=2, bias=False)
self.layer1_downsample = nn.Conv2d(
m_channels * 4, m_channels * 8, kernel_size=3, padding=1, stride=2, bias=False
)
self.layer2_downsample = nn.Conv2d(
m_channels * 8, m_channels * 16, kernel_size=3, padding=1, stride=2, bias=False
)
self.layer3_downsample = nn.Conv2d(
m_channels * 16, m_channels * 32, kernel_size=3, padding=1, stride=2, bias=False
)
self.fuse_mode12 = AFF(channels=m_channels * 8)
self.fuse_mode123 = AFF(channels=m_channels * 16)
self.fuse_mode1234 = AFF(channels=m_channels * 32)
self.n_stats = 1 if pooling_func == 'TAP' or pooling_func == "TSDP" else 2
self.pool = getattr(pooling_layers, pooling_func)(
in_dim=self.stats_dim * block.expansion)
self.n_stats = 1 if pooling_func == "TAP" or pooling_func == "TSDP" else 2
self.pool = getattr(pooling_layers, pooling_func)(in_dim=self.stats_dim * block.expansion)
self.seg_1 = nn.Linear(self.stats_dim * block.expansion * self.n_stats, embedding_size)
if self.two_emb_layer:
self.seg_bn_1 = nn.BatchNorm1d(embedding_size, affine=False)
@ -229,7 +237,7 @@ class ERes2Net(nn.Module):
else:
return embed_a
def forward2(self, x,if_mean):
def forward2(self, x, if_mean):
x = x.permute(0, 2, 1) # (B,T,F) => (B,F,T)
x = x.unsqueeze_(1)
@ -243,14 +251,13 @@ class ERes2Net(nn.Module):
fuse_out123 = self.fuse_mode123(out3, fuse_out12_downsample)
out4 = self.layer4(out3)
fuse_out123_downsample = self.layer3_downsample(fuse_out123)
fuse_out1234 = self.fuse_mode1234(out4, fuse_out123_downsample).flatten(start_dim=1,end_dim=2)#bs,20480,T
if(if_mean==False):
mean=fuse_out1234[0].transpose(1,0)#(T,20480),bs=T
fuse_out1234 = self.fuse_mode1234(out4, fuse_out123_downsample).flatten(start_dim=1, end_dim=2) # bs,20480,T
if if_mean == False:
mean = fuse_out1234[0].transpose(1, 0) # (T,20480),bs=T
else:
mean = fuse_out1234.mean(2)#bs,20480
mean_std=torch.cat([mean,torch.zeros_like(mean)],1)
return self.seg_1(mean_std)#(T,192)
mean = fuse_out1234.mean(2) # bs,20480
mean_std = torch.cat([mean, torch.zeros_like(mean)], 1)
return self.seg_1(mean_std) # (T,192)
# stats = self.pool(fuse_out1234)
# if self.two_emb_layer:
@ -275,12 +282,8 @@ class ERes2Net(nn.Module):
fuse_out123 = self.fuse_mode123(out3, fuse_out12_downsample)
out4 = self.layer4(out3)
fuse_out123_downsample = self.layer3_downsample(fuse_out123)
fuse_out1234 = self.fuse_mode1234(out4, fuse_out123_downsample).flatten(start_dim=1,end_dim=2).mean(-1)
fuse_out1234 = self.fuse_mode1234(out4, fuse_out123_downsample).flatten(start_dim=1, end_dim=2).mean(-1)
return fuse_out1234
# print(fuse_out1234.shape)
# print(fuse_out1234.flatten(start_dim=1,end_dim=2).shape)
# pdb.set_trace()

4
GPT_SoVITS/eres2net/fusion.py
View File

@ -6,7 +6,6 @@ import torch.nn as nn
class AFF(nn.Module):
def __init__(self, channels=64, r=4):
super(AFF, self).__init__()
inter_channels = int(channels // r)
@ -23,7 +22,6 @@ class AFF(nn.Module):
xa = torch.cat((x, ds_y), dim=1)
x_att = self.local_att(xa)
x_att = 1.0 + torch.tanh(x_att)
xo = torch.mul(x, x_att) + torch.mul(ds_y, 2.0-x_att)
xo = torch.mul(x, x_att) + torch.mul(ds_y, 2.0 - x_att)
return xo

49
GPT_SoVITS/eres2net/kaldi.py
View File

@ -144,7 +144,7 @@ def _get_waveform_and_window_properties(
)
assert 0 < window_shift, "`window_shift` must be greater than 0"
assert padded_window_size % 2 == 0, (
"the padded `window_size` must be divisible by two." " use `round_to_power_of_two` or change `frame_length`"
"the padded `window_size` must be divisible by two. use `round_to_power_of_two` or change `frame_length`"
)
assert 0.0 <= preemphasis_coefficient <= 1.0, "`preemphasis_coefficient` must be between [0,1]"
assert sample_frequency > 0, "`sample_frequency` must be greater than zero"
@ -441,7 +441,9 @@ def get_mel_banks(
high_freq: float,
vtln_low: float,
vtln_high: float,
vtln_warp_factor: float,device=None,dtype=None
vtln_warp_factor: float,
device=None,
dtype=None,
) -> Tuple[Tensor, Tensor]:
"""
Returns:
@ -457,9 +459,9 @@ def get_mel_banks(
if high_freq <= 0.0:
high_freq += nyquist
assert (
(0.0 <= low_freq < nyquist) and (0.0 < high_freq <= nyquist) and (low_freq < high_freq)
), "Bad values in options: low-freq {} and high-freq {} vs. nyquist {}".format(low_freq, high_freq, nyquist)
assert (0.0 <= low_freq < nyquist) and (0.0 < high_freq <= nyquist) and (low_freq < high_freq), (
"Bad values in options: low-freq {} and high-freq {} vs. nyquist {}".format(low_freq, high_freq, nyquist)
)
# fft-bin width [think of it as Nyquist-freq / half-window-length]
fft_bin_width = sample_freq / window_length_padded
@ -475,7 +477,7 @@ def get_mel_banks(
assert vtln_warp_factor == 1.0 or (
(low_freq < vtln_low < high_freq) and (0.0 < vtln_high < high_freq) and (vtln_low < vtln_high)
), "Bad values in options: vtln-low {} and vtln-high {}, versus " "low-freq {} and high-freq {}".format(
), "Bad values in options: vtln-low {} and vtln-high {}, versus low-freq {} and high-freq {}".format(
vtln_low, vtln_high, low_freq, high_freq
)
@ -508,9 +510,12 @@ def get_mel_banks(
bins[up_idx] = up_slope[up_idx]
bins[down_idx] = down_slope[down_idx]
return bins.to(device=device,dtype=dtype)#, center_freqs
return bins.to(device=device, dtype=dtype) # , center_freqs
cache = {}
cache={}
def fbank(
waveform: Tensor,
blackman_coeff: float = 0.42,
@ -620,14 +625,34 @@ def fbank(
# size (num_mel_bins, padded_window_size // 2)
# print(num_mel_bins, padded_window_size, sample_frequency, low_freq, high_freq, vtln_low, vtln_high, vtln_warp)
cache_key="%s-%s-%s-%s-%s-%s-%s-%s-%s-%s"%(num_mel_bins, padded_window_size, sample_frequency, low_freq, high_freq, vtln_low, vtln_high, vtln_warp,device,dtype)
cache_key = "%s-%s-%s-%s-%s-%s-%s-%s-%s-%s" % (
num_mel_bins,
padded_window_size,
sample_frequency,
low_freq,
high_freq,
vtln_low,
vtln_high,
vtln_warp,
device,
dtype,
)
if cache_key not in cache:
mel_energies = get_mel_banks(
num_mel_bins, padded_window_size, sample_frequency, low_freq, high_freq, vtln_low, vtln_high, vtln_warp,device,dtype
num_mel_bins,
padded_window_size,
sample_frequency,
low_freq,
high_freq,
vtln_low,
vtln_high,
vtln_warp,
device,
dtype,
)
cache[cache_key]=mel_energies
cache[cache_key] = mel_energies
else:
mel_energies=cache[cache_key]
mel_energies = cache[cache_key]
# pad right column with zeros and add dimension, size (num_mel_bins, padded_window_size // 2 + 1)
mel_energies = torch.nn.functional.pad(mel_energies, (0, 1), mode="constant", value=0)

27
GPT_SoVITS/eres2net/pooling_layers.py
View File

@ -1,7 +1,7 @@
# Copyright 3D-Speaker (https://github.com/alibaba-damo-academy/3D-Speaker). All Rights Reserved.
# Licensed under the Apache License, Version 2.0 (http://www.apache.org/licenses/LICENSE-2.0)
""" This implementation is adapted from https://github.com/wenet-e2e/wespeaker."""
"""This implementation is adapted from https://github.com/wenet-e2e/wespeaker."""
import torch
import torch.nn as nn
@ -11,6 +11,7 @@ class TAP(nn.Module):
"""
Temporal average pooling, only first-order mean is considered
"""
def __init__(self, **kwargs):
super(TAP, self).__init__()
@ -25,6 +26,7 @@ class TSDP(nn.Module):
"""
Temporal standard deviation pooling, only second-order std is considered
"""
def __init__(self, **kwargs):
super(TSDP, self).__init__()
@ -41,6 +43,7 @@ class TSTP(nn.Module):
x-vector
Comment: simple concatenation can not make full use of both statistics
"""
def __init__(self, **kwargs):
super(TSTP, self).__init__()
@ -56,9 +59,10 @@ class TSTP(nn.Module):
class ASTP(nn.Module):
""" Attentive statistics pooling: Channel- and context-dependent
statistics pooling, first used in ECAPA_TDNN.
"""Attentive statistics pooling: Channel- and context-dependent
statistics pooling, first used in ECAPA_TDNN.
"""
def __init__(self, in_dim, bottleneck_dim=128, global_context_att=False):
super(ASTP, self).__init__()
self.global_context_att = global_context_att
@ -66,15 +70,10 @@ class ASTP(nn.Module):
# Use Conv1d with stride == 1 rather than Linear, then we don't
# need to transpose inputs.
if global_context_att:
self.linear1 = nn.Conv1d(
in_dim * 3, bottleneck_dim,
kernel_size=1) # equals W and b in the paper
self.linear1 = nn.Conv1d(in_dim * 3, bottleneck_dim, kernel_size=1) # equals W and b in the paper
else:
self.linear1 = nn.Conv1d(
in_dim, bottleneck_dim,
kernel_size=1) # equals W and b in the paper
self.linear2 = nn.Conv1d(bottleneck_dim, in_dim,
kernel_size=1) # equals V and k in the paper
self.linear1 = nn.Conv1d(in_dim, bottleneck_dim, kernel_size=1) # equals W and b in the paper
self.linear2 = nn.Conv1d(bottleneck_dim, in_dim, kernel_size=1) # equals V and k in the paper
def forward(self, x):
"""
@ -88,15 +87,13 @@ class ASTP(nn.Module):
if self.global_context_att:
context_mean = torch.mean(x, dim=-1, keepdim=True).expand_as(x)
context_std = torch.sqrt(
torch.var(x, dim=-1, keepdim=True) + 1e-10).expand_as(x)
context_std = torch.sqrt(torch.var(x, dim=-1, keepdim=True) + 1e-10).expand_as(x)
x_in = torch.cat((x, context_mean, context_std), dim=1)
else:
x_in = x
# DON'T use ReLU here! ReLU may be hard to converge.
alpha = torch.tanh(
self.linear1(x_in)) # alpha = F.relu(self.linear1(x_in))
alpha = torch.tanh(self.linear1(x_in)) # alpha = F.relu(self.linear1(x_in))
alpha = torch.softmax(self.linear2(alpha), dim=2)
mean = torch.sum(alpha * x, dim=2)
var = torch.sum(alpha * (x**2), dim=2) - mean**2

20
GPT_SoVITS/export_torch_script_v3v4.py
View File

@ -402,7 +402,7 @@ class GPTSoVITSV3(torch.nn.Module):
chunk_len = 934 - fea_ref.shape[2]
wav_gen_list = []
idx = 0
fea_todo = fea_todo[:,:,:-5]
fea_todo = fea_todo[:, :, :-5]
wav_gen_length = fea_todo.shape[2] * 256
while 1:
# current_time = datetime.now()
@ -435,6 +435,7 @@ class GPTSoVITSV3(torch.nn.Module):
wav_gen = torch.cat(wav_gen_list, 2)
return wav_gen[0][0][:wav_gen_length]
class GPTSoVITSV4(torch.nn.Module):
def __init__(self, gpt_sovits_half, cfm, hifigan):
super().__init__()
@ -461,7 +462,7 @@ class GPTSoVITSV4(torch.nn.Module):
chunk_len = 1000 - fea_ref.shape[2]
wav_gen_list = []
idx = 0
fea_todo = fea_todo[:,:,:-10]
fea_todo = fea_todo[:, :, :-10]
wav_gen_length = fea_todo.shape[2] * 480
while 1:
# current_time = datetime.now()
@ -577,6 +578,7 @@ from process_ckpt import get_sovits_version_from_path_fast, load_sovits_new
v3v4set = {"v3", "v4"}
def get_sovits_weights(sovits_path):
path_sovits_v3 = "GPT_SoVITS/pretrained_models/s2Gv3.pth"
is_exist_s2gv3 = os.path.exists(path_sovits_v3)
@ -699,7 +701,7 @@ def export_cfm(
return export_cfm
def export_1(ref_wav_path,ref_wav_text,version="v3"):
def export_1(ref_wav_path, ref_wav_text, version="v3"):
if version == "v3":
sovits = get_sovits_weights("GPT_SoVITS/pretrained_models/s2Gv3.pth")
init_bigvgan()
@ -707,7 +709,6 @@ def export_1(ref_wav_path,ref_wav_text,version="v3"):
sovits = get_sovits_weights("GPT_SoVITS/pretrained_models/gsv-v4-pretrained/s2Gv4.pth")
init_hifigan()
dict_s1 = torch.load("GPT_SoVITS/pretrained_models/s1v3.ckpt")
raw_t2s = get_raw_t2s_model(dict_s1).to(device)
print("#### get_raw_t2s_model ####")
@ -751,9 +752,7 @@ def export_1(ref_wav_path,ref_wav_text,version="v3"):
# phones1, bert1, norm_text1 = get_phones_and_bert(
# "你这老坏蛋，我找了你这么久，真没想到在这里找到你。他说。", "all_zh", "v3"
# )
phones1, bert1, norm_text1 = get_phones_and_bert(
ref_wav_text, "auto", "v3"
)
phones1, bert1, norm_text1 = get_phones_and_bert(ref_wav_text, "auto", "v3")
phones2, bert2, norm_text2 = get_phones_and_bert(
"这是一个简单的示例，真没想到这么简单就完成了。The King and His Stories.Once there was a king. He likes to write stories, but his stories were not good. As people were afraid of him, they all said his stories were good.After reading them, the writer at once turned to the soldiers and said: Take me back to prison, please.",
"auto",
@ -1201,7 +1200,6 @@ def export_2(version="v3"):
gpt_sovits_v3v4 = gpt_sovits_v3 if version == "v3" else gpt_sovits_v4
sr = 24000 if version == "v3" else 48000
time.sleep(5)
# print("thread:", torch.get_num_threads())
# print("thread:", torch.get_num_interop_threads())
@ -1212,14 +1210,14 @@ def export_2(version="v3"):
"汗流浃背了呀!老弟~ My uncle has two dogs. One is big and the other is small. He likes them very much. He often plays with them. He takes them for a walk every day. He says they are his good friends. He is very happy with them. 最后还是我得了 MVP....",
gpt_sovits_v3v4,
"out.wav",
sr
sr,
)
test_export(
"你小子是什么来路.汗流浃背了呀!老弟~ My uncle has two dogs. He is very happy with them. 最后还是我得了 MVP!",
gpt_sovits_v3v4,
"out2.wav",
sr
sr,
)
# test_export(
@ -1251,6 +1249,6 @@ def test_export_gpt_sovits_v3():
with torch.no_grad():
export_1("onnx/ad/ref.wav","你这老坏蛋，我找了你这么久，真没想到在这里找到你。他说。","v4")
export_1("onnx/ad/ref.wav", "你这老坏蛋，我找了你这么久，真没想到在这里找到你。他说。", "v4")
# export_2("v4")
# test_export_gpt_sovits_v3()

6
GPT_SoVITS/f5_tts/model/backbones/dit.py
View File

@ -143,9 +143,9 @@ class DiT(nn.Module):
drop_audio_cond=False, # cfg for cond audio
drop_text=False, # cfg for text
# mask: bool["b n"] | None = None, # noqa: F722
infer=False, # bool
text_cache=None, # torch tensor as text_embed
dt_cache=None, # torch tensor as dt
infer=False, # bool
text_cache=None, # torch tensor as text_embed
dt_cache=None, # torch tensor as dt
):
x = x0.transpose(2, 1)
cond = cond0.transpose(2, 1)

1
GPT_SoVITS/inference_webui.py
View File

@ -1337,5 +1337,6 @@ if __name__ == "__main__":
inbrowser=True,
share=is_share,
server_port=infer_ttswebui,
show_api=False,
# quiet=True,
)

1
GPT_SoVITS/inference_webui_fast.py
View File

@ -505,5 +505,6 @@ if __name__ == "__main__":
inbrowser=True,
share=is_share,
server_port=infer_ttswebui,
show_api=False,
# quiet=True,
)

41
GPT_SoVITS/module/data_utils.py
View File

@ -21,7 +21,7 @@ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
3) computes spectrograms from audio files.
"""
def __init__(self, hparams, version=None,val=False):
def __init__(self, hparams, version=None, val=False):
exp_dir = hparams.exp_dir
self.path2 = "%s/2-name2text.txt" % exp_dir
self.path4 = "%s/4-cnhubert" % exp_dir
@ -29,7 +29,7 @@ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
assert os.path.exists(self.path2)
assert os.path.exists(self.path4)
assert os.path.exists(self.path5)
self.is_v2Pro=version in {"v2Pro","v2ProPlus"}
self.is_v2Pro = version in {"v2Pro", "v2ProPlus"}
if self.is_v2Pro:
self.path7 = "%s/7-sv_cn" % exp_dir
assert os.path.exists(self.path7)
@ -118,7 +118,7 @@ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
ssl = F.pad(ssl.float(), (0, 1), mode="replicate").to(typee)
ssl.requires_grad = False
if self.is_v2Pro:
sv_emb=torch.load("%s/%s.pt" % (self.path7, audiopath), map_location="cpu")
sv_emb = torch.load("%s/%s.pt" % (self.path7, audiopath), map_location="cpu")
except:
traceback.print_exc()
spec = torch.zeros(1025, 100)
@ -126,10 +126,10 @@ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
ssl = torch.zeros(1, 768, 100)
text = text[-1:]
if self.is_v2Pro:
sv_emb=torch.zeros(1,20480)
sv_emb = torch.zeros(1, 20480)
print("load audio or ssl error!!!!!!", audiopath)
if self.is_v2Pro:
return (ssl, spec, wav, text,sv_emb)
return (ssl, spec, wav, text, sv_emb)
else:
return (ssl, spec, wav, text)
@ -192,9 +192,9 @@ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
class TextAudioSpeakerCollate:
"""Zero-pads model inputs and targets"""
def __init__(self, return_ids=False,version=None):
def __init__(self, return_ids=False, version=None):
self.return_ids = return_ids
self.is_v2Pro=version in {"v2Pro","v2ProPlus"}
self.is_v2Pro = version in {"v2Pro", "v2ProPlus"}
def __call__(self, batch):
"""Collate's training batch from normalized text, audio and speaker identities
@ -228,7 +228,7 @@ class TextAudioSpeakerCollate:
text_padded.zero_()
if self.is_v2Pro:
sv_embs=torch.FloatTensor(len(batch),20480)
sv_embs = torch.FloatTensor(len(batch), 20480)
for i in range(len(ids_sorted_decreasing)):
row = batch[ids_sorted_decreasing[i]]
@ -250,11 +250,30 @@ class TextAudioSpeakerCollate:
text_lengths[i] = text.size(0)
if self.is_v2Pro:
sv_embs[i]=row[4]
sv_embs[i] = row[4]
if self.is_v2Pro:
return ssl_padded, ssl_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, text_padded, text_lengths,sv_embs
return (
ssl_padded,
ssl_lengths,
spec_padded,
spec_lengths,
wav_padded,
wav_lengths,
text_padded,
text_lengths,
sv_embs,
)
else:
return ssl_padded, ssl_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, text_padded, text_lengths
return (
ssl_padded,
ssl_lengths,
spec_padded,
spec_lengths,
wav_padded,
wav_lengths,
text_padded,
text_lengths,
)
class TextAudioSpeakerLoaderV3(torch.utils.data.Dataset):

71
GPT_SoVITS/module/models.py
View File

@ -586,12 +586,17 @@ class DiscriminatorS(torch.nn.Module):
return x, fmap
v2pro_set={"v2Pro","v2ProPlus"}
v2pro_set = {"v2Pro", "v2ProPlus"}
class MultiPeriodDiscriminator(torch.nn.Module):
def __init__(self, use_spectral_norm=False,version=None):
def __init__(self, use_spectral_norm=False, version=None):
super(MultiPeriodDiscriminator, self).__init__()
if version in v2pro_set:periods = [2, 3, 5, 7, 11,17,23]
else:periods = [2, 3, 5, 7, 11]
if version in v2pro_set:
periods = [2, 3, 5, 7, 11, 17, 23]
else:
periods = [2, 3, 5, 7, 11]
discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
@ -787,6 +792,7 @@ class CodePredictor(nn.Module):
return pred_codes.transpose(0, 1)
class SynthesizerTrn(nn.Module):
"""
Synthesizer for Training
@ -886,13 +892,13 @@ class SynthesizerTrn(nn.Module):
self.quantizer = ResidualVectorQuantizer(dimension=ssl_dim, n_q=1, bins=1024)
self.freeze_quantizer = freeze_quantizer
self.is_v2pro=self.version in v2pro_set
self.is_v2pro = self.version in v2pro_set
if self.is_v2pro:
self.sv_emb = nn.Linear(20480, gin_channels)
self.ge_to512 = nn.Linear(gin_channels, 512)
self.prelu = nn.PReLU(num_parameters=gin_channels)
def forward(self, ssl, y, y_lengths, text, text_lengths,sv_emb=None):
def forward(self, ssl, y, y_lengths, text, text_lengths, sv_emb=None):
y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, y.size(2)), 1).to(y.dtype)
if self.version == "v1":
ge = self.ref_enc(y * y_mask, y_mask)
@ -952,7 +958,7 @@ class SynthesizerTrn(nn.Module):
return o, y_mask, (z, z_p, m_p, logs_p)
@torch.no_grad()
def decode(self, codes, text, refer,noise_scale=0.5, speed=1, sv_emb=None):
def decode(self, codes, text, refer, noise_scale=0.5, speed=1, sv_emb=None):
def get_ge(refer, sv_emb):
ge = None
if refer is not None:
@ -970,8 +976,8 @@ class SynthesizerTrn(nn.Module):
if type(refer) == list:
ges = []
for idx,_refer in enumerate(refer):
ge = get_ge(_refer, sv_emb[idx]if self.is_v2pro else None)
for idx, _refer in enumerate(refer):
ge = get_ge(_refer, sv_emb[idx] if self.is_v2pro else None)
ges.append(ge)
ge = torch.stack(ges, 0).mean(0)
else:
@ -983,7 +989,14 @@ class SynthesizerTrn(nn.Module):
quantized = self.quantizer.decode(codes)
if self.semantic_frame_rate == "25hz":
quantized = F.interpolate(quantized, size=int(quantized.shape[-1] * 2), mode="nearest")
x, m_p, logs_p, y_mask = self.enc_p(quantized, y_lengths, text, text_lengths, self.ge_to512(ge.transpose(2,1)).transpose(2,1)if self.is_v2pro else ge, speed)
x, m_p, logs_p, y_mask = self.enc_p(
quantized,
y_lengths,
text,
text_lengths,
self.ge_to512(ge.transpose(2, 1)).transpose(2, 1) if self.is_v2pro else ge,
speed,
)
z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
z = self.flow(z_p, y_mask, g=ge, reverse=True)
@ -996,6 +1009,7 @@ class SynthesizerTrn(nn.Module):
quantized, codes, commit_loss, quantized_list = self.quantizer(ssl)
return codes.transpose(0, 1)
class CFM(torch.nn.Module):
def __init__(self, in_channels, dit):
super().__init__()
@ -1029,7 +1043,18 @@ class CFM(torch.nn.Module):
t_tensor = torch.ones(x.shape[0], device=x.device, dtype=mu.dtype) * t
# v_pred = model(x, t_tensor, d_tensor, **extra_args)
v_pred, text_emb, dt = self.estimator(
x, prompt_x, x_lens, t_tensor, d_tensor, mu, use_grad_ckpt=False, drop_audio_cond=False, drop_text=False, infer=True, text_cache=text_cache, dt_cache=dt_cache
x,
prompt_x,
x_lens,
t_tensor,
d_tensor,
mu,
use_grad_ckpt=False,
drop_audio_cond=False,
drop_text=False,
infer=True,
text_cache=text_cache,
dt_cache=dt_cache,
)
v_pred = v_pred.transpose(2, 1)
if self.use_conditioner_cache:
@ -1037,18 +1062,18 @@ class CFM(torch.nn.Module):
dt_cache = dt
if inference_cfg_rate > 1e-5:
neg, text_cfg_emb, _ = self.estimator(
x,
prompt_x,
x_lens,
t_tensor,
d_tensor,
mu,
use_grad_ckpt=False,
drop_audio_cond=True,
drop_text=True,
infer=True,
text_cache=text_cfg_cache,
dt_cache=dt_cache
x,
prompt_x,
x_lens,
t_tensor,
d_tensor,
mu,
use_grad_ckpt=False,
drop_audio_cond=True,
drop_text=True,
infer=True,
text_cache=text_cfg_cache,
dt_cache=dt_cache,
)
neg = neg.transpose(2, 1)
if self.use_conditioner_cache:

7
GPT_SoVITS/module/modules.py
View File

@ -1,5 +1,4 @@
import math
import pdb
import numpy as np
import torch
@ -720,10 +719,10 @@ class MelStyleEncoder(nn.Module):
else:
len_ = (~mask).sum(dim=1).unsqueeze(1)
x = x.masked_fill(mask.unsqueeze(-1), 0)
dtype=x.dtype
dtype = x.dtype
x = x.float()
x=torch.div(x,len_.unsqueeze(1))
out=x.sum(dim=1).to(dtype)
x = torch.div(x, len_.unsqueeze(1))
out = x.sum(dim=1).to(dtype)
return out
def forward(self, x, mask=None):

50
GPT_SoVITS/prepare_datasets/2-get-sv.py
View File

@ -10,7 +10,6 @@ i_part = os.environ.get("i_part")
all_parts = os.environ.get("all_parts")
if "_CUDA_VISIBLE_DEVICES" in os.environ:
os.environ["CUDA_VISIBLE_DEVICES"] = os.environ["_CUDA_VISIBLE_DEVICES"]
from feature_extractor import cnhubert
opt_dir = os.environ.get("opt_dir")
sv_path = os.environ.get("sv_path")
@ -19,19 +18,18 @@ import torch
is_half = eval(os.environ.get("is_half", "True")) and torch.cuda.is_available()
import traceback
import numpy as np
from scipy.io import wavfile
import torchaudio
now_dir = os.getcwd()
sys.path.append(now_dir)
sys.path.append(f"{now_dir}/GPT_SoVITS/eres2net")
from tools.my_utils import load_audio, clean_path
from tools.my_utils import clean_path
from time import time as ttime
import shutil
from ERes2NetV2 import ERes2NetV2
import kaldi as Kaldi
def my_save(fea, path): #####fix issue: torch.save doesn't support chinese path
dir = os.path.dirname(path)
name = os.path.basename(path)
@ -56,37 +54,45 @@ if torch.cuda.is_available():
else:
device = "cpu"
class SV:
def __init__(self,device,is_half):
pretrained_state = torch.load(sv_path, map_location='cpu')
embedding_model = ERes2NetV2(baseWidth=24,scale=4,expansion=4)
def __init__(self, device, is_half):
pretrained_state = torch.load(sv_path, map_location="cpu")
embedding_model = ERes2NetV2(baseWidth=24, scale=4, expansion=4)
embedding_model.load_state_dict(pretrained_state)
embedding_model.eval()
self.embedding_model=embedding_model
self.res=torchaudio.transforms.Resample(32000, 16000).to(device)
self.embedding_model = embedding_model
self.res = torchaudio.transforms.Resample(32000, 16000).to(device)
if is_half == False:
self.embedding_model=self.embedding_model.to(device)
self.embedding_model = self.embedding_model.to(device)
else:
self.embedding_model=self.embedding_model.half().to(device)
self.is_half=is_half
self.embedding_model = self.embedding_model.half().to(device)
self.is_half = is_half
def compute_embedding3(self,wav):#(1,x)#-1~1
def compute_embedding3(self, wav): # (1,x)#-1~1
with torch.no_grad():
wav=self.res(wav)
if self.is_half==True:wav=wav.half()
feat = torch.stack([Kaldi.fbank(wav0.unsqueeze(0), num_mel_bins=80, sample_frequency=16000, dither=0) for wav0 in wav])
wav = self.res(wav)
if self.is_half == True:
wav = wav.half()
feat = torch.stack(
[Kaldi.fbank(wav0.unsqueeze(0), num_mel_bins=80, sample_frequency=16000, dither=0) for wav0 in wav]
)
sv_emb = self.embedding_model.forward3(feat)
return sv_emb
sv=SV(device,is_half)
sv = SV(device, is_half)
def name2go(wav_name, wav_path):
sv_cn_path = "%s/%s.pt" % (sv_cn_dir, wav_name)
if os.path.exists(sv_cn_path):return
wav_path="%s/%s" % (wav32dir, wav_name)
wav32k,sr0 = torchaudio.load(wav_path)
assert sr0==32000
if os.path.exists(sv_cn_path):
return
wav_path = "%s/%s" % (wav32dir, wav_name)
wav32k, sr0 = torchaudio.load(wav_path)
assert sr0 == 32000
wav32k = wav32k.to(device)
emb=sv.compute_embedding3(wav32k).cpu() # torch.Size([1, 20480])
emb = sv.compute_embedding3(wav32k).cpu() # torch.Size([1, 20480])
my_save(emb, sv_cn_path)

16
GPT_SoVITS/process_ckpt.py
View File

@ -17,15 +17,16 @@ def my_save(fea, path): #####fix issue: torch.save doesn't support chinese path
shutil.move(tmp_path, "%s/%s" % (dir, name))
from io import BytesIO
model_version2byte={
"v3":b"03",
"v4":b"04",
"v2Pro":b"05",
"v2ProPlus":b"06",
model_version2byte = {
"v3": b"03",
"v4": b"04",
"v2Pro": b"05",
"v2ProPlus": b"06",
}
def my_save2(fea, path, model_version):
bio = BytesIO()
torch.save(fea, bio)
@ -50,7 +51,7 @@ def savee(ckpt, name, epoch, steps, hps, model_version=None, lora_rank=None):
if lora_rank:
opt["lora_rank"] = lora_rank
my_save2(opt, "%s/%s.pth" % (hps.save_weight_dir, name), model_version)
elif (model_version!=None and "Pro"in model_version):
elif model_version != None and "Pro" in model_version:
my_save2(opt, "%s/%s.pth" % (hps.save_weight_dir, name), model_version)
else:
my_save(opt, "%s/%s.pth" % (hps.save_weight_dir, name))
@ -58,6 +59,7 @@ def savee(ckpt, name, epoch, steps, hps, model_version=None, lora_rank=None):
except:
return traceback.format_exc()
"""
00:v1
01:v2

93
GPT_SoVITS/s2_train.py
View File

@ -36,7 +36,7 @@ from module.models import (
MultiPeriodDiscriminator,
SynthesizerTrn,
)
from process_ckpt import savee,my_save2
from process_ckpt import savee
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = False
@ -87,11 +87,30 @@ def run(rank, n_gpus, hps):
if torch.cuda.is_available():
torch.cuda.set_device(rank)
train_dataset = TextAudioSpeakerLoader(hps.data,version=hps.model.version)
train_dataset = TextAudioSpeakerLoader(hps.data, version=hps.model.version)
train_sampler = DistributedBucketSampler(
train_dataset,
hps.train.batch_size,
[32,300,400,500,600,700,800,900,1000,1100,1200,1300,1400,1500,1600,1700,1800,1900,],
[
32,
300,
400,
500,
600,
700,
800,
900,
1000,
1100,
1200,
1300,
1400,
1500,
1600,
1700,
1800,
1900,
],
num_replicas=n_gpus,
rank=rank,
shuffle=True,
@ -130,9 +149,9 @@ def run(rank, n_gpus, hps):
)
net_d = (
MultiPeriodDiscriminator(hps.model.use_spectral_norm,version=hps.model.version).cuda(rank)
MultiPeriodDiscriminator(hps.model.use_spectral_norm, version=hps.model.version).cuda(rank)
if torch.cuda.is_available()
else MultiPeriodDiscriminator(hps.model.use_spectral_norm,version=hps.model.version).to(device)
else MultiPeriodDiscriminator(hps.model.use_spectral_norm, version=hps.model.version).to(device)
)
for name, param in net_g.named_parameters():
if not param.requires_grad:
@ -235,7 +254,7 @@ def run(rank, n_gpus, hps):
print(
"loaded pretrained %s" % hps.train.pretrained_s2D,
net_d.module.load_state_dict(
torch.load(hps.train.pretrained_s2D, map_location="cpu", weights_only=False)["weight"],strict=False
torch.load(hps.train.pretrained_s2D, map_location="cpu", weights_only=False)["weight"], strict=False
)
if torch.cuda.is_available()
else net_d.load_state_dict(
@ -310,17 +329,44 @@ def train_and_evaluate(rank, epoch, hps, nets, optims, schedulers, scaler, loade
net_g.train()
net_d.train()
for batch_idx, data in enumerate(tqdm(train_loader)):
if hps.model.version in {"v2Pro","v2ProPlus"}:
ssl, ssl_lengths, spec, spec_lengths, y, y_lengths, text, text_lengths,sv_emb=data
if hps.model.version in {"v2Pro", "v2ProPlus"}:
ssl, ssl_lengths, spec, spec_lengths, y, y_lengths, text, text_lengths, sv_emb = data
else:
ssl, ssl_lengths, spec, spec_lengths, y, y_lengths, text, text_lengths=data
ssl, ssl_lengths, spec, spec_lengths, y, y_lengths, text, text_lengths = data
if torch.cuda.is_available():
spec, spec_lengths = (spec.cuda(rank,non_blocking=True,),spec_lengths.cuda(rank,non_blocking=True,),)
y, y_lengths = (y.cuda(rank,non_blocking=True,),y_lengths.cuda(rank,non_blocking=True,),)
spec, spec_lengths = (
spec.cuda(
rank,
non_blocking=True,
),
spec_lengths.cuda(
rank,
non_blocking=True,
),
)
y, y_lengths = (
y.cuda(
rank,
non_blocking=True,
),
y_lengths.cuda(
rank,
non_blocking=True,
),
)
ssl = ssl.cuda(rank, non_blocking=True)
ssl.requires_grad = False
# ssl_lengths = ssl_lengths.cuda(rank, non_blocking=True)
text, text_lengths = (text.cuda(rank,non_blocking=True,),text_lengths.cuda(rank,non_blocking=True,),)
text, text_lengths = (
text.cuda(
rank,
non_blocking=True,
),
text_lengths.cuda(
rank,
non_blocking=True,
),
)
if hps.model.version in {"v2Pro", "v2ProPlus"}:
sv_emb = sv_emb.cuda(rank, non_blocking=True)
else:
@ -334,9 +380,19 @@ def train_and_evaluate(rank, epoch, hps, nets, optims, schedulers, scaler, loade
sv_emb = sv_emb.to(device)
with autocast(enabled=hps.train.fp16_run):
if hps.model.version in {"v2Pro", "v2ProPlus"}:
(y_hat,kl_ssl,ids_slice,x_mask,z_mask,(z, z_p, m_p, logs_p, m_q, logs_q),stats_ssl) = net_g(ssl, spec, spec_lengths, text, text_lengths,sv_emb)
(y_hat, kl_ssl, ids_slice, x_mask, z_mask, (z, z_p, m_p, logs_p, m_q, logs_q), stats_ssl) = net_g(
ssl, spec, spec_lengths, text, text_lengths, sv_emb
)
else:
(y_hat,kl_ssl,ids_slice,x_mask,z_mask,(z, z_p, m_p, logs_p, m_q, logs_q),stats_ssl,) = net_g(ssl, spec, spec_lengths, text, text_lengths)
(
y_hat,
kl_ssl,
ids_slice,
x_mask,
z_mask,
(z, z_p, m_p, logs_p, m_q, logs_q),
stats_ssl,
) = net_g(ssl, spec, spec_lengths, text, text_lengths)
mel = spec_to_mel_torch(
spec,
@ -508,7 +564,14 @@ def train_and_evaluate(rank, epoch, hps, nets, optims, schedulers, scaler, loade
% (
hps.name,
epoch,
savee(ckpt,hps.name + "_e%s_s%s" % (epoch, global_step),epoch,global_step,hps,model_version=None if hps.model.version not in {"v2Pro","v2ProPlus"}else hps.model.version),
savee(
ckpt,
hps.name + "_e%s_s%s" % (epoch, global_step),
epoch,
global_step,
hps,
model_version=None if hps.model.version not in {"v2Pro", "v2ProPlus"} else hps.model.version,
),
)
)

30
GPT_SoVITS/sv.py
View File

@ -1,24 +1,32 @@
import sys,os,torch
import sys
import os
import torch
sys.path.append(f"{os.getcwd()}/GPT_SoVITS/eres2net")
sv_path = "GPT_SoVITS/pretrained_models/sv/pretrained_eres2netv2w24s4ep4.ckpt"
from ERes2NetV2 import ERes2NetV2
import kaldi as Kaldi
class SV:
def __init__(self,device,is_half):
pretrained_state = torch.load(sv_path, map_location='cpu', weights_only=False)
embedding_model = ERes2NetV2(baseWidth=24,scale=4,expansion=4)
def __init__(self, device, is_half):
pretrained_state = torch.load(sv_path, map_location="cpu", weights_only=False)
embedding_model = ERes2NetV2(baseWidth=24, scale=4, expansion=4)
embedding_model.load_state_dict(pretrained_state)
embedding_model.eval()
self.embedding_model=embedding_model
self.embedding_model = embedding_model
if is_half == False:
self.embedding_model=self.embedding_model.to(device)
self.embedding_model = self.embedding_model.to(device)
else:
self.embedding_model=self.embedding_model.half().to(device)
self.is_half=is_half
self.embedding_model = self.embedding_model.half().to(device)
self.is_half = is_half
def compute_embedding3(self,wav):
def compute_embedding3(self, wav):
with torch.no_grad():
if self.is_half==True:wav=wav.half()
feat = torch.stack([Kaldi.fbank(wav0.unsqueeze(0), num_mel_bins=80, sample_frequency=16000, dither=0) for wav0 in wav])
if self.is_half == True:
wav = wav.half()
feat = torch.stack(
[Kaldi.fbank(wav0.unsqueeze(0), num_mel_bins=80, sample_frequency=16000, dither=0) for wav0 in wav]
)
sv_emb = self.embedding_model.forward3(feat)
return sv_emb

103
GPT_SoVITS/text/LangSegmenter/langsegmenter.py
View File

@ -3,38 +3,44 @@ import re
# jieba静音
import jieba
jieba.setLogLevel(logging.CRITICAL)
# 更改fast_langdetect大模型位置
from pathlib import Path
import fast_langdetect
fast_langdetect.infer._default_detector = fast_langdetect.infer.LangDetector(fast_langdetect.infer.LangDetectConfig(cache_dir=Path(__file__).parent.parent.parent / "pretrained_models" / "fast_langdetect"))
fast_langdetect.infer._default_detector = fast_langdetect.infer.LangDetector(
fast_langdetect.infer.LangDetectConfig(
cache_dir=Path(__file__).parent.parent.parent / "pretrained_models" / "fast_langdetect"
)
)
from split_lang import LangSplitter
def full_en(text):
pattern = r'^(?=.*[A-Za-z])[A-Za-z0-9\s\u0020-\u007E\u2000-\u206F\u3000-\u303F\uFF00-\uFFEF]+$'
pattern = r"^(?=.*[A-Za-z])[A-Za-z0-9\s\u0020-\u007E\u2000-\u206F\u3000-\u303F\uFF00-\uFFEF]+$"
return bool(re.match(pattern, text))
def full_cjk(text):
# 来自wiki
cjk_ranges = [
(0x4E00, 0x9FFF), # CJK Unified Ideographs
(0x3400, 0x4DB5), # CJK Extension A
(0x20000, 0x2A6DD), # CJK Extension B
(0x2A700, 0x2B73F), # CJK Extension C
(0x2B740, 0x2B81F), # CJK Extension D
(0x2B820, 0x2CEAF), # CJK Extension E
(0x2CEB0, 0x2EBEF), # CJK Extension F
(0x30000, 0x3134A), # CJK Extension G
(0x31350, 0x323AF), # CJK Extension H
(0x2EBF0, 0x2EE5D), # CJK Extension H
(0x4E00, 0x9FFF), # CJK Unified Ideographs
(0x3400, 0x4DB5), # CJK Extension A
(0x20000, 0x2A6DD), # CJK Extension B
(0x2A700, 0x2B73F), # CJK Extension C
(0x2B740, 0x2B81F), # CJK Extension D
(0x2B820, 0x2CEAF), # CJK Extension E
(0x2CEB0, 0x2EBEF), # CJK Extension F
(0x30000, 0x3134A), # CJK Extension G
(0x31350, 0x323AF), # CJK Extension H
(0x2EBF0, 0x2EE5D), # CJK Extension H
]
pattern = r'[0-9、-〜。！？.!?… /]+$'
pattern = r"[0-9、-〜。！？.!?… /]+$"
cjk_text = ""
for char in text:
@ -45,7 +51,7 @@ def full_cjk(text):
return cjk_text
def split_jako(tag_lang,item):
def split_jako(tag_lang, item):
if tag_lang == "ja":
pattern = r"([\u3041-\u3096\u3099\u309A\u30A1-\u30FA\u30FC]+(?:[0-9、-〜。！？.!?… ]+[\u3041-\u3096\u3099\u309A\u30A1-\u30FA\u30FC]*)*)"
else:
@ -53,41 +59,40 @@ def split_jako(tag_lang,item):
lang_list: list[dict] = []
tag = 0
for match in re.finditer(pattern, item['text']):
for match in re.finditer(pattern, item["text"]):
if match.start() > tag:
lang_list.append({'lang':item['lang'],'text':item['text'][tag:match.start()]})
lang_list.append({"lang": item["lang"], "text": item["text"][tag : match.start()]})
tag = match.end()
lang_list.append({'lang':tag_lang,'text':item['text'][match.start():match.end()]})
lang_list.append({"lang": tag_lang, "text": item["text"][match.start() : match.end()]})
if tag < len(item['text']):
lang_list.append({'lang':item['lang'],'text':item['text'][tag:len(item['text'])]})
if tag < len(item["text"]):
lang_list.append({"lang": item["lang"], "text": item["text"][tag : len(item["text"])]})
return lang_list
def merge_lang(lang_list, item):
if lang_list and item['lang'] == lang_list[-1]['lang']:
lang_list[-1]['text'] += item['text']
if lang_list and item["lang"] == lang_list[-1]["lang"]:
lang_list[-1]["text"] += item["text"]
else:
lang_list.append(item)
return lang_list
class LangSegmenter():
class LangSegmenter:
# 默认过滤器, 基于gsv目前四种语言
DEFAULT_LANG_MAP = {
"zh": "zh",
"yue": "zh", # 粤语
"wuu": "zh", # 吴语
"zh-cn": "zh",
"zh-tw": "x", # 繁体设置为x
"zh-tw": "x", # 繁体设置为x
"ko": "ko",
"ja": "ja",
"en": "en",
}
def getTexts(text):
lang_splitter = LangSplitter(lang_map=LangSegmenter.DEFAULT_LANG_MAP)
substr = lang_splitter.split_by_lang(text=text)
@ -95,18 +100,18 @@ class LangSegmenter():
lang_list: list[dict] = []
for _, item in enumerate(substr):
dict_item = {'lang':item.lang,'text':item.text}
dict_item = {"lang": item.lang, "text": item.text}
# 处理短英文被识别为其他语言的问题
if full_en(dict_item['text']):
dict_item['lang'] = 'en'
lang_list = merge_lang(lang_list,dict_item)
if full_en(dict_item["text"]):
dict_item["lang"] = "en"
lang_list = merge_lang(lang_list, dict_item)
continue
# 处理非日语夹日文的问题(不包含CJK)
ja_list: list[dict] = []
if dict_item['lang'] != 'ja':
ja_list = split_jako('ja',dict_item)
if dict_item["lang"] != "ja":
ja_list = split_jako("ja", dict_item)
if not ja_list:
ja_list.append(dict_item)
@ -115,8 +120,8 @@ class LangSegmenter():
ko_list: list[dict] = []
temp_list: list[dict] = []
for _, ko_item in enumerate(ja_list):
if ko_item["lang"] != 'ko':
ko_list = split_jako('ko',ko_item)
if ko_item["lang"] != "ko":
ko_list = split_jako("ko", ko_item)
if ko_list:
temp_list.extend(ko_list)
@ -126,43 +131,43 @@ class LangSegmenter():
# 未存在非日韩文夹日韩文
if len(temp_list) == 1:
# 未知语言检查是否为CJK
if dict_item['lang'] == 'x':
cjk_text = full_cjk(dict_item['text'])
if dict_item["lang"] == "x":
cjk_text = full_cjk(dict_item["text"])
if cjk_text:
dict_item = {'lang':'zh','text':cjk_text}
lang_list = merge_lang(lang_list,dict_item)
dict_item = {"lang": "zh", "text": cjk_text}
lang_list = merge_lang(lang_list, dict_item)
else:
lang_list = merge_lang(lang_list,dict_item)
lang_list = merge_lang(lang_list, dict_item)
continue
else:
lang_list = merge_lang(lang_list,dict_item)
lang_list = merge_lang(lang_list, dict_item)
continue
# 存在非日韩文夹日韩文
for _, temp_item in enumerate(temp_list):
# 未知语言检查是否为CJK
if temp_item['lang'] == 'x':
cjk_text = full_cjk(dict_item['text'])
if temp_item["lang"] == "x":
cjk_text = full_cjk(dict_item["text"])
if cjk_text:
dict_item = {'lang':'zh','text':cjk_text}
lang_list = merge_lang(lang_list,dict_item)
dict_item = {"lang": "zh", "text": cjk_text}
lang_list = merge_lang(lang_list, dict_item)
else:
lang_list = merge_lang(lang_list,dict_item)
lang_list = merge_lang(lang_list, dict_item)
else:
lang_list = merge_lang(lang_list,temp_item)
lang_list = merge_lang(lang_list, temp_item)
temp_list = lang_list
lang_list = []
for _, temp_item in enumerate(temp_list):
if temp_item['lang'] == 'x':
if temp_item["lang"] == "x":
if lang_list:
temp_item['lang'] = lang_list[-1]['lang']
temp_item["lang"] = lang_list[-1]["lang"]
elif len(temp_list) > 1:
temp_item['lang'] = temp_list[1]['lang']
temp_item["lang"] = temp_list[1]["lang"]
else:
temp_item['lang'] = 'zh'
temp_item["lang"] = "zh"
lang_list = merge_lang(lang_list,temp_item)
lang_list = merge_lang(lang_list, temp_item)
return lang_list

6
GPT_SoVITS/text/g2pw/onnx_api.py
View File

@ -3,7 +3,6 @@
import json
import os
import traceback
import warnings
import zipfile
from typing import Any, Dict, List, Tuple
@ -23,8 +22,9 @@ from .utils import load_config
onnxruntime.set_default_logger_severity(3)
try:
onnxruntime.preload_dlls()
except:pass
#traceback.print_exc()
except:
pass
# traceback.print_exc()
warnings.filterwarnings("ignore")
model_version = "1.1"

6
GPT_SoVITS/text/tone_sandhi.py
View File

@ -655,11 +655,7 @@ class ToneSandhi:
while i < len(seg):
word, pos = seg[i]
merged = False
if (
i - 1 >= 0
and word == "一"
and i + 1 < len(seg)
):
if i - 1 >= 0 and word == "一" and i + 1 < len(seg):
last = new_seg[-1] if new_seg else seg[i - 1]
if last[0] == seg[i + 1][0] and last[1] == "v" and seg[i + 1][1] == "v":
combined = last[0] + "一" + seg[i + 1][0]

55
README.md
View File

@ -1,3 +1,5 @@
#
<div align="center">
<h1>GPT-SoVITS-WebUI</h1>
@ -7,12 +9,17 @@ A Powerful Few-shot Voice Conversion and Text-to-Speech WebUI.<br><br>
<a href="https://trendshift.io/repositories/7033" target="_blank"><img src="https://trendshift.io/api/badge/repositories/7033" alt="RVC-Boss%2FGPT-SoVITS | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
<!-- img src="https://counter.seku.su/cmoe?name=gptsovits&theme=r34" /><br> -->
[![Train In Colab](https://img.shields.io/badge/Colab-Training-F9AB00?style=for-the-badge&logo=googlecolab)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/Colab-WebUI.ipynb)
[![Infer In Colab](https://img.shields.io/badge/Colab-Inference-F9AB00?style=for-the-badge&logo=googlecolab)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/Colab-Inference.ipynb)
[![Huggingface](https://img.shields.io/badge/HuggingFace-online%20demo-blue.svg?style=for-the-badge&logo=huggingface)](https://huggingface.co/spaces/lj1995/GPT-SoVITS-v2)
[![Open In Colab](https://img.shields.io/badge/Colab-F9AB00?style=for-the-badge&logo=googlecolab&color=525252)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/colab_webui.ipynb)
[![License](https://img.shields.io/badge/LICENSE-MIT-green.svg?style=for-the-badge)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/LICENSE)
[![Huggingface](https://img.shields.io/badge/🤗%20-online%20demo-yellow.svg?style=for-the-badge)](https://huggingface.co/spaces/lj1995/GPT-SoVITS-v2)
[![Discord](https://img.shields.io/discord/1198701940511617164?color=%23738ADB&label=Discord&style=for-the-badge)](https://discord.gg/dnrgs5GHfG)
[![GitHub release](https://img.shields.io/github/v/release/RVC-Boss/gpt-sovits?style=for-the-badge&logo=github)](https://github.com/RVC-Boss/gpt-sovits/releases)
[![Image Size](https://img.shields.io/docker/image-size/xxxxrt666/gpt-sovits/latest?style=for-the-badge&logo=docker)](https://hub.docker.com/r/xxxxrt666/gpt-sovits)
[![License](https://img.shields.io/badge/LICENSE-MIT-green.svg?style=for-the-badge&logo=opensourceinitiative)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/LICENSE)
[![简体中文](https://img.shields.io/badge/简体中文-阅读文档-blue?style=for-the-badge&logo=googledocs&logoColor=white)](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e)
[![English](https://img.shields.io/badge/English-Read%20Docs-blue?style=for-the-badge&logo=googledocs&logoColor=white)](https://rentry.co/GPT-SoVITS-guide#/)
[![Change Log](https://img.shields.io/badge/Change%20Log-View%20Updates-blue?style=for-the-badge&logo=googledocs&logoColor=white)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/docs/en/Changelog_EN.md)
**English** | [**中文简体**](./docs/cn/README.md) | [**日本語**](./docs/ja/README.md) | [**한국어**](./docs/ko/README.md) | [**Türkçe**](./docs/tr/README.md)
@ -20,7 +27,7 @@ A Powerful Few-shot Voice Conversion and Text-to-Speech WebUI.<br><br>
---
## Features:
## Features
1. **Zero-shot TTS:** Input a 5-second vocal sample and experience instant text-to-speech conversion.
@ -34,13 +41,13 @@ A Powerful Few-shot Voice Conversion and Text-to-Speech WebUI.<br><br>
Unseen speakers few-shot fine-tuning demo:
https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-80c060ab47fb
<https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-80c060ab47fb>
**User guide: [简体中文](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e) | [English](https://rentry.co/GPT-SoVITS-guide#/)**
<!-- **User guide: [简体中文](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e) | [English](https://rentry.co/GPT-SoVITS-guide#/)** -->
## Installation
For users in China, you can [click here](https://www.codewithgpu.com/i/RVC-Boss/GPT-SoVITS/GPT-SoVITS-Official) to use AutoDL Cloud Docker to experience the full functionality online.
For users in China, you can use [AutoDL Cloud Docker](https://www.codewithgpu.com/i/RVC-Boss/GPT-SoVITS/GPT-SoVITS-Official) to experience the full functionality online.
### Tested Environments
@ -171,7 +178,7 @@ docker exec -it <GPT-SoVITS-CU126-Lite|GPT-SoVITS-CU128-Lite|GPT-SoVITS-CU126|GP
## Pretrained Models
**If `install.sh` runs successfully, you may skip No.1,2,3**
**If `install.sh` runs successfully, you may skip No.1, 2, 3**
**Users in China can [download all these models here](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e/dkxgpiy9zb96hob4#nVNhX).**
@ -193,10 +200,8 @@ docker exec -it <GPT-SoVITS-CU126-Lite|GPT-SoVITS-CU128-Lite|GPT-SoVITS-CU126|GP
The TTS annotation .list file format:
```
```text
vocal_path|speaker_name|language|text
```
Language dictionary:
@ -209,10 +214,8 @@ Language dictionary:
Example:
```
```text
D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
```
## Finetune and inference
@ -222,7 +225,6 @@ D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
#### Integrated Package Users
Double-click `go-webui.bat`or use `go-webui.ps1`
if you want to switch to V1,then double-click`go-webui-v1.bat` or use `go-webui-v1.ps1`
#### Others
@ -230,14 +232,6 @@ if you want to switch to V1,then double-click`go-webui-v1.bat` or use `go-webui-
python webui.py <language(optional)>
```
if you want to switch to V1,then
```bash
python webui.py v1 <language(optional)>
```
Or maunally switch version in WebUI
### Finetune
#### Path Auto-filling is now supported
@ -253,7 +247,7 @@ Or maunally switch version in WebUI
#### Integrated Package Users
Double-click `go-webui-v2.bat` or use `go-webui-v2.ps1` ,then open the inference webui at `1-GPT-SoVITS-TTS/1C-inference`
Double-click `go-webui.bat` or use `go-webui.ps1` , then open the inference webui at `1-GPT-SoVITS-TTS/1C-inference`
#### Others
@ -333,7 +327,7 @@ Use v4 from v1/v2/v3 environment:
New Features:
1. Slightly higher VRAM usage than v2, surpassing v4's performance, with v2's hardware cost and speed.
[more details](https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90features-(%E5%90%84%E7%89%88%E6%9C%AC%E7%89%B9%E6%80%A7))
[more details](<https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90features-(%E5%90%84%E7%89%88%E6%9C%AC%E7%89%B9%E6%80%A7)>)
2.v1/v2 and the v2Pro series share the same characteristics, while v3/v4 have similar features. For training sets with average audio quality, v1/v2/v2Pro can deliver decent results, but v3/v4 cannot. Additionally, the synthesized tone and timebre of v3/v4 lean more toward the reference audio rather than the overall training set.
@ -373,11 +367,6 @@ Use the command line to open the WebUI for UVR5
python tools/uvr5/webui.py "<infer_device>" <is_half> <webui_port_uvr5>
```
<!-- If you can't open a browser, follow the format below for UVR processing,This is using mdxnet for audio processing
```
python mdxnet.py --model --input_root --output_vocal --output_ins --agg_level --format --device --is_half_precision
``` -->
This is how the audio segmentation of the dataset is done using the command line
```bash
@ -453,5 +442,5 @@ Thankful to @Naozumi520 for providing the Cantonese training set and for the gui
## Thanks to all contributors for their efforts
<a href="https://github.com/RVC-Boss/GPT-SoVITS/graphs/contributors" target="_blank">
<img src="https://contrib.rocks/image?repo=RVC-Boss/GPT-SoVITS" />
<img src="https://contrib.rocks/image?repo=RVC-Boss/GPT-SoVITS" alt="Contributors"/>
</a>

84
api.py
View File

@ -199,6 +199,8 @@ def is_full(*items): # 任意一项为空返回False
bigvgan_model = hifigan_model = sv_cn_model = None
def clean_hifigan_model():
global hifigan_model
if hifigan_model:
@ -208,6 +210,8 @@ def clean_hifigan_model():
torch.cuda.empty_cache()
except:
pass
def clean_bigvgan_model():
global bigvgan_model
if bigvgan_model:
@ -217,6 +221,8 @@ def clean_bigvgan_model():
torch.cuda.empty_cache()
except:
pass
def clean_sv_cn_model():
global sv_cn_model
if sv_cn_model:
@ -229,7 +235,7 @@ def clean_sv_cn_model():
def init_bigvgan():
global bigvgan_model, hifigan_model,sv_cn_model
global bigvgan_model, hifigan_model, sv_cn_model
from BigVGAN import bigvgan
bigvgan_model = bigvgan.BigVGAN.from_pretrained(
@ -247,7 +253,7 @@ def init_bigvgan():
def init_hifigan():
global hifigan_model, bigvgan_model,sv_cn_model
global hifigan_model, bigvgan_model, sv_cn_model
hifigan_model = Generator(
initial_channel=100,
resblock="1",
@ -262,7 +268,9 @@ def init_hifigan():
hifigan_model.eval()
hifigan_model.remove_weight_norm()
state_dict_g = torch.load(
"%s/GPT_SoVITS/pretrained_models/gsv-v4-pretrained/vocoder.pth" % (now_dir,), map_location="cpu", weights_only=False
"%s/GPT_SoVITS/pretrained_models/gsv-v4-pretrained/vocoder.pth" % (now_dir,),
map_location="cpu",
weights_only=False,
)
print("loading vocoder", hifigan_model.load_state_dict(state_dict_g))
if is_half == True:
@ -272,19 +280,21 @@ def init_hifigan():
from sv import SV
def init_sv_cn():
global hifigan_model, bigvgan_model, sv_cn_model
sv_cn_model = SV(device, is_half)
resample_transform_dict={}
def resample(audio_tensor, sr0,sr1,device):
resample_transform_dict = {}
def resample(audio_tensor, sr0, sr1, device):
global resample_transform_dict
key="%s-%s-%s"%(sr0,sr1,str(device))
key = "%s-%s-%s" % (sr0, sr1, str(device))
if key not in resample_transform_dict:
resample_transform_dict[key] = torchaudio.transforms.Resample(
sr0, sr1
).to(device)
resample_transform_dict[key] = torchaudio.transforms.Resample(sr0, sr1).to(device)
return resample_transform_dict[key](audio_tensor)
@ -370,6 +380,7 @@ from process_ckpt import get_sovits_version_from_path_fast, load_sovits_new
def get_sovits_weights(sovits_path):
from config import pretrained_sovits_name
path_sovits_v3 = pretrained_sovits_name["v3"]
path_sovits_v4 = pretrained_sovits_name["v4"]
is_exist_s2gv3 = os.path.exists(path_sovits_v3)
@ -628,15 +639,17 @@ class DictToAttrRecursive(dict):
def get_spepc(hps, filename, dtype, device, is_v2pro=False):
sr1=int(hps.data.sampling_rate)
audio, sr0=torchaudio.load(filename)
if sr0!=sr1:
audio=audio.to(device)
if(audio.shape[0]==2):audio=audio.mean(0).unsqueeze(0)
audio=resample(audio,sr0,sr1,device)
sr1 = int(hps.data.sampling_rate)
audio, sr0 = torchaudio.load(filename)
if sr0 != sr1:
audio = audio.to(device)
if audio.shape[0] == 2:
audio = audio.mean(0).unsqueeze(0)
audio = resample(audio, sr0, sr1, device)
else:
audio=audio.to(device)
if(audio.shape[0]==2):audio=audio.mean(0).unsqueeze(0)
audio = audio.to(device)
if audio.shape[0] == 2:
audio = audio.mean(0).unsqueeze(0)
maxx = audio.abs().max()
if maxx > 1:
@ -649,10 +662,10 @@ def get_spepc(hps, filename, dtype, device, is_v2pro=False):
hps.data.win_length,
center=False,
)
spec=spec.to(dtype)
if is_v2pro==True:
audio=resample(audio,sr1,16000,device).to(dtype)
return spec,audio
spec = spec.to(dtype)
if is_v2pro == True:
audio = resample(audio, sr1, 16000, device).to(dtype)
return spec, audio
def pack_audio(audio_bytes, data, rate):
@ -872,29 +885,29 @@ def get_tts_wav(
prompt_semantic = codes[0, 0]
prompt = prompt_semantic.unsqueeze(0).to(device)
is_v2pro = version in {"v2Pro","v2ProPlus"}
is_v2pro = version in {"v2Pro", "v2ProPlus"}
if version not in {"v3", "v4"}:
refers = []
if is_v2pro:
sv_emb= []
sv_emb = []
if sv_cn_model == None:
init_sv_cn()
if inp_refs:
for path in inp_refs:
try:#####这里加上提取sv的逻辑，要么一堆sv一堆refer，要么单个sv单个refer
refer,audio_tensor = get_spepc(hps, path.name, dtype, device, is_v2pro)
try: #####这里加上提取sv的逻辑，要么一堆sv一堆refer，要么单个sv单个refer
refer, audio_tensor = get_spepc(hps, path.name, dtype, device, is_v2pro)
refers.append(refer)
if is_v2pro:
sv_emb.append(sv_cn_model.compute_embedding3(audio_tensor))
except Exception as e:
logger.error(e)
if len(refers) == 0:
refers,audio_tensor = get_spepc(hps, ref_wav_path, dtype, device, is_v2pro)
refers=[refers]
refers, audio_tensor = get_spepc(hps, ref_wav_path, dtype, device, is_v2pro)
refers = [refers]
if is_v2pro:
sv_emb=[sv_cn_model.compute_embedding3(audio_tensor)]
sv_emb = [sv_cn_model.compute_embedding3(audio_tensor)]
else:
refer,audio_tensor = get_spepc(hps, ref_wav_path, dtype, device)
refer, audio_tensor = get_spepc(hps, ref_wav_path, dtype, device)
t1 = ttime()
# os.environ['version'] = version
@ -937,14 +950,22 @@ def get_tts_wav(
if version not in {"v3", "v4"}:
if is_v2pro:
audio = (
vq_model.decode(pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refers, speed=speed,sv_emb=sv_emb)
vq_model.decode(
pred_semantic,
torch.LongTensor(phones2).to(device).unsqueeze(0),
refers,
speed=speed,
sv_emb=sv_emb,
)
.detach()
.cpu()
.numpy()[0, 0]
)
else:
audio = (
vq_model.decode(pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refers, speed=speed)
vq_model.decode(
pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refers, speed=speed
)
.detach()
.cpu()
.numpy()[0, 0]
@ -1108,7 +1129,6 @@ def handle(
if not default_refer.is_ready():
return JSONResponse({"code": 400, "message": "未指定参考音频且接口无预设"}, status_code=400)
if cut_punc == None:
text = cut_text(text, default_cut_punc)
else:

14
config.py
View File

@ -144,7 +144,8 @@ webui_port_subfix = 9871
api_port = 9880
#Thanks to the contribution of @Karasukaigan and @XXXXRT666
# Thanks to the contribution of @Karasukaigan and @XXXXRT666
def get_device_dtype_sm(idx: int) -> tuple[torch.device, torch.dtype, float, float]:
cpu = torch.device("cpu")
cuda = torch.device(f"cuda:{idx}")
@ -157,10 +158,13 @@ def get_device_dtype_sm(idx: int) -> tuple[torch.device, torch.dtype, float, flo
mem_gb = mem_bytes / (1024**3) + 0.4
major, minor = capability
sm_version = major + minor / 10.0
is_16_series = bool(re.search(r"16\d{2}", name))and sm_version == 7.5
if mem_gb < 4 or sm_version < 5.3:return cpu, torch.float32, 0.0, 0.0
if sm_version == 6.1 or is_16_series==True:return cuda, torch.float32, sm_version, mem_gb
if sm_version > 6.1:return cuda, torch.float16, sm_version, mem_gb
is_16_series = bool(re.search(r"16\d{2}", name)) and sm_version == 7.5
if mem_gb < 4 or sm_version < 5.3:
return cpu, torch.float32, 0.0, 0.0
if sm_version == 6.1 or is_16_series == True:
return cuda, torch.float32, sm_version, mem_gb
if sm_version > 6.1:
return cuda, torch.float16, sm_version, mem_gb
return cpu, torch.float32, 0.0, 0.0

73
docker_build.ps1 Normal file
View File

@ -0,0 +1,73 @@
$ErrorActionPreference = "Stop"
$ScriptDir = Split-Path -Parent $MyInvocation.MyCommand.Definition
Set-Location $ScriptDir
if (-not (Get-Command "docker" -ErrorAction SilentlyContinue)) {
Write-Host "Docker Not Found"
exit 1
}
$Lite = $false
$CudaVersion = "12.6"
function Write-Help {
Write-Host @"
Usage: powershell -File docker_build.ps1 [OPTIONS]
Options:
--cuda 12.6|12.8 Specify the CUDA VERSION (REQUIRED)
--lite Build a Lite Image
-h, --help Show this help message and exit
Examples:
powershell -File docker_build.ps1 --cuda 12.6 --lite
"@
}
if ($args.Count -eq 0) {
Write-Help
exit 0
}
for ($i = 0; $i -lt $args.Count; $i++) {
switch ($args[$i]) {
'--cuda' {
$i++
$val = $args[$i]
if ($val -ne "12.6" -and $val -ne "12.8") {
Write-Host "Error: Invalid CUDA_VERSION: $val"
Write-Host "Choose From: [12.6, 12.8]"
exit 1
}
$CudaVersion = $val
}
'--lite' {
$Lite = $true
}
'-h' { Write-Help; exit 0 }
'--help' { Write-Help; exit 0 }
default {
Write-Host "Unknown Argument: $($args[$i])"
Write-Host "Use -h or --help to see available options."
exit 1
}
}
}
$arch = (Get-CimInstance Win32_Processor).Architecture
$TargetPlatform = if ($arch -eq 9) { "linux/amd64" } else { "linux/arm64" }
if ($Lite) {
$TorchBase = "lite"
} else {
$TorchBase = "full"
}
docker build `
--build-arg CUDA_VERSION=$CudaVersion `
--build-arg LITE=$Lite `
--build-arg TARGETPLATFORM=$TargetPlatform `
--build-arg TORCH_BASE=$TorchBase `
-t "$env:USERNAME/gpt-sovits:local" `
.

2
docker_build.sh
View File

@ -25,7 +25,7 @@ print_help() {
echo " -h, --help Show this help message and exit"
echo ""
echo "Examples:"
echo " bash docker_build.sh --cuda 12.6 --funasr --faster-whisper"
echo " bash docker_build.sh --cuda 12.6"
}
# Show help if no arguments provided

10
docs/cn/Changelog_CN.md
View File

@ -157,7 +157,7 @@
- 类型: 修复
- 提交: StaryLan
- 2024.02.07 [Commit#14a28510](https://github.com/RVC-Boss/GPT-SoVITS/commit/14a285109a521679f8846589c22da8f656a46ad8)
- 内容: 修复 UVR5 `inf everywhere` 报错的问题 (`is_half` 传参未转换布尔类型导致恒定半精度推理, 16系显卡会 `inf`).
- 内容: 修复 UVR5 `inf everywhere` 报错的问题 (`is_half` 传参未转换布尔类型导致恒定半精度推理, 16 系显卡会 `inf`).
- 类型: 修复
- 提交: RVC-Boss
- 2024.02.07 [Commit#d74f888e](https://github.com/RVC-Boss/GPT-SoVITS/commit/d74f888e7ac86063bfeacef95d0e6ddafe42b3b2)
@ -230,6 +230,7 @@
## 202403
- 2024.03.06 [PR#675](https://github.com/RVC-Boss/GPT-SoVITS/pull/675)
- 内容: Faster Whisper 在没有 CUDA 可用时自动使用 CPU 推理.
- 类型: 优化
- 提交: ShiroDoMain
@ -347,7 +348,7 @@
- 2024.07.23 [Commit#9588a3c5](https://github.com/RVC-Boss/GPT-SoVITS/commit/9588a3c52d9ebdb20b3c5d74f647d12e7c1171c2), [PR#1340](https://github.com/RVC-Boss/GPT-SoVITS/pull/1340)
- 内容: 支持合成语速调节, 支持冻结随机性只调节语速, 并将其更新到`api.py` 上.
- 类型: 新功能
- 提交: RVC-Boss, 红血球AE3803
- 提交: RVC-Boss, 红血球 AE3803
- 2024.07.27 [PR#1306](https://github.com/RVC-Boss/GPT-SoVITS/pull/1306), [PR#1356](https://github.com/RVC-Boss/GPT-SoVITS/pull/1356)
- 内容: 增加 BS-Roformer 人声伴奏分离模型支持.
- 类型: 新功能
@ -409,7 +410,7 @@
- 2025.02.11 [Commit#ed207c4b](https://github.com/RVC-Boss/GPT-SoVITS/commit/ed207c4b879d5296e9be3ae5f7b876729a2c43b8)~[Commit#6e2b4918](https://github.com/RVC-Boss/GPT-SoVITS/commit/6e2b49186c5b961f0de41ea485d398dffa9787b4)
- 内容: **新增 GPT-SoVITS V3 模型, 需要 14G 显存进行微调.**
- 类型: 新功能 (特性参阅 [Wiki](https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)))
- 类型: 新功能 (特性参阅 [Wiki](<https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)>))
- 提交: RVC-Boss
- 2025.02.12 [PR#2032](https://github.com/RVC-Boss/GPT-SoVITS/pull/2032)
- 内容: 更新项目多语言文档.
@ -453,7 +454,7 @@
- 类型: 文档
- 提交: StaryLan
- 2025.02.28 [PR#2122](https://github.com/RVC-Boss/GPT-SoVITS/pull/2122)
- 内容: 对于模型无法判断的CJK短字符采用规则判断.
- 内容: 对于模型无法判断的 CJK 短字符采用规则判断.
- 类型: 修复
- 提交: KamioRinn
- 关联: [Issue#2116](https://github.com/RVC-Boss/GPT-SoVITS/issues/2116)
@ -475,6 +476,7 @@
- Pydantic: [Issue#2230](https://github.com/RVC-Boss/GPT-SoVITS/issues/2230), [Issue#2239](https://github.com/RVC-Boss/GPT-SoVITS/issues/2239).
- PyTorch-Lightning: [Issue#2174](https://github.com/RVC-Boss/GPT-SoVITS/issues/2174).
- 2025.03.31 [PR#2241](https://github.com/RVC-Boss/GPT-SoVITS/pull/2241)
- 内容: **为 SoVITS v3 适配并行推理**.
- 类型: 新功能
- 提交: ChasonJiang

50
docs/cn/README.md
View File

@ -1,18 +1,25 @@
#
<div align="center">
<h1>GPT-SoVITS-WebUI</h1>
强大的少样本语音转换与语音合成Web用户界面.<br><br>
强大的少样本语音转换与语音合成Web用户界面. <br><br>
[![madewithlove](https://img.shields.io/badge/made_with-%E2%9D%A4-red?style=for-the-badge&labelColor=orange)](https://github.com/RVC-Boss/GPT-SoVITS)
<a href="https://trendshift.io/repositories/7033" target="_blank"><img src="https://trendshift.io/api/badge/repositories/7033" alt="RVC-Boss%2FGPT-SoVITS | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
<!-- img src="https://counter.seku.su/cmoe?name=gptsovits&theme=r34" /><br> -->
[![Train In Colab](https://img.shields.io/badge/Colab-训练-F9AB00?style=for-the-badge&logo=googlecolab)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/Colab-WebUI.ipynb)
[![Infer In Colab](https://img.shields.io/badge/Colab-推理-F9AB00?style=for-the-badge&logo=googlecolab)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/Colab-Inference.ipynb)
[![Huggingface](https://img.shields.io/badge/HuggingFace-online%20demo-blue.svg?style=for-the-badge&logo=huggingface)](https://huggingface.co/spaces/lj1995/GPT-SoVITS-v2)
[![Open In Colab](https://img.shields.io/badge/Colab-F9AB00?style=for-the-badge&logo=googlecolab&color=525252)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/colab_webui.ipynb)
[![License](https://img.shields.io/badge/LICENSE-MIT-green.svg?style=for-the-badge)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/LICENSE)
[![Huggingface](https://img.shields.io/badge/🤗%20-online%20demo-yellow.svg?style=for-the-badge)](https://huggingface.co/spaces/lj1995/GPT-SoVITS-v2)
[![Discord](https://img.shields.io/discord/1198701940511617164?color=%23738ADB&label=Discord&style=for-the-badge)](https://discord.gg/dnrgs5GHfG)
[![GitHub release](https://img.shields.io/github/v/release/RVC-Boss/gpt-sovits?style=for-the-badge&logo=github)](https://github.com/RVC-Boss/gpt-sovits/releases)
[![Image Size](https://img.shields.io/docker/image-size/xxxxrt666/gpt-sovits/latest?style=for-the-badge&logo=docker)](https://hub.docker.com/r/xxxxrt666/gpt-sovits)
[![License](https://img.shields.io/badge/LICENSE-MIT-green.svg?style=for-the-badge&logo=opensourceinitiative)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/LICENSE)
[![简体中文](https://img.shields.io/badge/简体中文-阅读文档-blue?style=for-the-badge&logo=googledocs&logoColor=white)](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e)
[![English](https://img.shields.io/badge/English-Read%20Docs-blue?style=for-the-badge&logo=googledocs&logoColor=white)](https://rentry.co/GPT-SoVITS-guide#/)
[![Change Log](https://img.shields.io/badge/更新日志-查看更新-blue?style=for-the-badge&logo=googledocs&logoColor=white)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/docs/cn/Changelog_CN.md)
[**English**](../../README.md) | **中文简体** | [**日本語**](../ja/README.md) | [**한국어**](../ko/README.md) | [**Türkçe**](../tr/README.md)
@ -36,7 +43,7 @@
<https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-80c060ab47fb>
**用户手册: [简体中文](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e) | [English](https://rentry.co/GPT-SoVITS-guide#/)**
<!-- **用户手册: [简体中文](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e) | [English](https://rentry.co/GPT-SoVITS-guide#/)** -->
## 安装
@ -171,7 +178,7 @@ docker exec -it <GPT-SoVITS-CU126-Lite|GPT-SoVITS-CU128-Lite|GPT-SoVITS-CU126|GP
## 预训练模型
**若成功运行`install.sh`可跳过 No.1,2,3**
**若成功运行`install.sh`可跳过 No.1, 2, 3**
**中国地区的用户可以[在此处下载这些模型](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e/dkxgpiy9zb96hob4#nVNhX).**
@ -181,19 +188,19 @@ docker exec -it <GPT-SoVITS-CU126-Lite|GPT-SoVITS-CU128-Lite|GPT-SoVITS-CU126|GP
3. 对于 UVR5 (人声/伴奏分离和混响移除, 额外功能), 从 [UVR5 Weights](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/uvr5_weights) 下载模型, 并将其放置在 `tools/uvr5/uvr5_weights` 目录中.
- 如果你在 UVR5 中使用 `bs_roformer` 或 `mel_band_roformer`模型, 你可以手动下载模型和相应的配置文件, 并将它们放在 `tools/UVR5/UVR5_weights` 中.**重命名模型文件和配置文件, 确保除后缀外**, 模型和配置文件具有相同且对应的名称.此外, 模型和配置文件名**必须包含"roformer"**, 才能被识别为 roformer 类的模型.
- 如果你在 UVR5 中使用 `bs_roformer` 或 `mel_band_roformer`模型, 你可以手动下载模型和相应的配置文件, 并将它们放在 `tools/UVR5/UVR5_weights` 中. **重命名模型文件和配置文件, 确保除后缀外**, 模型和配置文件具有相同且对应的名称. 此外, 模型和配置文件名**必须包含"roformer"**, 才能被识别为 roformer 类的模型.
- 建议在模型名称和配置文件名中**直接指定模型类型**, 例如`mel_mand_roformer`、`bs_roformer`.如果未指定, 将从配置文中比对特征, 以确定它是哪种类型的模型.例如, 模型`bs_roformer_ep_368_sdr_12.9628.ckpt` 和对应的配置文件`bs_roformer_ep_368_sdr_12.9628.yaml` 是一对.`kim_mel_band_roformer.ckpt` 和 `kim_mel_band_roformer.yaml` 也是一对.
- 建议在模型名称和配置文件名中**直接指定模型类型**, 例如`mel_mand_roformer`、`bs_roformer`. 如果未指定, 将从配置文中比对特征, 以确定它是哪种类型的模型. 例如, 模型`bs_roformer_ep_368_sdr_12.9628.ckpt` 和对应的配置文件`bs_roformer_ep_368_sdr_12.9628.yaml` 是一对.`kim_mel_band_roformer.ckpt` 和 `kim_mel_band_roformer.yaml` 也是一对.
4. 对于中文 ASR (额外功能), 从 [Damo ASR Model](https://modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/files)、[Damo VAD Model](https://modelscope.cn/models/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/files) 和 [Damo Punc Model](https://modelscope.cn/models/damo/punc_ct-transformer_zh-cn-common-vocab272727-pytorch/files) 下载模型, 并将它们放置在 `tools/asr/models` 目录中.
5. 对于英语或日语 ASR (额外功能), 从 [Faster Whisper Large V3](https://huggingface.co/Systran/faster-whisper-large-v3) 下载模型, 并将其放置在 `tools/asr/models` 目录中.此外, [其他模型](https://huggingface.co/Systran) 可能具有类似效果且占用更少的磁盘空间.
5. 对于英语或日语 ASR (额外功能), 从 [Faster Whisper Large V3](https://huggingface.co/Systran/faster-whisper-large-v3) 下载模型, 并将其放置在 `tools/asr/models` 目录中. 此外, [其他模型](https://huggingface.co/Systran) 可能具有类似效果且占用更少的磁盘空间.
## 数据集格式
文本到语音 (TTS) 注释 .list 文件格式:
```
```text
vocal_path|speaker_name|language|text
```
@ -207,7 +214,7 @@ vocal_path|speaker_name|language|text
示例:
```
```text
D:\GPT-SoVITS\xxx/xxx.wav|xxx|zh|我爱玩原神.
```
@ -218,7 +225,6 @@ D:\GPT-SoVITS\xxx/xxx.wav|xxx|zh|我爱玩原神.
#### 整合包用户
双击`go-webui.bat`或者使用`go-webui.ps1`
若想使用 V1,则双击`go-webui-v1.bat`或者使用`go-webui-v1.ps1`
#### 其他
@ -226,14 +232,6 @@ D:\GPT-SoVITS\xxx/xxx.wav|xxx|zh|我爱玩原神.
python webui.py <language(optional)>
```
若想使用 V1,则
```bash
python webui.py v1 <language(optional)>
```
或者在 webUI 内动态切换
### 微调
#### 现已支持自动填充路径
@ -243,13 +241,13 @@ python webui.py v1 <language(optional)>
3. 进行降噪(可选)
4. 进行 ASR
5. 校对标注
6. 前往下一个窗口,点击训练
6. 前往下一个窗口, 点击训练
### 打开推理 WebUI
#### 整合包用户
双击 `go-webui.bat` 或者使用 `go-webui.ps1` ,然后在 `1-GPT-SoVITS-TTS/1C-推理` 中打开推理 webUI
双击 `go-webui.bat` 或者使用 `go-webui.ps1` , 然后在 `1-GPT-SoVITS-TTS/1C-推理` 中打开推理 webUI
#### 其他
@ -287,7 +285,7 @@ python webui.py
3. 需要从[huggingface](https://huggingface.co/lj1995/GPT-SoVITS/tree/main/gsv-v2final-pretrained) 下载预训练模型文件放到 GPT_SoVITS/pretrained_models/gsv-v2final-pretrained 下
中文额外需要下载[G2PWModel.zip(HF)](https://huggingface.co/XXXXRT/GPT-SoVITS-Pretrained/resolve/main/G2PWModel.zip)| [G2PWModel.zip(ModelScope)](https://www.modelscope.cn/models/XXXXRT/GPT-SoVITS-Pretrained/resolve/master/G2PWModel.zip) (下载 G2PW 模型,解压并重命名为`G2PWModel`,将其放到`GPT_SoVITS/text`目录下)
中文额外需要下载[G2PWModel.zip(HF)](https://huggingface.co/XXXXRT/GPT-SoVITS-Pretrained/resolve/main/G2PWModel.zip)| [G2PWModel.zip(ModelScope)](https://www.modelscope.cn/models/XXXXRT/GPT-SoVITS-Pretrained/resolve/master/G2PWModel.zip) (下载 G2PW 模型, 解压并重命名为`G2PWModel`, 将其放到`GPT_SoVITS/text`目录下)
## V3 更新说明
@ -449,5 +447,5 @@ python ./tools/asr/fasterwhisper_asr.py -i <input> -o <output> -l <language> -p
## 感谢所有贡献者的努力
<a href="https://github.com/RVC-Boss/GPT-SoVITS/graphs/contributors" target="_blank">
<img src="https://contrib.rocks/image?repo=RVC-Boss/GPT-SoVITS" />
<img src="https://contrib.rocks/image?repo=RVC-Boss/GPT-SoVITS" alt="Contributors"/>
</a>

9
docs/en/Changelog_EN.md
View File

@ -285,7 +285,7 @@
- 2024.06.06 [Commit#](https://github.com/RVC-Boss/GPT-SoVITS/commit/99f09c8bdc155c1f4272b511940717705509582a)
- Content: Fixed the issue of WebUI's GPT fine-tuning not reading BERT feature of Chinese input texts, causing inconsistency with inference and potential quality degradation.
**Caution: If you have previously fine-tuned with a large amount of data, it is recommended to retune the model to improve quality.**
**Caution: If you have previously fine-tuned with a large amount of data, it is recommended to retune the model to improve quality.**
- Type: Fix
- Contributor: RVC-Boss
- 2024.06.07 [PR#1159](https://github.com/RVC-Boss/GPT-SoVITS/pull/1159)
@ -347,7 +347,7 @@
- 2024.07.23 [Commit#9588a3c5](https://github.com/RVC-Boss/GPT-SoVITS/commit/9588a3c52d9ebdb20b3c5d74f647d12e7c1171c2), [PR#1340](https://github.com/RVC-Boss/GPT-SoVITS/pull/1340)
- Content: Support adjusting speech speed during synthesis, including an option to freeze randomness and only control speed. This feature has been updated to `api.py`.
- Type: Feature
- Contributor: RVC-Boss, 红血球AE3803
- Contributor: RVC-Boss, 红血球 AE3803
- 2024.07.27 [PR#1306](https://github.com/RVC-Boss/GPT-SoVITS/pull/1306), [PR#1356](https://github.com/RVC-Boss/GPT-SoVITS/pull/1356)
- Content: Added support for the BS-RoFormer vocal accompaniment separation model.
- Type: Feature
@ -409,7 +409,7 @@
- 2025.02.11 [Commit#ed207c4b](https://github.com/RVC-Boss/GPT-SoVITS/commit/ed207c4b879d5296e9be3ae5f7b876729a2c43b8)~[Commit#6e2b4918](https://github.com/RVC-Boss/GPT-SoVITS/commit/6e2b49186c5b961f0de41ea485d398dffa9787b4)
- Content: **Added GPT-SoVITS V3 model, which requires 14GB VRAM for fine-tuning.**
- Type: Feature (Refer to [Wiki](https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)))
- Type: Feature (Refer to [Wiki](<https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)>))
- Contributor: RVC-Boss
- 2025.02.12 [PR#2032](https://github.com/RVC-Boss/GPT-SoVITS/pull/2032)
- Content: Updated multilingual project documentation.
@ -478,9 +478,6 @@
- Content: **Enabled parallel inference for SoVITS v3.**
- Type: Feature
- Contributor: ChasonJiang
- Fixed other minor bugs.
- Integrated package fixes for ONNX runtime GPU inference support:
- Type: Fix
- Details:

179
docs/ja/Changelog_JA.md
View File

@ -3,11 +3,11 @@
## 202401
- 2024.01.21 [PR#108](https://github.com/RVC-Boss/GPT-SoVITS/pull/108)
- 内容: WebUIに英語システム翻訳サポートを追加。
- 内容: WebUI に英語システム翻訳サポートを追加。
- タイプ: ドキュメンテーション
- 貢献者: D3lik
- 2024.01.21 [Commit#7b89c9ed](https://github.com/RVC-Boss/GPT-SoVITS/commit/7b89c9ed5669f63c4ed6ae791408969640bdcf3e)
- 内容: SoVITSトレーニングのZeroDivisionError修正を試みた。
- 内容: SoVITS トレーニングの ZeroDivisionError 修正を試みた。
- タイプ: 修正
- 貢献者: RVC-Boss, Tybost
- 関連: [Issue#79](https://github.com/RVC-Boss/GPT-SoVITS/issues/79)
@ -16,11 +16,11 @@
- タイプ: 最適化
- 貢献者: RVC-Boss
- 2024.01.21 [Commit#a87ad522](https://github.com/RVC-Boss/GPT-SoVITS/commit/a87ad5228ed2d729da42019ae1b93171f6a745ef)
- 内容: `cmd-asr.py`がデフォルトディレクトリにFunASRモデルが含まれているか確認し、ない場合はModelScopeからダウンロードするようになった。
- 内容: `cmd-asr.py`がデフォルトディレクトリに FunASR モデルが含まれているか確認し、ない場合は ModelScope からダウンロードするようになった。
- タイプ: 機能
- 貢献者: RVC-Boss
- 2024.01.21 [Commit#f6147116](https://github.com/RVC-Boss/GPT-SoVITS/commit/f61471166c107ba56ccb7a5137fa9d7c09b2830d)
- 内容: `Config.py`に`is_share`パラメータを追加、`True`に設定するとWebUIを公開ネットワークにマッピング可能。
- 内容: `Config.py`に`is_share`パラメータを追加、`True`に設定すると WebUI を公開ネットワークにマッピング可能。
- タイプ: 機能
- 貢献者: RVC-Boss
- 2024.01.21 [Commit#102d5081](https://github.com/RVC-Boss/GPT-SoVITS/commit/102d50819e5d24580d6e96085b636b25533ecc7f)
@ -33,11 +33,11 @@
- 貢献者: RVC-Boss
- 2024.01.22 英語と日本語トレーニングのネイティブサポートをテスト（日本語トレーニングはルートディレクトリに非英語特殊文字がないことが必要）。
- 2024.01.22 [PR#124](https://github.com/RVC-Boss/GPT-SoVITS/pull/124)
- 内容: 音声パスチェックを改善。不正な入力パスから読み取ろうとすると、ffmpegエラーではなくパスが存在しないと報告するようになった。
- 内容: 音声パスチェックを改善。不正な入力パスから読み取ろうとすると、ffmpeg エラーではなくパスが存在しないと報告するようになった。
- タイプ: 最適化
- 貢献者: xmimu
- 2024.01.23 [Commit#93c47cd9](https://github.com/RVC-Boss/GPT-SoVITS/commit/93c47cd9f0c53439536eada18879b4ec5a812ae1)
- 内容: Hubert抽出がNaNエラーを引き起こし、SoVITS/GPTトレーニングでZeroDivisionErrorが発生する問題を解決。
- 内容: Hubert 抽出が NaN エラーを引き起こし、SoVITS/GPT トレーニングで ZeroDivisionError が発生する問題を解決。
- タイプ: 修正
- 貢献者: RVC-Boss
- 2024.01.23 [Commit#80fffb0a](https://github.com/RVC-Boss/GPT-SoVITS/commit/80fffb0ad46e4e7f27948d5a57c88cf342088d50)
@ -49,19 +49,19 @@
- タイプ: 最適化
- 貢献者: RVC-Boss
- 2024.01.23 [Commit#0c691191](https://github.com/RVC-Boss/GPT-SoVITS/commit/0c691191e894c15686e88279745712b3c6dc232f)
- 内容: 推論WebUIでクイックモデル切り替えをサポート追加。
- 内容: 推論 WebUI でクイックモデル切り替えをサポート追加。
- タイプ: 機能
- 貢献者: RVC-Boss
- 2024.01.25 [Commit#249561e5](https://github.com/RVC-Boss/GPT-SoVITS/commit/249561e5a18576010df6587c274d38cbd9e18b4b)
- 内容: 推論WebUIの冗長なログを削除。
- 内容: 推論 WebUI の冗長なログを削除。
- タイプ: 最適化
- 貢献者: RVC-Boss
- 2024.01.25 [PR#183](https://github.com/RVC-Boss/GPT-SoVITS/pull/183), [PR#200](https://github.com/RVC-Boss/GPT-SoVITS/pull/200)
- 内容: Macでのトレーニングと推論をサポート。
- 内容: Mac でのトレーニングと推論をサポート。
- タイプ: 機能
- 貢献者: Lion-Wu
- 2024.01.26 [Commit#813cf96e](https://github.com/RVC-Boss/GPT-SoVITS/commit/813cf96e508ba1bb2c658f38c7cc77b797fb4082), [Commit#2d1ddeca](https://github.com/RVC-Boss/GPT-SoVITS/commit/2d1ddeca42db90c3fe2d0cd79480fd544d87f02b)
- 内容: UVR5の読み取り時にディレクトリが自動的に飛び出す問題を修正。
- 内容: UVR5 の読み取り時にディレクトリが自動的に飛び出す問題を修正。
- タイプ: 修正
- 貢献者: RVC-Boss
- 2024.01.26 [PR#204](https://github.com/RVC-Boss/GPT-SoVITS/pull/204)
@ -77,11 +77,11 @@
- タイプ: 修正
- 貢献者: RVC-Boss
- 2024.01.26 [Commit#84ee4719](https://github.com/RVC-Boss/GPT-SoVITS/commit/84ee471936b332bc2ccee024d6dfdedab4f0dc7b)
- 内容: 半精度をサポートしないGPU向けに自動的に単精度を強制。CPU推論時も単精度を強制。
- 内容: 半精度をサポートしない GPU 向けに自動的に単精度を強制。CPU 推論時も単精度を強制。
- タイプ: 最適化
- 貢献者: RVC-Boss
- 2024.01.28 [PR#238](https://github.com/RVC-Boss/GPT-SoVITS/pull/238)
- 内容: Dockerfile内のモデルダウンロードプロセスを完了。
- 内容: Dockerfile 内のモデルダウンロードプロセスを完了。
- タイプ: 修正
- 貢献者: breakstring
- 2024.01.28 [PR#257](https://github.com/RVC-Boss/GPT-SoVITS/pull/257)
@ -89,7 +89,7 @@
- タイプ: 修正
- 貢献者: duliangang
- 2024.01.28 [Commit#f0cfe397](https://github.com/RVC-Boss/GPT-SoVITS/commit/f0cfe397089a6fd507d678c71adeaab5e7ed0683)
- 内容: GPTトレーニングがチェックポイントを保存しない問題を修正。
- 内容: GPT トレーニングがチェックポイントを保存しない問題を修正。
- タイプ: 修正
- 貢献者: RVC-Boss
- 2024.01.28 [Commit#b8ae5a27](https://github.com/RVC-Boss/GPT-SoVITS/commit/b8ae5a2761e2654fc0c905498009d3de9de745a8)
@ -101,15 +101,15 @@
- タイプ: 修正
- 貢献者: RVC-Boss
- 2024.01.29 [Commit#ff977a5f](https://github.com/RVC-Boss/GPT-SoVITS/commit/ff977a5f5dc547e0ad82b9e0f1cd95fbc830b2b0)
- 内容: 16シリーズのような半精度トレーニングに問題があるGPU向けに、トレーニング設定を単精度に変更。
- 内容: 16 シリーズのような半精度トレーニングに問題がある GPU 向けに、トレーニング設定を単精度に変更。
- タイプ: 修正
- 貢献者: RVC-Boss
- 2024.01.29 [Commit#172e139f](https://github.com/RVC-Boss/GPT-SoVITS/commit/172e139f45ac26723bc2cf7fac0112f69d6b46ec)
- 内容: 利用可能なColabバージョンをテストして更新。
- 内容: 利用可能な Colab バージョンをテストして更新。
- タイプ: 機能
- 貢献者: RVC-Boss
- 2024.01.29 [PR#135](https://github.com/RVC-Boss/GPT-SoVITS/pull/135)
- 内容: FunASRをバージョン1.0に更新し、インターフェース不一致によるエラーを修正。
- 内容: FunASR をバージョン 1.0 に更新し、インターフェース不一致によるエラーを修正。
- タイプ: 修正
- 貢献者: LauraGPT
- 2024.01.30 [Commit#1c2fa98c](https://github.com/RVC-Boss/GPT-SoVITS/commit/1c2fa98ca8c325dcfb32797d22ff1c2a726d1cb4)
@ -128,11 +128,11 @@
## 202402
- 2024.02.01 [Commit#45f73519](https://github.com/RVC-Boss/GPT-SoVITS/commit/45f73519cc41cd17cf816d8b997a9dcb0bee04b6)
- 内容: ASRパス末尾のスラッシュによるファイル名保存エラーの修正
- 内容: ASR パス末尾のスラッシュによるファイル名保存エラーの修正
- タイプ: 修正
- 貢献者: RVC-Boss
- 2024.02.03 [Commit#dba1a74c](https://github.com/RVC-Boss/GPT-SoVITS/commit/dba1a74ccb0cf19a1b4eb93faf11d4ec2b1fc5d7)
- 内容: UVR5のフォーマット読み取りエラーによる音声分離失敗の修正
- 内容: UVR5 のフォーマット読み取りエラーによる音声分離失敗の修正
- タイプ: 修正
- 貢献者: RVC-Boss
- 2024.02.03 [Commit#3ebff70b](https://github.com/RVC-Boss/GPT-SoVITS/commit/3ebff70b71580ee1f97b3238c9442cbc5aef47c7)
@ -140,7 +140,7 @@
- タイプ: 機能改善
- 貢献者: RVC-Boss
- 2024.02.03 [PR#377](https://github.com/RVC-Boss/GPT-SoVITS/pull/377)
- 内容: PaddleSpeechのテキスト正規化を導入(例: xx.xx%表記、「元/吨」を「元每吨」と正確に読む、アンダースコア問題の解消)
- 内容: PaddleSpeech のテキスト正規化を導入(例: xx.xx%表記、「元/吨」を「元每吨」と正確に読む、アンダースコア問題の解消)
- タイプ: 機能改善
- 貢献者: KamioRinn
- 2024.02.05 [PR#395](https://github.com/RVC-Boss/GPT-SoVITS/pull/395)
@ -153,36 +153,36 @@
- 貢献者: RVC-Boss
- 関連: [Issue#391](https://github.com/RVC-Boss/GPT-SoVITS/issues/391)
- 2024.02.06 [PR#403](https://github.com/RVC-Boss/GPT-SoVITS/pull/403)
- 内容: UVR5の高バージョンLibrosaへの対応
- 内容: UVR5 の高バージョン Librosa への対応
- タイプ: 修正
- 貢献者: StaryLan
- 2024.02.07 [Commit#14a28510](https://github.com/RVC-Boss/GPT-SoVITS/commit/14a285109a521679f8846589c22da8f656a46ad8)
- 内容: UVR5の「inf everywhere」エラーの修正(ブール型変換不足による半精度推論問題、16シリーズGPUで発生)
- 内容: UVR5 の「inf everywhere」エラーの修正(ブール型変換不足による半精度推論問題、16 シリーズ GPU で発生)
- タイプ: 修正
- 貢献者: RVC-Boss
- 2024.02.07 [Commit#d74f888e](https://github.com/RVC-Boss/GPT-SoVITS/commit/d74f888e7ac86063bfeacef95d0e6ddafe42b3b2)
- 内容: Gradio依存関係の修正
- 内容: Gradio 依存関係の修正
- タイプ: 修正
- 貢献者: RVC-Boss
- 2024.02.07 [PR#400](https://github.com/RVC-Boss/GPT-SoVITS/pull/400)
- 内容: Faster Whisperの統合による日本語・英語音声認識機能の追加
- 内容: Faster Whisper の統合による日本語・英語音声認識機能の追加
- タイプ: 新機能
- 貢献者: Shadow
- 2024.02.07 [Commit#6469048d](https://github.com/RVC-Boss/GPT-SoVITS/commit/6469048de12a8d6f0bd05d07f031309e61575a38)～[Commit#94ee71d9](https://github.com/RVC-Boss/GPT-SoVITS/commit/94ee71d9d562d10c9a1b96e745c6a6575aa66a10)
- 内容: 3連ルートディレクトリ空欄時の自動.listファイルパス読み込み機能
- 内容: 3 連ルートディレクトリ空欄時の自動.list ファイルパス読み込み機能
- タイプ: 機能改善
- 貢献者: RVC-Boss
- 2024.02.08 [Commit#59f35ada](https://github.com/RVC-Boss/GPT-SoVITS/commit/59f35adad85815df27e9c6b33d420f5ebfd8376b)
- 内容: GPTトレーニングのフリーズ問題(Windows10 1909)と繁体字システム言語時のエラー修正
- 内容: GPT トレーニングのフリーズ問題(Windows10 1909)と繁体字システム言語時のエラー修正
- タイプ: 修正
- 貢献者: RVC-Boss
- 関連: [Issue#232](https://github.com/RVC-Boss/GPT-SoVITS/issues/232)
- 2024.02.12 [PR#457](https://github.com/RVC-Boss/GPT-SoVITS/pull/457)
- 内容: DPO損失実験的トレーニングオプションの追加(ネガティブサンプル構築によるGPTの繰り返し・文字抜け問題改善)、推論インターフェースの複数パラメータ公開
- 内容: DPO 損失実験的トレーニングオプションの追加(ネガティブサンプル構築による GPT の繰り返し・文字抜け問題改善)、推論インターフェースの複数パラメータ公開
- タイプ: 新機能
- 貢献者: liufenghua
- 2024.02.12 [Commit#2fa74ecb](https://github.com/RVC-Boss/GPT-SoVITS/commit/2fa74ecb941db27d9015583a9be6962898d66730), [Commit#d82f6bbb](https://github.com/RVC-Boss/GPT-SoVITS/commit/d82f6bbb98ba725e6725dcee99b80ce71fb0bf28)
- 内容: 音声認識ロジックの最適化、Faster Whisperのミラーサイトダウンロード対応(HuggingFace接続問題回避)
- 内容: 音声認識ロジックの最適化、Faster Whisper のミラーサイトダウンロード対応(HuggingFace 接続問題回避)
- タイプ: 機能改善
- 貢献者: RVC-Boss
- 2024.02.15 [Commit#dd2c4d6d](https://github.com/RVC-Boss/GPT-SoVITS/commit/dd2c4d6d7121bf82d29d0f0e4d788f3b231997c8)
@ -190,7 +190,7 @@
- タイプ: 修正
- 貢献者: RVC-Boss
- 2024.02.15 [Commit#ccb9b08b](https://github.com/RVC-Boss/GPT-SoVITS/commit/ccb9b08be3c58e102defcc94ff4fd609da9e27ee)～[Commit#895fde46](https://github.com/RVC-Boss/GPT-SoVITS/commit/895fde46e420040ed26aaf0c5b7e99359d9b199b)
- 内容: DPOトレーニングを必須からオプションに変更(選択時は自動的にバッチサイズ半減)、推論インターフェースの新パラメータ未伝達問題の修正
- 内容: DPO トレーニングを必須からオプションに変更(選択時は自動的にバッチサイズ半減)、推論インターフェースの新パラメータ未伝達問題の修正
- タイプ: 機能改善
- 貢献者: RVC-Boss
- 2024.02.15 [Commit#7b0c3c67](https://github.com/RVC-Boss/GPT-SoVITS/commit/7b0c3c676495c64b2064aa472bff14b5c06206a5)
@ -207,38 +207,38 @@
- タイプ: 機能改善
- 貢献者: KamioRinn, v3cun
- 2024.02.17 [PR#510](https://github.com/RVC-Boss/GPT-SoVITS/pull/511), [PR#511](https://github.com/RVC-Boss/GPT-SoVITS/pull/511)
- 内容: Colabの公開URL未開始問題の修正
- 内容: Colab の公開 URL 未開始問題の修正
- タイプ: 修正
- 貢献者: ChanningWang2018, RVC-Boss
- 2024.02.21 [PR#557](https://github.com/RVC-Boss/GPT-SoVITS/pull/557)
- 内容: MacOS推論デバイスをMPSからCPUに変更(CPU推論の方が高速)
- 内容: MacOS 推論デバイスを MPS から CPU に変更(CPU 推論の方が高速)
- タイプ: 機能改善
- 貢献者: XXXXRT666
- 2024.02.21 [Commit#6da486c1](https://github.com/RVC-Boss/GPT-SoVITS/commit/6da486c15d09e3d99fa42c5e560aaac56b6b4ce1), [Commit#5a171773](https://github.com/RVC-Boss/GPT-SoVITS/commit/5a17177342d2df1e11369f2f4f58d34a3feb1a35)
- 内容: データ前処理に音声ノイズ除去オプション追加(16Kサンプルレートにダウンサンプリング、高ノイズ時以外は非推奨)
- 内容: データ前処理に音声ノイズ除去オプション追加(16K サンプルレートにダウンサンプリング、高ノイズ時以外は非推奨)
- タイプ: 新機能
- 貢献者: RVC-Boss
- 2024.02.28 [PR#573](https://github.com/RVC-Boss/GPT-SoVITS/pull/573)
- 内容: is_half判定の修正によるMacOSの正常なCPU推論対応
- 内容: is_half 判定の修正による MacOS の正常な CPU 推論対応
- タイプ: 修正
- 貢献者: XXXXRT666
- 2024.02.28 [PR#610](https://github.com/RVC-Boss/GPT-SoVITS/pull/610)
- 内容: UVR5 MDXNetパラメータ順序エラーによる出力フォルダ逆転問題の修正
- 内容: UVR5 MDXNet パラメータ順序エラーによる出力フォルダ逆転問題の修正
- タイプ: 修正
- 貢献者: Yuze Wang
## 202403
- 2024.03.06 [PR#675](https://github.com/RVC-Boss/GPT-SoVITS/pull/675)
- 内容: CUDAが利用できない場合、Faster Whisperの自動CPU推論を有効化
- 内容: CUDA が利用できない場合、Faster Whisper の自動 CPU 推論を有効化
- タイプ: 機能改善
- 貢献者: ShiroDoMain
- 2024.03.06 [Commit#616be20d](https://github.com/RVC-Boss/GPT-SoVITS/commit/616be20db3cf94f1cd663782fea61b2370704193)
- 内容: Faster Whisper非中国語ASR使用時、中国語FunASRモデルの事前ダウンロードが不要に
- 内容: Faster Whisper 非中国語 ASR 使用時、中国語 FunASR モデルの事前ダウンロードが不要に
- タイプ: 機能改善
- 貢献者: RVC-Boss
- 2024.03.09 [PR#672](https://github.com/RVC-Boss/GPT-SoVITS/pull/672)
- 内容: 推論速度を50%向上（RTX3090 + PyTorch 2.2.1 + CU11.8 + Win10 + Py39環境で検証）
- 内容: 推論速度を 50%向上（RTX3090 + PyTorch 2.2.1 + CU11.8 + Win10 + Py39 環境で検証）
- タイプ: 機能改善
- 貢献者: GoHomeToMacDonal
- 2024.03.10 [PR#721](https://github.com/RVC-Boss/GPT-SoVITS/pull/721)
@ -246,7 +246,7 @@
- タイプ: 新機能
- 貢献者: ChasonJiang
- 2024.03.13 [PR#761](https://github.com/RVC-Boss/GPT-SoVITS/pull/761)
- 内容: CPUトレーニングをサポート（macOSでCPUを使用したトレーニングが可能に）
- 内容: CPU トレーニングをサポート（macOS で CPU を使用したトレーニングが可能に）
- タイプ: 新機能
- 貢献者: Lion-Wu
- 2024.03.19 [PR#804](https://github.com/RVC-Boss/GPT-SoVITS/pull/804), [PR#812](https://github.com/RVC-Boss/GPT-SoVITS/pull/812), [PR#821](https://github.com/RVC-Boss/GPT-SoVITS/pull/821)
@ -254,21 +254,21 @@
- タイプ: 機能改善
- 貢献者: KamioRinn
- 2024.03.30 [PR#894](https://github.com/RVC-Boss/GPT-SoVITS/pull/894)
- 内容: APIフォーマットの改善
- 内容: API フォーマットの改善
- タイプ: 機能改善
- 貢献者: KamioRinn
## 202404
- 2024.04.03 [PR#917](https://github.com/RVC-Boss/GPT-SoVITS/pull/917)
- 内容: UVR5 WebUIにおけるFFmpegコマンド文字列フォーマットの修正
- 内容: UVR5 WebUI における FFmpeg コマンド文字列フォーマットの修正
- タイプ: 修正
- 貢献者: StaryLan
## 202405
- 2024.05.02 [PR#953](https://github.com/RVC-Boss/GPT-SoVITS/pull/953)
- 内容: SoVITSトレーニング時のVQ凍結漏れ問題を修正（品質劣化の原因となる）
- 内容: SoVITS トレーニング時の VQ 凍結漏れ問題を修正（品質劣化の原因となる）
- タイプ: 修正
- 貢献者: hcwu1993
- 関連: [Issue#747](https://github.com/RVC-Boss/GPT-SoVITS/issues/747)
@ -277,23 +277,23 @@
- タイプ: 機能改善
- 貢献者: StaryLan
- 2024.05.27 [PR#1132](https://github.com/RVC-Boss/GPT-SoVITS/pull/1132)
- 内容: Hubert抽出におけるバグ修正
- 内容: Hubert 抽出におけるバグ修正
- タイプ: 修正
- 貢献者: XXXXRT666
## 202406
- 2024.06.06 [Commit#](https://github.com/RVC-Boss/GPT-SoVITS/commit/99f09c8bdc155c1f4272b511940717705509582a)
- 内容: WebUIのGPTファインチューニング時に中国語入力テキストのBERT特徴量を読み取れない問題を修正（推論時との不一致や品質劣化の原因となる）
**注意: 既に大量データでファインチューニング済みの場合は、品質向上のためモデルの再チューニングを推奨**
- 内容: WebUI の GPT ファインチューニング時に中国語入力テキストの BERT 特徴量を読み取れない問題を修正（推論時との不一致や品質劣化の原因となる）
**注意: 既に大量データでファインチューニング済みの場合は、品質向上のためモデルの再チューニングを推奨**
- タイプ: 修正
- 貢献者: RVC-Boss
- 2024.06.07 [PR#1159](https://github.com/RVC-Boss/GPT-SoVITS/pull/1159)
- 内容: `s2_train.py`におけるSoVITSトレーニングの進捗バー処理を修正
- 内容: `s2_train.py`における SoVITS トレーニングの進捗バー処理を修正
- タイプ: 修正
- 貢献者: pengzhendong
- 2024.06.10 [Commit#501a74ae](https://github.com/RVC-Boss/GPT-SoVITS/commit/501a74ae96789a26b48932babed5eb4e9483a232)
- 内容: UVR5 MDXNetがFFmpegを呼び出す際の文字列フォーマットを修正（スペースを含むパスに対応）
- 内容: UVR5 MDXNet が FFmpeg を呼び出す際の文字列フォーマットを修正（スペースを含むパスに対応）
- タイプ: 修正
- 貢献者: RVC-Boss
- 2024.06.10 [PR#1168](https://github.com/RVC-Boss/GPT-SoVITS/pull/1168), [PR#1169](https://github.com/RVC-Boss/GPT-SoVITS/pull/1169)
@ -302,15 +302,15 @@
- 貢献者: XXXXRT666
- 関連: [Issue#1165](https://github.com/RVC-Boss/GPT-SoVITS/issues/1165)
- 2024.06.13 [Commit#db506705](https://github.com/RVC-Boss/GPT-SoVITS/commit/db50670598f0236613eefa6f2d5a23a271d82041)
- 内容: CPU推論におけるデフォルトバッチサイズの小数点問題を修正
- 内容: CPU 推論におけるデフォルトバッチサイズの小数点問題を修正
- タイプ: 修正
- 貢献者: RVC-Boss
- 2024.06.28 [PR#1258](https://github.com/RVC-Boss/GPT-SoVITS/pull/1258), [PR#1265](https://github.com/RVC-Boss/GPT-SoVITS/pull/1265), [PR#1267](https://github.com/RVC-Boss/GPT-SoVITS/pull/1267)
- 内容: ノイズ除去やASRで例外が発生した場合に保留中の全オーディオファイル処理が終了してしまう問題を修正
- 内容: ノイズ除去や ASR で例外が発生した場合に保留中の全オーディオファイル処理が終了してしまう問題を修正
- タイプ: 修正
- 貢献者: XXXXRT666
- 2024.06.29 [Commit#a208698e](https://github.com/RVC-Boss/GPT-SoVITS/commit/a208698e775155efc95b187b746d153d0f2847ca)
- 内容: マルチGPUトレーニング時のマルチプロセス保存ロジックを修正
- 内容: マルチ GPU トレーニング時のマルチプロセス保存ロジックを修正
- タイプ: 修正
- 貢献者: RVC-Boss
- 2024.06.29 [PR#1251](https://github.com/RVC-Boss/GPT-SoVITS/pull/1251)
@ -326,14 +326,14 @@
- タイプ: 修正
- 貢献者: aoguai
- 2024.07.06 [Commit#](https://github.com/RVC-Boss/GPT-SoVITS/commit/b0786f2998f1b2fce6678434524b4e0e8cc716f5)
- 内容: 高速化推論コードが検証済みでmainブランチにマージされ、ベースと同等の推論効果を保証。テキスト未参照モードでも高速推論をサポート
- 内容: 高速化推論コードが検証済みで main ブランチにマージされ、ベースと同等の推論効果を保証。テキスト未参照モードでも高速推論をサポート
- タイプ: 最適化
- 貢献者: RVC-Boss, GoHomeToMacDonal
- 関連: [PR#672](https://github.com/RVC-Boss/GPT-SoVITS/pull/672)
- 今後も`fast_inference`ブランチでの変更整合性を継続検証
- 2024.07.13 [PR#1294](https://github.com/RVC-Boss/GPT-SoVITS/pull/1294), [PR#1298](https://github.com/RVC-Boss/GPT-SoVITS/pull/1298)
- 内容: i18nスキャンのリファクタリングと多言語設定ファイルの更新
- 内容: i18n スキャンのリファクタリングと多言語設定ファイルの更新
- タイプ: ドキュメンテーション
- 貢献者: StaryLan
- 2024.07.13 [PR#1299](https://github.com/RVC-Boss/GPT-SoVITS/pull/1299)
@ -341,15 +341,15 @@
- タイプ: 修正
- 貢献者: XXXXRT666
- 2024.07.19 [PR#756](https://github.com/RVC-Boss/GPT-SoVITS/pull/756)
- 内容: GPTトレーニング時、カスタム`bucket_sampler`使用時のステップ数不一致を修正
- 内容: GPT トレーニング時、カスタム`bucket_sampler`使用時のステップ数不一致を修正
- タイプ: 修正
- 貢献者: huangxu1991
- 2024.07.23 [Commit#9588a3c5](https://github.com/RVC-Boss/GPT-SoVITS/commit/9588a3c52d9ebdb20b3c5d74f647d12e7c1171c2), [PR#1340](https://github.com/RVC-Boss/GPT-SoVITS/pull/1340)
- 内容: 合成時の話速調整をサポート（ランダム性を固定して速度のみ制御するオプション含む）。`api.py`に更新済み
- タイプ: 新機能
- 貢献者: RVC-Boss, 红血球AE3803
- 貢献者: RVC-Boss, 红血球 AE3803
- 2024.07.27 [PR#1306](https://github.com/RVC-Boss/GPT-SoVITS/pull/1306), [PR#1356](https://github.com/RVC-Boss/GPT-SoVITS/pull/1356)
- 内容: BS-RoFormerボーカル・伴奏分離モデルのサポートを追加。
- 内容: BS-RoFormer ボーカル・伴奏分離モデルのサポートを追加。
- タイプ: 新機能
- 貢献者: KamioRinn
- 2024.07.27 [PR#1351](https://github.com/RVC-Boss/GPT-SoVITS/pull/1351)
@ -360,56 +360,56 @@
## 202408 (V2 バージョン)
- 2024.08.01 [PR#1355](https://github.com/RVC-Boss/GPT-SoVITS/pull/1355)
- 内容: WebUIでファイル処理時にパスを自動入力するように変更。
- 内容: WebUI でファイル処理時にパスを自動入力するように変更。
- タイプ: 雑務
- 貢献者: XXXXRT666
- 2024.08.01 [Commit#e62e9653](https://github.com/RVC-Boss/GPT-SoVITS/commit/e62e965323a60a76a025bcaa45268c1ddcbcf05c)
- 内容: BS-RoformerのFP16推論サポートを有効化。
- 内容: BS-Roformer の FP16 推論サポートを有効化。
- タイプ: パフォーマンス最適化
- 貢献者: RVC-Boss
- 2024.08.01 [Commit#bce451a2](https://github.com/RVC-Boss/GPT-SoVITS/commit/bce451a2d1641e581e200297d01f219aeaaf7299), [Commit#4c8b7612](https://github.com/RVC-Boss/GPT-SoVITS/commit/4c8b7612206536b8b4435997acb69b25d93acb78)
- 内容: GPU認識ロジックを最適化、ユーザーが入力した任意のGPUインデックスを処理するユーザーフレンドリーなロジックを追加。
- 内容: GPU 認識ロジックを最適化、ユーザーが入力した任意の GPU インデックスを処理するユーザーフレンドリーなロジックを追加。
- タイプ: 雑務
- 貢献者: RVC-Boss
- 2024.08.02 [Commit#ff6c193f](https://github.com/RVC-Boss/GPT-SoVITS/commit/ff6c193f6fb99d44eea3648d82ebcee895860a22)~[Commit#de7ee7c7](https://github.com/RVC-Boss/GPT-SoVITS/commit/de7ee7c7c15a2ec137feb0693b4ff3db61fad758)
- 内容: **GPT-SoVITS V2モデルを追加。**
- 内容: **GPT-SoVITS V2 モデルを追加。**
- タイプ: 新機能
- 貢献者: RVC-Boss
- 2024.08.03 [Commit#8a101474](https://github.com/RVC-Boss/GPT-SoVITS/commit/8a101474b5a4f913b4c94fca2e3ca87d0771bae3)
- 内容: FunASRを使用して広東語ASRをサポート。
- 内容: FunASR を使用して広東語 ASR をサポート。
- タイプ: 新機能
- 貢献者: RVC-Boss
- 2024.08.03 [PR#1387](https://github.com/RVC-Boss/GPT-SoVITS/pull/1387), [PR#1388](https://github.com/RVC-Boss/GPT-SoVITS/pull/1388)
- 内容: UIとタイミングロジックを最適化。
- 内容: UI とタイミングロジックを最適化。
- タイプ: 雑務
- 貢献者: XXXXRT666
- 2024.08.06 [PR#1404](https://github.com/RVC-Boss/GPT-SoVITS/pull/1404), [PR#987](https://github.com/RVC-Boss/GPT-SoVITS/pull/987), [PR#488](https://github.com/RVC-Boss/GPT-SoVITS/pull/488)
- 内容: 多音字処理ロジックを最適化（V2のみ）。
- 内容: 多音字処理ロジックを最適化（V2 のみ）。
- タイプ: 修正、新機能
- 貢献者: KamioRinn、RVC-Boss
- 2024.08.13 [PR#1422](https://github.com/RVC-Boss/GPT-SoVITS/pull/1422)
- 内容: 参照音声が1つしかアップロードできないバグを修正。欠損ファイルがある場合に警告ポップアップを表示するデータセット検証を追加。
- 内容: 参照音声が 1 つしかアップロードできないバグを修正。欠損ファイルがある場合に警告ポップアップを表示するデータセット検証を追加。
- タイプ: 修正、雑務
- 貢献者: XXXXRT666
- 2024.08.20 [Issue#1508](https://github.com/RVC-Boss/GPT-SoVITS/issues/1508)
- 内容: 上流のLangSegmentライブラリがSSMLタグを使用した数字、電話番号、日付、時刻の最適化をサポート。
- 内容: 上流の LangSegment ライブラリが SSML タグを使用した数字、電話番号、日付、時刻の最適化をサポート。
- タイプ: 新機能
- 貢献者: juntaosun
- 2024.08.20 [PR#1503](https://github.com/RVC-Boss/GPT-SoVITS/pull/1503)
- 内容: APIを修正・最適化。
- 内容: API を修正・最適化。
- タイプ: 修正
- 貢献者: KamioRinn
- 2024.08.20 [PR#1490](https://github.com/RVC-Boss/GPT-SoVITS/pull/1490)
- 内容: `fast_inference`ブランチをメインブランチにマージ。
- タイプ: リファクタリング
- 貢献者: ChasonJiang
- 2024.08.21 **GPT-SoVITS V2バージョンを正式リリース。**
- 2024.08.21 **GPT-SoVITS V2 バージョンを正式リリース。**
## 202502 (V3 バージョン)
- 2025.02.11 [Commit#ed207c4b](https://github.com/RVC-Boss/GPT-SoVITS/commit/ed207c4b879d5296e9be3ae5f7b876729a2c43b8)~[Commit#6e2b4918](https://github.com/RVC-Boss/GPT-SoVITS/commit/6e2b49186c5b961f0de41ea485d398dffa9787b4)
- 内容: **GPT-SoVITS V3モデルを追加。ファインチューニングには14GBのVRAMが必要。**
- タイプ: 新機能（[Wiki](https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7))参照）
- 内容: **GPT-SoVITS V3 モデルを追加。ファインチューニングには 14GB の VRAM が必要。**
- タイプ: 新機能（[Wiki](<https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)>)参照）
- 貢献者: RVC-Boss
- 2025.02.12 [PR#2032](https://github.com/RVC-Boss/GPT-SoVITS/pull/2032)
- 内容: 多言語プロジェクトドキュメントを更新。
@ -424,7 +424,7 @@
- タイプ: パフォーマンス最適化
- 貢献者: wzy3650
- 2025.02.12 [PR#2040](https://github.com/RVC-Boss/GPT-SoVITS/pull/2040)
- 内容: ファインチューニング用に勾配チェックポイントサポートを追加。12GB VRAMが必要。
- 内容: ファインチューニング用に勾配チェックポイントサポートを追加。12GB VRAM が必要。
- タイプ: 新機能
- 貢献者: Kakaru Hayate
- 2025.02.14 [PR#2047](https://github.com/RVC-Boss/GPT-SoVITS/pull/2047), [PR#2062](https://github.com/RVC-Boss/GPT-SoVITS/pull/2062), [PR#2073](https://github.com/RVC-Boss/GPT-SoVITS/pull/2073)
@ -432,19 +432,19 @@
- タイプ: 新機能
- 貢献者: KamioRinn
- 2025.02.23 [Commit#56509a17](https://github.com/RVC-Boss/GPT-SoVITS/commit/56509a17c918c8d149c48413a672b8ddf437495b)~[Commit#514fb692](https://github.com/RVC-Boss/GPT-SoVITS/commit/514fb692db056a06ed012bc3a5bca2a5b455703e)
- 内容: **GPT-SoVITS V3モデルがLoRAトレーニングをサポート。ファインチューニングに8GB GPUメモリが必要。**
- 内容: **GPT-SoVITS V3 モデルが LoRA トレーニングをサポート。ファインチューニングに 8GB GPU メモリが必要。**
- タイプ: 新機能
- 貢献者: RVC-Boss
- 2025.02.23 [PR#2078](https://github.com/RVC-Boss/GPT-SoVITS/pull/2078)
- 内容: ボーカルと楽器分離のためのMel Band Roformerモデルサポートを追加。
- 内容: ボーカルと楽器分離のための Mel Band Roformer モデルサポートを追加。
- タイプ: 新機能
- 貢献者: Sucial
- 2025.02.26 [PR#2112](https://github.com/RVC-Boss/GPT-SoVITS/pull/2112), [PR#2114](https://github.com/RVC-Boss/GPT-SoVITS/pull/2114)
- 内容: 中国語パス下でのMeCabエラーを修正（日本語/韓国語または多言語テキスト分割用）。
- 内容: 中国語パス下での MeCab エラーを修正（日本語/韓国語または多言語テキスト分割用）。
- タイプ: 修正
- 貢献者: KamioRinn
- 2025.02.27 [Commit#92961c3f](https://github.com/RVC-Boss/GPT-SoVITS/commit/92961c3f68b96009ff2cd00ce614a11b6c4d026f)~[Commit#250b1c73](https://github.com/RVC-Boss/GPT-SoVITS/commit/250b1c73cba60db18148b21ec5fbce01fd9d19bc)
- 内容: **24kHzから48kHzへのオーディオ超解像モデルを追加**。V3モデルで24Kオーディオを生成する際の「こもった」オーディオ問題を緩和。
- 内容: **24kHz から 48kHz へのオーディオ超解像モデルを追加**。V3 モデルで 24K オーディオを生成する際の「こもった」オーディオ問題を緩和。
- タイプ: 新機能
- 貢献者: RVC-Boss
- 関連: [Issue#2085](https://github.com/RVC-Boss/GPT-SoVITS/issues/2085), [Issue#2117](https://github.com/RVC-Boss/GPT-SoVITS/issues/2117)
@ -453,7 +453,7 @@
- タイプ: ドキュメント
- 貢献者: StaryLan
- 2025.02.28 [PR#2122](https://github.com/RVC-Boss/GPT-SoVITS/pull/2122)
- 内容: モデルが識別できない短いCJK文字に対してルールベースの検出を適用。
- 内容: モデルが識別できない短い CJK 文字に対してルールベースの検出を適用。
- タイプ: 修正
- 貢献者: KamioRinn
- 関連: [Issue#2116](https://github.com/RVC-Boss/GPT-SoVITS/issues/2116)
@ -461,7 +461,7 @@
- 内容: 合成速度を制御するための発話速度パラメータを追加。
- タイプ: 修正
- 貢献者: RVC-Boss
- 2025.02.28 **GPT-SoVITS V3を正式リリース**。
- 2025.02.28 **GPT-SoVITS V3 を正式リリース**。
## 202503
@ -475,30 +475,31 @@
- Pydantic: [Issue#2230](https://github.com/RVC-Boss/GPT-SoVITS/issues/2230), [Issue#2239](https://github.com/RVC-Boss/GPT-SoVITS/issues/2239).
- PyTorch-Lightning: [Issue#2174](https://github.com/RVC-Boss/GPT-SoVITS/issues/2174).
- 2025.03.31 [PR#2241](https://github.com/RVC-Boss/GPT-SoVITS/pull/2241)
- 内容: **SoVITS v3の並列推論を有効化。**
- 内容: **SoVITS v3 の並列推論を有効化。**
- タイプ: 新機能
- 貢献者: ChasonJiang
- その他の軽微なバグを修正。
- ONNXランタイムGPU推論サポートのための統合パッケージ修正:
- ONNX ランタイム GPU 推論サポートのための統合パッケージ修正:
- タイプ: 修正
- 詳細:
- G2PW内のONNXモデルをCPUからGPU推論に切り替え、CPUボトルネックを大幅に削減;
- foxjoy dereverberationモデルがGPU推論をサポート。
- G2PW 内の ONNX モデルを CPU から GPU 推論に切り替え、CPU ボトルネックを大幅に削減;
- foxjoy dereverberation モデルが GPU 推論をサポート。
## 202504 (V4 バージョン)
- 2025.04.01 [Commit#6a60e5ed](https://github.com/RVC-Boss/GPT-SoVITS/commit/6a60e5edb1817af4a61c7a5b196c0d0f1407668f)
- 内容: SoVITS v3並列推論のロックを解除。非同期モデル読み込みロジックを修正。
- 内容: SoVITS v3 並列推論のロックを解除。非同期モデル読み込みロジックを修正。
- タイプ: 修正
- 貢献者: RVC-Boss
- 2025.04.07 [PR#2255](https://github.com/RVC-Boss/GPT-SoVITS/pull/2255)
- 内容: Ruffを使用したコードフォーマット。G2PWリンクを更新。
- 内容: Ruff を使用したコードフォーマット。G2PW リンクを更新。
- タイプ: スタイル
- 貢献者: XXXXRT666
- 2025.04.15 [PR#2290](https://github.com/RVC-Boss/GPT-SoVITS/pull/2290)
- 内容: ドキュメントを整理。Python 3.11サポートを追加。インストーラーを更新。
- 内容: ドキュメントを整理。Python 3.11 サポートを追加。インストーラーを更新。
- タイプ: 雑務
- 貢献者: XXXXRT666
- 2025.04.20 [PR#2300](https://github.com/RVC-Boss/GPT-SoVITS/pull/2300)
@ -506,11 +507,11 @@
- タイプ: 雑務
- 貢献者: XXXXRT666
- 2025.04.20 [Commit#e0c452f0](https://github.com/RVC-Boss/GPT-SoVITS/commit/e0c452f0078e8f7eb560b79a54d75573fefa8355)~[Commit#9d481da6](https://github.com/RVC-Boss/GPT-SoVITS/commit/9d481da610aa4b0ef8abf5651fd62800d2b4e8bf)
- 内容: **GPT-SoVITS V4モデルを追加。**
- 内容: **GPT-SoVITS V4 モデルを追加。**
- タイプ: 新機能
- 貢献者: RVC-Boss
- 2025.04.21 [Commit#8b394a15](https://github.com/RVC-Boss/GPT-SoVITS/commit/8b394a15bce8e1d85c0b11172442dbe7a6017ca2)~[Commit#bc2fe5ec](https://github.com/RVC-Boss/GPT-SoVITS/commit/bc2fe5ec86536c77bb3794b4be263ac87e4fdae6), [PR#2307](https://github.com/RVC-Boss/GPT-SoVITS/pull/2307)
- 内容: V4の並列推論を有効化。
- 内容: V4 の並列推論を有効化。
- タイプ: 新機能
- 貢献者: RVC-Boss、ChasonJiang
- 2025.04.22 [Commit#7405427a](https://github.com/RVC-Boss/GPT-SoVITS/commit/7405427a0ab2a43af63205df401fd6607a408d87)~[Commit#590c83d7](https://github.com/RVC-Boss/GPT-SoVITS/commit/590c83d7667c8d4908f5bdaf2f4c1ba8959d29ff), [PR#2309](https://github.com/RVC-Boss/GPT-SoVITS/pull/2309)
@ -518,24 +519,24 @@
- タイプ: 修正
- 貢献者: RVC-Boss、ChasonJiang
- 2025.04.22 [Commit#fbdab94e](https://github.com/RVC-Boss/GPT-SoVITS/commit/fbdab94e17d605d85841af6f94f40a45976dd1d9), [PR#2310](https://github.com/RVC-Boss/GPT-SoVITS/pull/2310)
- 内容: NumpyとNumbaのバージョン不一致問題を修正。librosaバージョンを更新。
- 内容: Numpy と Numba のバージョン不一致問題を修正。librosa バージョンを更新。
- タイプ: 修正
- 貢献者: RVC-Boss、XXXXRT666
- 関連: [Issue#2308](https://github.com/RVC-Boss/GPT-SoVITS/issues/2308)
- **2024.04.22 GPT-SoVITS V4を正式リリース**。
- **2024.04.22 GPT-SoVITS V4 を正式リリース**。
- 2025.04.22 [PR#2311](https://github.com/RVC-Boss/GPT-SoVITS/pull/2311)
- 内容: Gradioパラメータを更新。
- 内容: Gradio パラメータを更新。
- タイプ: 雑務
- 貢献者: XXXXRT666
- 2025.04.25 [PR#2322](https://github.com/RVC-Boss/GPT-SoVITS/pull/2322)
- 内容: Colab/Kaggleノートブックスクリプトを改善。
- 内容: Colab/Kaggle ノートブックスクリプトを改善。
- タイプ: 雑務
- 貢献者: XXXXRT666
## 202505
- 2025.05.26 [PR#2351](https://github.com/RVC-Boss/GPT-SoVITS/pull/2351)
- 内容: DockerとWindows自動ビルドスクリプトを改善。pre-commitフォーマットを追加。
- 内容: Docker と Windows 自動ビルドスクリプトを改善。pre-commit フォーマットを追加。
- タイプ: 雑務
- 貢献者: XXXXRT666
- 2025.05.26 [PR#2408](https://github.com/RVC-Boss/GPT-SoVITS/pull/2408)
@ -544,7 +545,7 @@
- 貢献者: KamioRinn
- 関連: [Issue#2404](https://github.com/RVC-Boss/GPT-SoVITS/issues/2404)
- 2025.05.26 [PR#2377](https://github.com/RVC-Boss/GPT-SoVITS/pull/2377)
- 内容: キャッシュ戦略を実装し、SoVITS V3/V4推論速度を10%向上。
- 内容: キャッシュ戦略を実装し、SoVITS V3/V4 推論速度を 10%向上。
- タイプ: パフォーマンス最適化
- 貢献者: Kakaru Hayate
- 2025.05.26 [Commit#4d9d56b1](https://github.com/RVC-Boss/GPT-SoVITS/commit/4d9d56b19638dc434d6eefd9545e4d8639a3e072), [Commit#8c705784](https://github.com/RVC-Boss/GPT-SoVITS/commit/8c705784c50bf438c7b6d0be33a9e5e3cb90e6b2), [Commit#fafe4e7f](https://github.com/RVC-Boss/GPT-SoVITS/commit/fafe4e7f120fba56c5f053c6db30aa675d5951ba)
@ -552,7 +553,7 @@
- タイプ: 修正
- 貢献者: RVC-Boss
- 2025.05.29 [Commit#1934fc1e](https://github.com/RVC-Boss/GPT-SoVITS/commit/1934fc1e1b22c4c162bba1bbe7d7ebb132944cdc)
- 内容: UVR5およびONNX dereverberationモデルのエラーを修正。FFmpegが元のパスにスペースを含むMP3/M4Aファイルをエンコードする場合の問題を解決。
- 内容: UVR5 および ONNX dereverberation モデルのエラーを修正。FFmpeg が元のパスにスペースを含む MP3/M4A ファイルをエンコードする場合の問題を解決。
- タイプ: 修正
- 貢献者: RVC-Boss

57
docs/ja/README.md
View File

@ -1,16 +1,23 @@
#
<div align="center">
<h1>GPT-SoVITS-WebUI</h1>
パワフルなFew-Shot音声変換・音声合成 WebUI.<br><br>
パワフルなFew-Shot音声変換・音声合成 WebUI. <br><br>
[![madewithlove](https://img.shields.io/badge/made_with-%E2%9D%A4-red?style=for-the-badge&labelColor=orange)](https://github.com/RVC-Boss/GPT-SoVITS)
<img src="https://counter.seku.su/cmoe?name=gptsovits&theme=r34" /><br>
[![Train In Colab](https://img.shields.io/badge/Colab-Training-F9AB00?style=for-the-badge&logo=googlecolab)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/Colab-WebUI.ipynb)
[![Infer In Colab](https://img.shields.io/badge/Colab-Inference-F9AB00?style=for-the-badge&logo=googlecolab)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/Colab-Inference.ipynb)
[![Huggingface](https://img.shields.io/badge/HuggingFace-online%20demo-blue.svg?style=for-the-badge&logo=huggingface)](https://huggingface.co/spaces/lj1995/GPT-SoVITS-v2)
[![Open In Colab](https://img.shields.io/badge/Colab-F9AB00?style=for-the-badge&logo=googlecolab&color=525252)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/colab_webui.ipynb)
[![License](https://img.shields.io/badge/LICENSE-MIT-green.svg?style=for-the-badge)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/LICENSE)
[![Huggingface](https://img.shields.io/badge/🤗%20-online%20demo-yellow.svg?style=for-the-badge)](https://huggingface.co/spaces/lj1995/GPT-SoVITS-v2)
[![Discord](https://img.shields.io/discord/1198701940511617164?color=%23738ADB&label=Discord&style=for-the-badge)](https://discord.gg/dnrgs5GHfG)
[![GitHub release](https://img.shields.io/github/v/release/RVC-Boss/gpt-sovits?style=for-the-badge&logo=github)](https://github.com/RVC-Boss/gpt-sovits/releases)
[![Image Size](https://img.shields.io/docker/image-size/xxxxrt666/gpt-sovits/latest?style=for-the-badge&logo=docker)](https://hub.docker.com/r/xxxxrt666/gpt-sovits)
[![License](https://img.shields.io/badge/LICENSE-MIT-green.svg?style=for-the-badge&logo=opensourceinitiative)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/LICENSE)
[![简体中文](https://img.shields.io/badge/简体中文-阅读文档-blue?style=for-the-badge&logo=googledocs&logoColor=white)](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e)
[![English](https://img.shields.io/badge/English-Read%20Docs-blue?style=for-the-badge&logo=googledocs&logoColor=white)](https://rentry.co/GPT-SoVITS-guide#/)
[![Change Log](https://img.shields.io/badge/Change%20Log-View%20Updates-blue?style=for-the-badge&logo=googledocs&logoColor=white)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/docs/en/Changelog_EN.md)
[**English**](../../README.md) | [**中文简体**](../cn/README.md) | **日本語** | [**한국어**](../ko/README.md) | [**Türkçe**](../tr/README.md)
@ -18,7 +25,7 @@
---
## 機能:
## 機能
1. **Zero-Shot TTS:** たった 5 秒間の音声サンプルで、即座にテキストからその音声に変換できます.
@ -32,9 +39,9 @@
声の事前学習無しかつ Few-Shot でトレーニングされたモデルのデモ:
https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-80c060ab47fb
<https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-80c060ab47fb>
**ユーザーマニュアル: [简体中文](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e) | [English](https://rentry.co/GPT-SoVITS-guide#/)**
<!-- **ユーザーマニュアル: [简体中文](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e) | [English](https://rentry.co/GPT-SoVITS-guide#/)** -->
## インストール
@ -165,7 +172,7 @@ docker exec -it <GPT-SoVITS-CU126-Lite|GPT-SoVITS-CU128-Lite|GPT-SoVITS-CU126|GP
## 事前訓練済みモデル
**`install.sh`が正常に実行された場合、No.1,2,3 はスキップしてかまいません.**
**`install.sh`が正常に実行された場合、No.1, 2, 3 はスキップしてかまいません.**
1. [GPT-SoVITS Models](https://huggingface.co/lj1995/GPT-SoVITS) から事前訓練済みモデルをダウンロードし、`GPT_SoVITS/pretrained_models` ディレクトリに配置してください.
@ -173,19 +180,19 @@ docker exec -it <GPT-SoVITS-CU126-Lite|GPT-SoVITS-CU128-Lite|GPT-SoVITS-CU126|GP
3. UVR5 (ボーカル/伴奏 (BGM 等) 分離 & リバーブ除去の追加機能) の場合は、[UVR5 Weights](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/uvr5_weights) からモデルをダウンロードし、`tools/uvr5/uvr5_weights` ディレクトリに配置してください.
- UVR5 で bs_roformer または mel_band_roformer モデルを使用する場合、モデルと対応する設定ファイルを手動でダウンロードし、`tools/UVR5/UVR5_weights`フォルダに配置することができます.**モデルファイルと設定ファイルの名前は、拡張子を除いて同じであることを確認してください**.さらに、モデルと設定ファイルの名前には**「roformer」が含まれている必要があります**.これにより、roformer クラスのモデルとして認識されます.
- UVR5 で bs_roformer または mel_band_roformer モデルを使用する場合、モデルと対応する設定ファイルを手動でダウンロードし、`tools/UVR5/UVR5_weights`フォルダに配置することができます. **モデルファイルと設定ファイルの名前は、拡張子を除いて同じであることを確認してください**. さらに、モデルと設定ファイルの名前には**「roformer」が含まれている必要があります**. これにより、roformer クラスのモデルとして認識されます.
- モデル名と設定ファイル名には、**直接モデルタイプを指定することをお勧めします**.例: mel_mand_roformer、bs_roformer.指定しない場合、設定文から特徴を照合して、モデルの種類を特定します.例えば、モデル`bs_roformer_ep_368_sdr_12.9628.ckpt`と対応する設定ファイル`bs_roformer_ep_368_sdr_12.9628.yaml`はペアです.同様に、`kim_mel_band_roformer.ckpt`と`kim_mel_band_roformer.yaml`もペアです.
- モデル名と設定ファイル名には、**直接モデルタイプを指定することをお勧めします**. 例: mel_mand_roformer、bs_roformer. 指定しない場合、設定文から特徴を照合して、モデルの種類を特定します. 例えば、モデル`bs_roformer_ep_368_sdr_12.9628.ckpt`と対応する設定ファイル`bs_roformer_ep_368_sdr_12.9628.yaml`はペアです. 同様に、`kim_mel_band_roformer.ckpt`と`kim_mel_band_roformer.yaml`もペアです.
4. 中国語 ASR (追加機能) の場合は、[Damo ASR Model](https://modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/files)、[Damo VAD Model](https://modelscope.cn/models/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/files)、および [Damo Punc Model](https://modelscope.cn/models/damo/punc_ct-transformer_zh-cn-common-vocab272727-pytorch/files) からモデルをダウンロードし、`tools/asr/models` ディレクトリに配置してください.
5. 英語または日本語の ASR (追加機能) を使用する場合は、[Faster Whisper Large V3](https://huggingface.co/Systran/faster-whisper-large-v3) からモデルをダウンロードし、`tools/asr/models` ディレクトリに配置してください.また、[他のモデル](https://huggingface.co/Systran) は、より小さいサイズで高クオリティな可能性があります.
5. 英語または日本語の ASR (追加機能) を使用する場合は、[Faster Whisper Large V3](https://huggingface.co/Systran/faster-whisper-large-v3) からモデルをダウンロードし、`tools/asr/models` ディレクトリに配置してください. また、[他のモデル](https://huggingface.co/Systran) は、より小さいサイズで高クオリティな可能性があります.
## データセット形式
TTS アノテーション .list ファイル形式:
```
```text
vocal_path|speaker_name|language|text
```
@ -197,7 +204,7 @@ vocal_path|speaker_name|language|text
例:
```
```text
D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
```
@ -208,7 +215,6 @@ D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
#### 統合パッケージ利用者
`go-webui.bat`をダブルクリックするか、`go-webui.ps1`を使用します.
V1 に切り替えたい場合は、`go-webui-v1.bat`をダブルクリックするか、`go-webui-v1.ps1`を使用してください.
#### その他
@ -216,14 +222,6 @@ V1 に切り替えたい場合は、`go-webui-v1.bat`をダブルクリックす
python webui.py <言語(オプション)>
```
V1 に切り替えたい場合は
```bash
python webui.py v1 <言語(オプション)>
```
または WebUI で手動でバージョンを切り替えてください.
### 微調整
#### パス自動補完のサポート
@ -239,7 +237,7 @@ python webui.py v1 <言語(オプション)>
#### 統合パッケージ利用者
`go-webui-v2.bat`をダブルクリックするか、`go-webui-v2.ps1`を使用して、`1-GPT-SoVITS-TTS/1C-inference`で推論 webui を開きます.
`go-webui.bat`をダブルクリックするか、`go-webui.ps1`を使用して、`1-GPT-SoVITS-TTS/1C-inference`で推論 webui を開きます.
#### その他
@ -359,11 +357,6 @@ V1/V2/V3/V4 環境から V2Pro への移行方法:
python tools/uvr5/webui.py "<infer_device>" <is_half> <webui_port_uvr5>
```
<!-- ブラウザを開けない場合は、以下の形式に従って UVR 処理を行ってください.これはオーディオ処理に mdxnet を使用しています.
```
python mdxnet.py --model --input_root --output_vocal --output_ins --agg_level --format --device --is_half_precision
``` -->
コマンド ラインを使用してデータセットのオーディオ セグメンテーションを行う方法は次のとおりです.
```bash
@ -384,7 +377,7 @@ python tools/asr/funasr_asr.py -i <input> -o <output>
ASR 処理は Faster_Whisper を通じて実行されます(中国語を除く ASR マーキング)
(進行状況バーは表示されません.GPU のパフォーマンスにより時間遅延が発生する可能性があります)
(進行状況バーは表示されません. GPU のパフォーマンスにより時間遅延が発生する可能性があります)
```bash
python ./tools/asr/fasterwhisper_asr.py -i <input> -o <output> -l <language> -p <precision>
@ -439,5 +432,5 @@ python ./tools/asr/fasterwhisper_asr.py -i <input> -o <output> -l <language> -p
## すべてのコントリビューターに感謝します
<a href="https://github.com/RVC-Boss/GPT-SoVITS/graphs/contributors" target="_blank">
<img src="https://contrib.rocks/image?repo=RVC-Boss/GPT-SoVITS" />
<img src="https://contrib.rocks/image?repo=RVC-Boss/GPT-SoVITS" alt="Contributors"/>
</a>

9
docs/ko/Changelog_KO.md
View File

@ -242,7 +242,7 @@
- 유형: 최적화
- 기여자: GoHomeToMacDonal
- 2024.03.10 [PR#721](https://github.com/RVC-Boss/GPT-SoVITS/pull/721)
- 내용: 빠른 추론 브랜치 'fast_inference_' 추가
- 내용: 빠른 추론 브랜치 'fast*inference*' 추가
- 유형: 기능
- 기여자: ChasonJiang
- 2024.03.13 [PR#761](https://github.com/RVC-Boss/GPT-SoVITS/pull/761)
@ -285,7 +285,7 @@
- 2024.06.06 [Commit#99f09c8b](https://github.com/RVC-Boss/GPT-SoVITS/commit/99f09c8bdc155c1f4272b511940717705509582a)
- 내용: WebUI의 GPT 미세조정 시 중국어 입력 텍스트의 BERT 특징을 읽지 않아 추론과 불일치 및 품질 저하가 발생하는 문제 수정
**주의: 이전에 대량의 데이터로 미세조정을 한 경우 품질 향상을 위해 모델 재조정 권장**
**주의: 이전에 대량의 데이터로 미세조정을 한 경우 품질 향상을 위해 모델 재조정 권장**
- 유형: 수정
- 기여자: RVC-Boss
- 2024.06.07 [PR#1159](https://github.com/RVC-Boss/GPT-SoVITS/pull/1159)
@ -347,7 +347,7 @@
- 2024.07.23 [Commit#9588a3c5](https://github.com/RVC-Boss/GPT-SoVITS/commit/9588a3c52d9ebdb20b3c5d74f647d12e7c1171c2), [PR#1340](https://github.com/RVC-Boss/GPT-SoVITS/pull/1340)
- 내용: 합성 중 음성 속도 조절 기능 추가(무작위성 고정 및 속도만 제어 옵션 포함). 이 기능은 `api.py`에 업데이트됨
- 유형: 기능
- 기여자: RVC-Boss, 红血球AE3803
- 기여자: RVC-Boss, 红血球 AE3803
- 2024.07.27 [PR#1306](https://github.com/RVC-Boss/GPT-SoVITS/pull/1306), [PR#1356](https://github.com/RVC-Boss/GPT-SoVITS/pull/1356)
- 내용: BS-RoFormer 보컬 분리 모델 지원 추가
- 유형: 기능
@ -409,7 +409,7 @@
- 2025.02.11 [Commit#ed207c4b](https://github.com/RVC-Boss/GPT-SoVITS/commit/ed207c4b879d5296e9be3ae5f7b876729a2c43b8)~[Commit#6e2b4918](https://github.com/RVC-Boss/GPT-SoVITS/commit/6e2b49186c5b961f0de41ea485d398dffa9787b4)
- 내용: **GPT-SoVITS V3 모델 추가, 파인튜닝 시 14GB VRAM 필요.**
- 유형: 신규 기능 ([위키 참조](https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)))
- 유형: 신규 기능 ([위키 참조](<https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)>))
- 기여자: RVC-Boss
- 2025.02.12 [PR#2032](https://github.com/RVC-Boss/GPT-SoVITS/pull/2032)
- 내용: 다국어 프로젝트 문서 업데이트.
@ -475,6 +475,7 @@
- Pydantic: [Issue#2230](https://github.com/RVC-Boss/GPT-SoVITS/issues/2230), [Issue#2239](https://github.com/RVC-Boss/GPT-SoVITS/issues/2239).
- PyTorch-Lightning: [Issue#2174](https://github.com/RVC-Boss/GPT-SoVITS/issues/2174).
- 2025.03.31 [PR#2241](https://github.com/RVC-Boss/GPT-SoVITS/pull/2241)
- 내용: **SoVITS v3 병렬 추론 지원 활성화.**
- 유형: 신규 기능
- 기여자: ChasonJiang

48
docs/ko/README.md
View File

@ -1,3 +1,5 @@
#
<div align="center">
<h1>GPT-SoVITS-WebUI</h1>
@ -5,12 +7,17 @@
[![madewithlove](https://img.shields.io/badge/made_with-%E2%9D%A4-red?style=for-the-badge&labelColor=orange)](https://github.com/RVC-Boss/GPT-SoVITS)
<img src="https://counter.seku.su/cmoe?name=gptsovits&theme=r34" /><br>
[![Train In Colab](https://img.shields.io/badge/Colab-Training-F9AB00?style=for-the-badge&logo=googlecolab)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/Colab-WebUI.ipynb)
[![Infer In Colab](https://img.shields.io/badge/Colab-Inference-F9AB00?style=for-the-badge&logo=googlecolab)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/Colab-Inference.ipynb)
[![Huggingface](https://img.shields.io/badge/HuggingFace-online%20demo-blue.svg?style=for-the-badge&logo=huggingface)](https://huggingface.co/spaces/lj1995/GPT-SoVITS-v2)
[![Open In Colab](https://img.shields.io/badge/Colab-F9AB00?style=for-the-badge&logo=googlecolab&color=525252)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/colab_webui.ipynb)
[![License](https://img.shields.io/badge/LICENSE-MIT-green.svg?style=for-the-badge)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/LICENSE)
[![Huggingface](https://img.shields.io/badge/🤗%20-online%20demo-yellow.svg?style=for-the-badge)](https://huggingface.co/spaces/lj1995/GPT-SoVITS-v2)
[![Discord](https://img.shields.io/discord/1198701940511617164?color=%23738ADB&label=Discord&style=for-the-badge)](https://discord.gg/dnrgs5GHfG)
[![GitHub release](https://img.shields.io/github/v/release/RVC-Boss/gpt-sovits?style=for-the-badge&logo=github)](https://github.com/RVC-Boss/gpt-sovits/releases)
[![Image Size](https://img.shields.io/docker/image-size/xxxxrt666/gpt-sovits/latest?style=for-the-badge&logo=docker)](https://hub.docker.com/r/xxxxrt666/gpt-sovits)
[![License](https://img.shields.io/badge/LICENSE-MIT-green.svg?style=for-the-badge&logo=opensourceinitiative)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/LICENSE)
[![简体中文](https://img.shields.io/badge/简体中文-阅读文档-blue?style=for-the-badge&logo=googledocs&logoColor=white)](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e)
[![English](https://img.shields.io/badge/English-Read%20Docs-blue?style=for-the-badge&logo=googledocs&logoColor=white)](https://rentry.co/GPT-SoVITS-guide#/)
[![Change Log](https://img.shields.io/badge/Change%20Log-View%20Updates-blue?style=for-the-badge&logo=googledocs&logoColor=white)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/docs/en/Changelog_EN.md)
[**English**](../../README.md) | [**中文简体**](../cn/README.md) | [**日本語**](../ja/README.md) | **한국어** | [**Türkçe**](../tr/README.md)
@ -18,7 +25,7 @@
---
## 기능:
## 기능
1. **제로샷 텍스트 음성 변환 (TTS):** 5초의 음성 샘플을 입력하면 즉시 텍스트를 음성으로 변환할 수 있습니다.
@ -32,9 +39,9 @@
보지 못한 발화자의 퓨샷(few-shot) 파인튜닝 데모:
https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-80c060ab47fb
<https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-80c060ab47fb>
**사용자 설명서: [简体中文](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e) | [English](https://rentry.co/GPT-SoVITS-guide#/)**
<!-- **사용자 설명서: [简体中文](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e) | [English](https://rentry.co/GPT-SoVITS-guide#/)** -->
## 설치
@ -165,7 +172,7 @@ docker exec -it <GPT-SoVITS-CU126-Lite|GPT-SoVITS-CU128-Lite|GPT-SoVITS-CU126|GP
## 사전 학습된 모델
**`install.sh`가 성공적으로 실행되면 No.1,2,3 은 건너뛰어도 됩니다.**
**`install.sh`가 성공적으로 실행되면 No.1, 2, 3 은 건너뛰어도 됩니다.**
1. [GPT-SoVITS Models](https://huggingface.co/lj1995/GPT-SoVITS) 에서 사전 학습된 모델을 다운로드하고, `GPT_SoVITS/pretrained_models` 디렉토리에 배치하세요.
@ -185,7 +192,7 @@ docker exec -it <GPT-SoVITS-CU126-Lite|GPT-SoVITS-CU128-Lite|GPT-SoVITS-CU126|GP
텍스트 음성 합성(TTS) 주석 .list 파일 형식:
```
```text
vocal_path|speaker_name|language|text
```
@ -197,7 +204,7 @@ vocal_path|speaker_name|language|text
예시:
```
```text
D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
```
@ -208,7 +215,6 @@ D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
#### 통합 패키지 사용자
`go-webui.bat`을 더블 클릭하거나 `go-webui.ps1`를 사용하십시오.
V1으로 전환하려면, `go-webui-v1.bat`을 더블 클릭하거나 `go-webui-v1.ps1`를 사용하십시오.
#### 기타
@ -216,14 +222,6 @@ V1으로 전환하려면, `go-webui-v1.bat`을 더블 클릭하거나 `go-webui-
python webui.py <언어(옵션)>
```
V1으로 전환하려면,
```bash
python webui.py v1 <언어(옵션)>
```
또는 WebUI에서 수동으로 버전을 전환하십시오.
### 미세 조정
#### 경로 자동 채우기가 지원됩니다
@ -239,7 +237,7 @@ python webui.py v1 <언어(옵션)>
#### 통합 패키지 사용자
`go-webui-v2.bat`을 더블 클릭하거나 `go-webui-v2.ps1`를 사용한 다음 `1-GPT-SoVITS-TTS/1C-inference`에서 추론 webui를 엽니다.
`go-webui.bat`을 더블 클릭하거나 `go-webui.ps1`를 사용한 다음 `1-GPT-SoVITS-TTS/1C-inference`에서 추론 webui를 엽니다.
#### 기타
@ -277,13 +275,13 @@ V1 환경에서 V2를 사용하려면:
3. [huggingface](https://huggingface.co/lj1995/GPT-SoVITS/tree/main/gsv-v2final-pretrained)에서 V2 사전 학습 모델을 다운로드하여 `GPT_SoVITS/pretrained_models/gsv-v2final-pretrained`에 넣으십시오.
중국어 V2 추가: [G2PWModel.zip(HF)](https://huggingface.co/XXXXRT/GPT-SoVITS-Pretrained/resolve/main/G2PWModel.zip)| [G2PWModel.zip(ModelScope)](https://www.modelscope.cn/models/XXXXRT/GPT-SoVITS-Pretrained/resolve/master/G2PWModel.zip) (G2PW 모델을 다운로드하여 압축을 풀고 `G2PWModel`로 이름을 변경한 다음 `GPT_SoVITS/text`에 배치합니다.)
중국어 V2 추가: [G2PWModel.zip(HF)](https://huggingface.co/XXXXRT/GPT-SoVITS-Pretrained/resolve/main/G2PWModel.zip)| [G2PWModel.zip(ModelScope)](https://www.modelscope.cn/models/XXXXRT/GPT-SoVITS-Pretrained/resolve/master/G2PWModel.zip) (G2PW 모델을 다운로드하여 압축을 풀고 `G2PWModel`로 이름을 변경한 다음 `GPT_SoVITS/text`에 배치합니다)
## V3 릴리스 노트
새로운 기능:
1. 음색 유사성이 더 높아져 목표 음성에 대한 학습 데이터가 적게 필요합니다. (기본 모델을 직접 사용하여 미세 조정 없이 음색 유사성이 크게 향상됩니다.)
1. 음색 유사성이 더 높아져 목표 음성에 대한 학습 데이터가 적게 필요합니다. (기본 모델을 직접 사용하여 미세 조정 없이 음색 유사성이 크게 향상됩니다)
2. GPT 모델이 더 안정적이며 반복 및 생략이 적고, 더 풍부한 감정 표현을 가진 음성을 생성하기가 더 쉽습니다.
@ -437,8 +435,8 @@ python ./tools/asr/fasterwhisper_asr.py -i <input> -o <output> -l <language> -p
@Naozumi520 님께 감사드립니다. 광둥어 학습 자료를 제공해 주시고, 광둥어 관련 지식을 지도해 주셔서 감사합니다.
## 모든 기여자들에게 감사드립니다 ;)
## 모든 기여자들에게 감사드립니다
<a href="https://github.com/RVC-Boss/GPT-SoVITS/graphs/contributors" target="_blank">
<img src="https://contrib.rocks/image?repo=RVC-Boss/GPT-SoVITS" />
<img src="https://contrib.rocks/image?repo=RVC-Boss/GPT-SoVITS" alt="Contributors"/>
</a>

9
docs/tr/Changelog_TR.md
View File

@ -244,7 +244,7 @@
- Tür: Optimizasyon
- Katkıda Bulunan: GoHomeToMacDonal
- 2024.03.10 [PR#721](https://github.com/RVC-Boss/GPT-SoVITS/pull/721)
- İçerik: Hızlı çıkarım dalı 'fast_inference_' eklendi
- İçerik: Hızlı çıkarım dalı 'fast*inference*' eklendi
- Tür: Özellik
- Katkıda Bulunan: ChasonJiang
- 2024.03.13 [PR#761](https://github.com/RVC-Boss/GPT-SoVITS/pull/761)
@ -287,7 +287,7 @@
- 2024.06.06 [Commit#99f09c8b](https://github.com/RVC-Boss/GPT-SoVITS/commit/99f09c8bdc155c1f4272b511940717705509582a)
- İçerik: WebUI'da GPT ince ayarında Çince metinlerin BERT özelliklerinin okunmaması nedeniyle çıkarım tutarsızlığı ve kalite düşüşü sorunu düzeltildi
**Uyarı: Daha önce büyük miktarda veriyle ince ayar yaptıysanız, kaliteyi artırmak için modeli yeniden ayarlamanız önerilir**
**Uyarı: Daha önce büyük miktarda veriyle ince ayar yaptıysanız, kaliteyi artırmak için modeli yeniden ayarlamanız önerilir**
- Tür: Düzeltme
- Katkıda Bulunan: RVC-Boss
- 2024.06.07 [PR#1159](https://github.com/RVC-Boss/GPT-SoVITS/pull/1159)
@ -347,7 +347,7 @@
- 2024.07.23 [Commit#9588a3c5](https://github.com/RVC-Boss/GPT-SoVITS/commit/9588a3c52d9ebdb20b3c5d74f647d12e7c1171c2), [PR#1340](https://github.com/RVC-Boss/GPT-SoVITS/pull/1340)
- İçerik: Sentez sırasında konuşma hızı ayarlama özelliği eklendi (rastgeleliği sabitleme ve sadece hızı kontrol etme seçeneği dahil). Bu özellik `api.py` dosyasına eklendi
- Tür: Özellik
- Katkıda Bulunan: RVC-Boss, 红血球AE3803
- Katkıda Bulunan: RVC-Boss, 红血球 AE3803
- 2024.07.27 [PR#1306](https://github.com/RVC-Boss/GPT-SoVITS/pull/1306), [PR#1356](https://github.com/RVC-Boss/GPT-SoVITS/pull/1356)
- İçerik: BS-RoFormer vokal eşlik ayırma modeli desteği eklendi.
- Tür: Yeni Özellik
@ -409,7 +409,7 @@
- 2025.02.11 [Commit#ed207c4b](https://github.com/RVC-Boss/GPT-SoVITS/commit/ed207c4b879d5296e9be3ae5f7b876729a2c43b8)~[Commit#6e2b4918](https://github.com/RVC-Boss/GPT-SoVITS/commit/6e2b49186c5b961f0de41ea485d398dffa9787b4)
- İçerik: **İnce ayar için 14GB VRAM gerektiren GPT-SoVITS V3 modeli eklendi.**
- Tür: Yeni Özellik ([Wiki](https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)) referans)
- Tür: Yeni Özellik ([Wiki](<https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)>) referans)
- Katkıda Bulunan: RVC-Boss
- 2025.02.12 [PR#2032](https://github.com/RVC-Boss/GPT-SoVITS/pull/2032)
- İçerik: Çok dilli proje dokümantasyonu güncellendi.
@ -475,6 +475,7 @@
- Pydantic: [Issue#2230](https://github.com/RVC-Boss/GPT-SoVITS/issues/2230), [Issue#2239](https://github.com/RVC-Boss/GPT-SoVITS/issues/2239).
- PyTorch-Lightning: [Issue#2174](https://github.com/RVC-Boss/GPT-SoVITS/issues/2174).
- 2025.03.31 [PR#2241](https://github.com/RVC-Boss/GPT-SoVITS/pull/2241)
- İçerik: **SoVITS v3 için paralel çıkarım etkinleştirildi.**
- Tür: Yeni Özellik
- Katkıda Bulunan: ChasonJiang

55
docs/tr/README.md
View File

@ -1,3 +1,5 @@
#
<div align="center">
<h1>GPT-SoVITS-WebUI</h1>
@ -7,12 +9,17 @@ Güçlü Birkaç Örnekli Ses Dönüştürme ve Metinden Konuşmaya Web Arayüz
<a href="https://trendshift.io/repositories/7033" target="_blank"><img src="https://trendshift.io/api/badge/repositories/7033" alt="RVC-Boss%2FGPT-SoVITS | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
<!-- img src="https://counter.seku.su/cmoe?name=gptsovits&theme=r34" /><br> -->
[![Train In Colab](https://img.shields.io/badge/Colab-Training-F9AB00?style=for-the-badge&logo=googlecolab)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/Colab-WebUI.ipynb)
[![Infer In Colab](https://img.shields.io/badge/Colab-Inference-F9AB00?style=for-the-badge&logo=googlecolab)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/Colab-Inference.ipynb)
[![Huggingface](https://img.shields.io/badge/HuggingFace-online%20demo-blue.svg?style=for-the-badge&logo=huggingface)](https://huggingface.co/spaces/lj1995/GPT-SoVITS-v2)
[![Open In Colab](https://img.shields.io/badge/Colab-F9AB00?style=for-the-badge&logo=googlecolab&color=525252)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/colab_webui.ipynb)
[![License](https://img.shields.io/badge/LICENSE-MIT-green.svg?style=for-the-badge)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/LICENSE)
[![Huggingface](https://img.shields.io/badge/🤗%20-online%20demo-yellow.svg?style=for-the-badge)](https://huggingface.co/spaces/lj1995/GPT-SoVITS-v2)
[![Discord](https://img.shields.io/discord/1198701940511617164?color=%23738ADB&label=Discord&style=for-the-badge)](https://discord.gg/dnrgs5GHfG)
[![GitHub release](https://img.shields.io/github/v/release/RVC-Boss/gpt-sovits?style=for-the-badge&logo=github)](https://github.com/RVC-Boss/gpt-sovits/releases)
[![Image Size](https://img.shields.io/docker/image-size/xxxxrt666/gpt-sovits/latest?style=for-the-badge&logo=docker)](https://hub.docker.com/r/xxxxrt666/gpt-sovits)
[![License](https://img.shields.io/badge/LICENSE-MIT-green.svg?style=for-the-badge&logo=opensourceinitiative)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/LICENSE)
[![简体中文](https://img.shields.io/badge/简体中文-阅读文档-blue?style=for-the-badge&logo=googledocs&logoColor=white)](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e)
[![English](https://img.shields.io/badge/English-Read%20Docs-blue?style=for-the-badge&logo=googledocs&logoColor=white)](https://rentry.co/GPT-SoVITS-guide#/)
[![Change Log](https://img.shields.io/badge/Change%20Log-View%20Updates-blue?style=for-the-badge&logo=googledocs&logoColor=white)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/docs/en/Changelog_EN.md)
[**English**](../../README.md) | [**中文简体**](../cn/README.md) | [**日本語**](../ja/README.md) | [**한국어**](../ko/README.md) | **Türkçe**
@ -20,7 +27,7 @@ Güçlü Birkaç Örnekli Ses Dönüştürme ve Metinden Konuşmaya Web Arayüz
---
## Özellikler:
## Özellikler
1. **Sıfır Örnekli Metinden Konuşmaya:** 5 saniyelik bir vokal örneği girin ve anında metinden konuşmaya dönüşümünü deneyimleyin.
@ -34,9 +41,9 @@ Güçlü Birkaç Örnekli Ses Dönüştürme ve Metinden Konuşmaya Web Arayüz
Görünmeyen konuşmacılar birkaç örnekli ince ayar demosu:
https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-80c060ab47fb
<https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-80c060ab47fb>
**Kullanıcı Kılavuzu: [简体中文](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e) | [English](https://rentry.co/GPT-SoVITS-guide#/)**
<!-- **Kullanıcı Kılavuzu: [简体中文](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e) | [English](https://rentry.co/GPT-SoVITS-guide#/)** -->
## Kurulum
@ -167,7 +174,7 @@ docker exec -it <GPT-SoVITS-CU126-Lite|GPT-SoVITS-CU128-Lite|GPT-SoVITS-CU126|GP
## Önceden Eğitilmiş Modeller
**Eğer `install.sh` başarıyla çalıştırılırsa, No.1,2,3 adımını atlayabilirsiniz.**
**Eğer `install.sh` başarıyla çalıştırılırsa, No.1, 2, 3 adımını atlayabilirsiniz.**
1. [GPT-SoVITS Models](https://huggingface.co/lj1995/GPT-SoVITS) üzerinden önceden eğitilmiş modelleri indirip `GPT_SoVITS/pretrained_models` dizinine yerleştirin.
@ -187,7 +194,7 @@ docker exec -it <GPT-SoVITS-CU126-Lite|GPT-SoVITS-CU128-Lite|GPT-SoVITS-CU126|GP
TTS açıklama .list dosya formatı:
```
```text
vocal_path|speaker_name|language|text
```
@ -201,7 +208,7 @@ Dil sözlüğü:
Örnek:
```
```text
D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
```
@ -212,7 +219,6 @@ D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
#### Entegre Paket Kullanıcıları
`go-webui.bat` dosyasına çift tıklayın veya `go-webui.ps1` kullanın.
V1'e geçmek istiyorsanız, `go-webui-v1.bat` dosyasına çift tıklayın veya `go-webui-v1.ps1` kullanın.
#### Diğerleri
@ -220,14 +226,6 @@ V1'e geçmek istiyorsanız, `go-webui-v1.bat` dosyasına çift tıklayın veya `
python webui.py <dil(isteğe bağlı)>
```
V1'e geçmek istiyorsanız,
```bash
python webui.py v1 <dil(isteğe bağlı)>
```
veya WebUI'de manuel olarak sürüm değiştirin.
### İnce Ayar
#### Yol Otomatik Doldurma artık destekleniyor
@ -243,7 +241,7 @@ veya WebUI'de manuel olarak sürüm değiştirin.
#### Entegre Paket Kullanıcıları
`go-webui-v2.bat` dosyasına çift tıklayın veya `go-webui-v2.ps1` kullanın, ardından çıkarım webui'sini `1-GPT-SoVITS-TTS/1C-inference` adresinde açın.
`go-webui.bat` dosyasına çift tıklayın veya `go-webui.ps1` kullanın, ardından çıkarım webui'sini `1-GPT-SoVITS-TTS/1C-inference` adresinde açın.
#### Diğerleri
@ -281,11 +279,11 @@ V1 ortamından V2'yi kullanmak için:
3. [huggingface](https://huggingface.co/lj1995/GPT-SoVITS/tree/main/gsv-v2final-pretrained) adresinden v2 önceden eğitilmiş modelleri indirin ve bunları `GPT_SoVITS/pretrained_models/gsv-v2final-pretrained` dizinine yerleştirin.
Ek olarak Çince V2: [G2PWModel.zip(HF)](https://huggingface.co/XXXXRT/GPT-SoVITS-Pretrained/resolve/main/G2PWModel.zip)| [G2PWModel.zip(ModelScope)](https://www.modelscope.cn/models/XXXXRT/GPT-SoVITS-Pretrained/resolve/master/G2PWModel.zip) (G2PW modellerini indirip, zipten çıkarıp, `G2PWModel` olarak yeniden adlandırıp `GPT_SoVITS/text` dizinine yerleştirin.)
Ek olarak Çince V2: [G2PWModel.zip(HF)](https://huggingface.co/XXXXRT/GPT-SoVITS-Pretrained/resolve/main/G2PWModel.zip)| [G2PWModel.zip(ModelScope)](https://www.modelscope.cn/models/XXXXRT/GPT-SoVITS-Pretrained/resolve/master/G2PWModel.zip) (G2PW modellerini indirip, zipten çıkarıp, `G2PWModel` olarak yeniden adlandırıp `GPT_SoVITS/text` dizinine yerleştirin)
## V3 Sürüm Notları
Yeni Özellikler:
### Yeni Özellikler
1. **Tını benzerliği** daha yüksek olup, hedef konuşmacıyı yakınsamak için daha az eğitim verisi gerekmektedir (tını benzerliği, base model doğrudan kullanılacak şekilde fine-tuning yapılmadan önemli ölçüde iyileştirilmiştir).
@ -293,7 +291,7 @@ Yeni Özellikler:
[daha fazla detay](<https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)>)
V2 ortamında V3 kullanımı:
### v2 ortamında v3 kullanımı
1. `pip install -r requirements.txt` ile bazı paketleri güncelleyin.
@ -323,7 +321,7 @@ V1/V2/V3 ortamından V4'e geçiş:
Yeni Özellikler:
1. **V2 ile karşılaştırıldığında biraz daha yüksek VRAM kullanımı sağlar ancak V4'ten daha iyi performans gösterir; aynı donanım maliyeti ve hız avantajını korur**.
[Daha fazla bilgi](https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90features-(%E5%90%84%E7%89%88%E6%9C%AC%E7%89%B9%E6%80%A7))
[Daha fazla bilgi](<https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90features-(%E5%90%84%E7%89%88%E6%9C%AC%E7%89%B9%E6%80%A7)>)
2. V1/V2 ve V2Pro serisi benzer özelliklere sahipken, V3/V4 de yakın işlevleri paylaşır. Ortalama kalite düşük olan eğitim setleriyle V1/V2/V2Pro iyi sonuçlar verebilir ama V3/V4 veremez. Ayrıca, V3/V4’ün ürettiği ses tonu genel eğitim setine değil, referans ses örneğine daha çok benzemektedir.
@ -363,11 +361,6 @@ UVR5 için Web Arayüzünü açmak için komut satırını kullanın
python tools/uvr5/webui.py "<infer_device>" <is_half> <webui_port_uvr5>
```
<!-- Bir tarayıcı açamıyorsanız, UVR işleme için aşağıdaki formatı izleyin,Bu ses işleme için mdxnet kullanıyor
```
python mdxnet.py --model --input_root --output_vocal --output_ins --agg_level --format --device --is_half_precision
``` -->
Veri setinin ses segmentasyonu komut satırı kullanılarak bu şekilde yapılır
```bash
@ -443,5 +436,5 @@ python ./tools/asr/fasterwhisper_asr.py -i <girdi> -o <çıktı> -l <dil>
## Tüm katkıda bulunanlara çabaları için teşekkürler
<a href="https://github.com/RVC-Boss/GPT-SoVITS/graphs/contributors" target="_blank">
<img src="https://contrib.rocks/image?repo=RVC-Boss/GPT-SoVITS" />
<img src="https://contrib.rocks/image?repo=RVC-Boss/GPT-SoVITS" alt="Contributors"/>
</a>

5
install.sh
View File

@ -34,8 +34,8 @@ print_help() {
echo " -h, --help Show this help message and exit"
echo ""
echo "Examples:"
echo " bash install.sh --source HF --download-uvr5"
echo " bash install.sh --source ModelScope"
echo " bash install.sh --device CU128 --source HF --download-uvr5"
echo " bash install.sh --device MPS --source ModelScope"
}
# Show help if no arguments provided
@ -149,7 +149,6 @@ else
echo "Installing，Please Wait..."
fi
done
conda install -c conda-forge -q -y
fi
echo "Installing ffmpeg and cmake..."

2
requirements.txt
View File

@ -8,7 +8,7 @@ pytorch-lightning>=2.4
gradio<5
ffmpeg-python
onnxruntime; platform_machine == "aarch64" or platform_machine == "arm64"
onnxruntime-gpu; platform_machine == "x86_64" or platform_machine == "AMD64"
onnxruntime-gpu; platform_machine == "x86_64" or platform_machine == "amd64"
tqdm
funasr==1.0.27
cn2an

2
tools/my_utils.py
View File

@ -109,7 +109,7 @@ def check_details(path_list=None, is_train=False, is_dataset_processing=False):
if os.path.exists(wav_path):
...
else:
gr.Warning(wav_path+i18n("路径错误"))
gr.Warning(wav_path + i18n("路径错误"))
return
if is_train:
path_list.append(os.path.join(path_list[0], "2-name2text.txt"))

544
tools/subfix.py Normal file
View File

@ -0,0 +1,544 @@
import datetime
import os
import threading
import traceback
from dataclasses import dataclass
from functools import partial
from typing import List
import click
import gradio as gr
import librosa
import numpy as np
import soundfile
from gradio.components.audio import WaveformOptions
from tools.i18n.i18n import I18nAuto
PARTIAL_EXIT = partial(os._exit, 0)
LANGUAGE_MAP: dict = {
"ZH": "ZH",
"zh": "ZH",
"JP": "JA",
"jp": "JA",
"JA": "JA",
"ja": "JA",
"EN": "EN",
"en": "EN",
"KO": "KO",
"ko": "KO",
"yue": "YUE",
"YUE": "YUE",
}
LOCK = threading.Lock()
IS_CLI = True
@dataclass
class SubfixErr:
error: Exception
tracebacks: str
class Subfix:
batch_size: int = 2
cur_idx: int = 0
list_path: str
textboxes: List[gr.Textbox] = []
audios: List[gr.Audio] = []
languages: List[gr.Dropdown] = []
selections: List[gr.Checkbox] = []
transcriptions_list: List[List[str]] = []
merge_audio_button: gr.Button
delete_audio_button: gr.Button
previous_index_button1: gr.Button
next_index_button1: gr.Button
previous_index_button2: gr.Button
next_index_button2: gr.Button
index_slider: gr.Slider
batch_size_slider: gr.Slider
close_button: gr.Button
def __init__(self, i18n: I18nAuto):
self.i18n = i18n
with gr.Row(equal_height=True):
with gr.Column(scale=2, min_width=160):
self.index_slider = gr.Slider(minimum=0, maximum=1, step=1, label=i18n("音频索引"))
with gr.Column(scale=1, min_width=160):
self.previous_index_button1 = gr.Button(value=i18n("上一页"), elem_id="btn_previous")
with gr.Column(scale=1, min_width=160):
self.next_index_button1 = gr.Button(value=i18n("下一页"), elem_id="btn_next")
with gr.Row(equal_height=True):
with gr.Column(scale=2, min_width=160):
self.batch_size_slider = gr.Slider(
minimum=4, maximum=20, step=2, value=self.batch_size, label=i18n("每页音频条数")
)
with gr.Column(scale=1, min_width=160):
self.merge_audio_button = gr.Button(value=i18n("合并选中音频"))
with gr.Column(scale=1, min_width=160):
self.delete_audio_button = gr.Button(value=i18n("删除选中音频"))
gr.render(
inputs=[self.index_slider, self.batch_size_slider],
triggers=[self.batch_size_slider.change],
)(self._render_text_area)
@property
def max_index(self):
return len(self.transcriptions_list) - 1
def load_list(self, list_path: str):
with open(list_path, mode="r", encoding="utf-8") as f:
list_data = f.readlines()
for idx, transcriptions in enumerate(list_data):
data = transcriptions.split("|")
if len(data) != 4:
print(f"Error Line {idx + 1}: {'|'.join(data)}")
continue
audio_name, audio_folder, text_language, text = data
self.transcriptions_list.append(
[
audio_name,
audio_folder,
LANGUAGE_MAP.get(text_language.upper(), text_language.upper()),
text.strip("\n").strip(),
]
)
self.list_path = list_path
def save_list(self):
data = []
for transcriptions in self.transcriptions_list:
data.append("|".join(transcriptions))
try:
with open(self.list_path, mode="w", encoding="utf-8") as f:
f.write("\n".join(data))
except Exception as e:
return SubfixErr(e, traceback.format_exc())
def change_index(self, index: int):
audios = []
texts = []
languages = []
checkboxs = []
with LOCK:
for i in range(index, index + self.batch_size):
if i <= self.max_index:
audios.append(gr.Audio(value=self.transcriptions_list[i][0]))
texts.append(gr.Textbox(value=self.transcriptions_list[i][3], label=self.i18n("Text") + f" {i}"))
languages.append(gr.Dropdown(value=self.transcriptions_list[i][2]))
else:
audios.append(gr.Audio(value=None, interactive=False))
texts.append(gr.Textbox(value=None, label=self.i18n("Text") + f" {i}", interactive=False))
languages.append(gr.Dropdown(value=None, interactive=False))
checkboxs = [gr.Checkbox(False) for i in range(self.batch_size)]
self.cur_idx = index
return *audios, *texts, *languages, *checkboxs
def next_page(self, index: int):
batch_size = self.batch_size
max_index = max(self.max_index - batch_size + 1, 0)
index = min(index + batch_size, max_index)
return gr.Slider(value=index), *self.change_index(index)
def previous_page(self, index: int):
batch_size = self.batch_size
index = max(index - batch_size, 0)
return gr.Slider(value=index), *self.change_index(index)
def delete_audio(self, index, *selected):
delete_index = [i + index for i, _ in enumerate(selected) if _]
delete_index = [i for i in delete_index if i < self.max_index]
for idx in delete_index[::-1]:
self.transcriptions_list.pop(idx)
self.save_list()
return gr.Slider(value=index, maximum=self.max_index), *self.change_index(index)
def submit(self, *input):
with LOCK:
index = self.cur_idx
batch_size = self.batch_size
texts = input[: len(input) // 2]
languages = input[len(input) // 2 :]
if texts is None or languages is None:
raise ValueError()
print(index, min(index + batch_size, self.max_index))
for idx in range(index, min(index + batch_size, self.max_index + 1)):
self.transcriptions_list[idx][3] = texts[idx - index].strip().strip("\n")
self.transcriptions_list[idx][2] = languages[idx - index]
result = self.save_list()
if isinstance(result, SubfixErr):
gr.Warning(str(result.error))
print(result.tracebacks)
def merge_audio(self, index, *selected):
batch_size = self.batch_size
merge_index = [i + index for i, _ in enumerate(selected) if _]
merge_index = [i for i in merge_index if i < self.max_index]
if len(merge_index) < 2:
return *(gr.skip() for _ in range(batch_size * 3 + 1)), *(gr.Checkbox(False) for _ in range(batch_size))
else:
merge_texts = []
merge_audios = []
first_itm_index = merge_index[0]
first_itm_path = f"{os.path.splitext(self.transcriptions_list[first_itm_index][0])[0]}_{str(datetime.datetime.now().strftime(r'%Y%m%d_%H%M%S'))}.wav"
final_audio_list = []
for idx in merge_index:
merge_texts.append(self.transcriptions_list[idx][3])
merge_audios.append(self.transcriptions_list[idx][0])
for idx in merge_index[:0:-1]:
self.transcriptions_list.pop(idx)
for audio_path in merge_audios:
final_audio_list.append(librosa.load(audio_path, sr=32000, mono=True)[0])
final_audio_list.append(np.zeros(int(32000 * 0.3)))
final_audio_list.pop()
final_audio = np.concatenate(final_audio_list)
soundfile.write(first_itm_path, final_audio, 32000)
self.transcriptions_list[first_itm_index][0] = first_itm_path
self.transcriptions_list[first_itm_index][3] = ",".join(merge_texts)
return gr.Slider(maximum=self.max_index), *self.change_index(index)
def _render_text_area(self, index, batch_size):
i18n = self.i18n
self.textboxes = []
self.audios = []
self.languages = []
self.selections = []
self.batch_size = batch_size
for i in range(index, index + batch_size):
with gr.Row(equal_height=True):
if i <= self.max_index:
with gr.Column(scale=2, min_width=160):
textbox_tmp = gr.Textbox(
value=self.transcriptions_list[i][3],
label=i18n("Text") + f" {i}",
lines=2,
max_lines=3,
interactive=True,
)
with gr.Column(scale=1, min_width=160):
audio_tmp = gr.Audio(
value=self.transcriptions_list[i][0],
show_label=False,
show_download_button=False,
editable=False,
waveform_options={"show_recording_waveform": False, "show_controls": False},
)
with gr.Column(scale=1, min_width=160):
with gr.Group():
with gr.Row():
language_tmp = gr.Dropdown(
choices=["ZH", "EN", "JA", "KO", "YUE"],
value=self.transcriptions_list[i][2],
allow_custom_value=True,
label=i18n("文本语言"),
interactive=True,
)
with gr.Row():
selection_tmp = gr.Checkbox(
label=i18n("选择音频"),
)
else:
with gr.Column(scale=2, min_width=160):
textbox_tmp = gr.Textbox(
label=i18n("Text") + f" {i}",
lines=2,
max_lines=3,
elem_id="subfix_textbox",
interactive=False,
)
with gr.Column(scale=1, min_width=160):
audio_tmp = gr.Audio(
streaming=True,
show_label=False,
show_download_button=False,
interactive=False,
waveform_options=WaveformOptions(show_recording_waveform=False, show_controls=False),
)
with gr.Column(scale=1, min_width=160):
with gr.Group():
with gr.Row():
language_tmp = gr.Dropdown(
choices=["ZH", "EN", "JA", "KO", "YUE"],
value=None,
allow_custom_value=True,
label=i18n("文本语言"),
interactive=False,
)
with gr.Row():
selection_tmp = gr.Checkbox(
label=i18n("选择音频"),
interactive=False,
)
self.textboxes.append(textbox_tmp)
self.audios.append(audio_tmp)
self.languages.append(language_tmp)
self.selections.append(selection_tmp)
with gr.Row(equal_height=True):
with gr.Column(scale=2, min_width=160):
self.close_button = gr.Button(value=i18n("保存并关闭打标WebUI"), variant="stop")
with gr.Column(scale=1, min_width=160):
self.previous_index_button2 = gr.Button(value=i18n("上一页"))
with gr.Column(scale=1, min_width=160):
self.next_index_button2 = gr.Button(value=i18n("下一页"))
# Event Trigger Binding
self.index_slider.release( # Change Index Button
fn=self.submit,
inputs=[
*self.textboxes,
*self.languages,
],
outputs=[],
).success(
fn=self.change_index,
inputs=[
self.index_slider,
],
outputs=[
*self.audios,
*self.textboxes,
*self.languages,
*self.selections,
],
max_batch_size=1,
trigger_mode="once",
)
self.next_index_button1.click( # Next Page Button on the Top
fn=self.submit,
inputs=[
*self.textboxes,
*self.languages,
],
outputs=[],
).success(
fn=self.next_page,
inputs=[
self.index_slider,
],
outputs=[
self.index_slider,
*self.audios,
*self.textboxes,
*self.languages,
*self.selections,
],
scroll_to_output=True,
trigger_mode="once",
)
self.next_index_button2.click( # Next Page Button on the Bottom, Binding to Next Page Button on the Top
lambda: None,
[],
[],
js="""
() => {
document.getElementById("btn_next").click();
}""",
trigger_mode="once",
)
self.previous_index_button1.click( # Previous Page Button on the Top
fn=self.submit,
inputs=[
*self.textboxes,
*self.languages,
],
outputs=[],
).success(
fn=self.previous_page,
inputs=[
self.index_slider,
],
outputs=[
self.index_slider,
*self.audios,
*self.textboxes,
*self.languages,
*self.selections,
],
scroll_to_output=True,
trigger_mode="once",
)
self.previous_index_button2.click( # Previous Page Button on the Bottom, Binding to Previous Page Button on the Top
lambda: None,
[],
[],
js="""
() => {
document.getElementById("btn_previous").click();
}""",
trigger_mode="once",
)
self.delete_audio_button.click( # Delete the Audio in the Transcription File
fn=self.submit,
inputs=[
*self.textboxes,
*self.languages,
],
outputs=[],
).success(
fn=self.delete_audio,
inputs=[
self.index_slider,
*self.selections,
],
outputs=[
self.index_slider,
*self.audios,
*self.textboxes,
*self.languages,
*self.selections,
],
scroll_to_output=True,
).success(
fn=self.submit,
inputs=[
*self.textboxes,
*self.languages,
],
outputs=[],
show_progress="hidden",
)
self.merge_audio_button.click( # Delete the Audio in the Transcription File
fn=self.submit,
inputs=[
*self.textboxes,
*self.languages,
],
outputs=[],
).success(
fn=self.merge_audio,
inputs=[
self.index_slider,
*self.selections,
],
outputs=[
self.index_slider,
*self.audios,
*self.textboxes,
*self.languages,
*self.selections,
],
scroll_to_output=True,
).success(
fn=self.submit,
inputs=[
*self.textboxes,
*self.languages,
],
outputs=[],
show_progress="hidden",
)
if not IS_CLI:
self.close_button.click( # Close the Subfix Tab, Binding to Close Button on Audio Processing Tab
fn=lambda: None,
inputs=[],
outputs=[],
js="""
() => {
document.getElementById("btn_close").click();
}""",
trigger_mode="once",
)
else:
self.close_button.click( # Close the Subfix Tab, Binding to Close Button on Audio Processing Tab
fn=self.submit,
inputs=[
*self.textboxes,
*self.languages,
],
outputs=[],
trigger_mode="once",
).then(
fn=PARTIAL_EXIT,
inputs=[],
outputs=[],
)
def render(self, list_path: str, batch_size: int = 10):
self.batch_size = batch_size
self.transcriptions_list = []
self.load_list(list_path=list_path)
@click.command(name="subfix")
@click.argument(
"list-path",
metavar="<Path>",
type=click.Path(exists=True, dir_okay=False, readable=True, writable=True),
required=True,
)
@click.option(
"--i18n-lang",
type=str,
default="Auto",
help="Languages for internationalisation",
show_default=True,
)
@click.option(
"--port",
type=int,
default="9871",
show_default=True,
)
@click.option(
"--share",
type=bool,
default=False,
show_default=True,
)
def main(list_path: str = "", i18n_lang="Auto", port=9871, share=False):
"""Web-Based audio subtitle editing and multilingual annotation Tool
Accept a transcription list path to launch a Gradio WebUI for text editing
"""
with gr.Blocks(analytics_enabled=False) as app:
subfix = Subfix(I18nAuto(i18n_lang))
subfix.render(list_path=list_path)
if subfix.max_index >= 0:
timer = gr.Timer(0.1)
timer.tick(
fn=lambda: (
gr.Slider(value=0, maximum=subfix.max_index, step=1),
gr.Slider(value=10),
gr.Timer(active=False),
),
inputs=[],
outputs=[
subfix.index_slider,
subfix.batch_size_slider,
timer,
],
)
else:
timer = gr.Timer(2)
timer.tick(
fn=lambda x: (_ for _ in ()).throw(gr.Error("Invalid List")) if x is None else None,
inputs=[],
outputs=[],
)
app.queue().launch(
server_name="0.0.0.0",
inbrowser=True,
share=share,
server_port=port,
quiet=False,
show_api=False,
)
if __name__ == "__main__":
main()

422
tools/subfix_webui.py
View File

@ -1,422 +0,0 @@
import sys
from tools.i18n.i18n import I18nAuto, scan_language_list
language = sys.argv[-1] if sys.argv[-1] in scan_language_list() else "Auto"
i18n = I18nAuto(language=language)
import argparse
import copy
import json
import os
import uuid
try:
import gradio.analytics as analytics
analytics.version_check = lambda: None
except:
...
import gradio as gr
import librosa
import numpy as np
import soundfile
g_json_key_text = ""
g_json_key_path = ""
g_load_file = ""
g_load_format = ""
g_max_json_index = 0
g_index = 0
g_batch = 10
g_text_list = []
g_audio_list = []
g_checkbox_list = []
g_data_json = []
def reload_data(index, batch):
global g_index
g_index = index
global g_batch
g_batch = batch
datas = g_data_json[index : index + batch]
output = []
for d in datas:
output.append({g_json_key_text: d[g_json_key_text], g_json_key_path: d[g_json_key_path]})
return output
def b_change_index(index, batch):
global g_index, g_batch
g_index, g_batch = index, batch
datas = reload_data(index, batch)
output = []
for i, _ in enumerate(datas):
output.append(
# gr.Textbox(
# label=f"Text {i+index}",
# value=_[g_json_key_text]#text
# )
{"__type__": "update", "label": f"Text {i + index}", "value": _[g_json_key_text]}
)
for _ in range(g_batch - len(datas)):
output.append(
# gr.Textbox(
# label=f"Text",
# value=""
# )
{"__type__": "update", "label": "Text", "value": ""}
)
for _ in datas:
output.append(_[g_json_key_path])
for _ in range(g_batch - len(datas)):
output.append(None)
for _ in range(g_batch):
output.append(False)
return output
def b_next_index(index, batch):
b_save_file()
if (index + batch) <= g_max_json_index:
return index + batch, *b_change_index(index + batch, batch)
else:
return index, *b_change_index(index, batch)
def b_previous_index(index, batch):
b_save_file()
if (index - batch) >= 0:
return index - batch, *b_change_index(index - batch, batch)
else:
return 0, *b_change_index(0, batch)
def b_submit_change(*text_list):
global g_data_json
change = False
for i, new_text in enumerate(text_list):
if g_index + i <= g_max_json_index:
new_text = new_text.strip() + " "
if g_data_json[g_index + i][g_json_key_text] != new_text:
g_data_json[g_index + i][g_json_key_text] = new_text
change = True
if change:
b_save_file()
return g_index, *b_change_index(g_index, g_batch)
def b_delete_audio(*checkbox_list):
global g_data_json, g_index, g_max_json_index
b_save_file()
change = False
for i, checkbox in reversed(list(enumerate(checkbox_list))):
if g_index + i < len(g_data_json):
if checkbox == True:
g_data_json.pop(g_index + i)
change = True
g_max_json_index = len(g_data_json) - 1
if g_index > g_max_json_index:
g_index = g_max_json_index
g_index = g_index if g_index >= 0 else 0
if change:
b_save_file()
# return gr.Slider(value=g_index, maximum=(g_max_json_index if g_max_json_index>=0 else 0)), *b_change_index(g_index, g_batch)
return {
"value": g_index,
"__type__": "update",
"maximum": (g_max_json_index if g_max_json_index >= 0 else 0),
}, *b_change_index(g_index, g_batch)
def b_invert_selection(*checkbox_list):
new_list = [not item if item is True else True for item in checkbox_list]
return new_list
def get_next_path(filename):
base_dir = os.path.dirname(filename)
base_name = os.path.splitext(os.path.basename(filename))[0]
for i in range(100):
new_path = os.path.join(base_dir, f"{base_name}_{str(i).zfill(2)}.wav")
if not os.path.exists(new_path):
return new_path
return os.path.join(base_dir, f"{str(uuid.uuid4())}.wav")
def b_audio_split(audio_breakpoint, *checkbox_list):
global g_data_json, g_max_json_index
checked_index = []
for i, checkbox in enumerate(checkbox_list):
if checkbox == True and g_index + i < len(g_data_json):
checked_index.append(g_index + i)
if len(checked_index) == 1:
index = checked_index[0]
audio_json = copy.deepcopy(g_data_json[index])
path = audio_json[g_json_key_path]
data, sample_rate = librosa.load(path, sr=None, mono=True)
audio_maxframe = len(data)
break_frame = int(audio_breakpoint * sample_rate)
if break_frame >= 1 and break_frame < audio_maxframe:
audio_first = data[0:break_frame]
audio_second = data[break_frame:]
nextpath = get_next_path(path)
soundfile.write(nextpath, audio_second, sample_rate)
soundfile.write(path, audio_first, sample_rate)
g_data_json.insert(index + 1, audio_json)
g_data_json[index + 1][g_json_key_path] = nextpath
b_save_file()
g_max_json_index = len(g_data_json) - 1
# return gr.Slider(value=g_index, maximum=g_max_json_index), *b_change_index(g_index, g_batch)
return {"value": g_index, "maximum": g_max_json_index, "__type__": "update"}, *b_change_index(g_index, g_batch)
def b_merge_audio(interval_r, *checkbox_list):
global g_data_json, g_max_json_index
b_save_file()
checked_index = []
audios_path = []
audios_text = []
for i, checkbox in enumerate(checkbox_list):
if checkbox == True and g_index + i < len(g_data_json):
checked_index.append(g_index + i)
if len(checked_index) > 1:
for i in checked_index:
audios_path.append(g_data_json[i][g_json_key_path])
audios_text.append(g_data_json[i][g_json_key_text])
for i in reversed(checked_index[1:]):
g_data_json.pop(i)
base_index = checked_index[0]
base_path = audios_path[0]
g_data_json[base_index][g_json_key_text] = "".join(audios_text)
audio_list = []
l_sample_rate = None
for i, path in enumerate(audios_path):
data, sample_rate = librosa.load(path, sr=l_sample_rate, mono=True)
l_sample_rate = sample_rate
if i > 0:
silence = np.zeros(int(l_sample_rate * interval_r))
audio_list.append(silence)
audio_list.append(data)
audio_concat = np.concatenate(audio_list)
soundfile.write(base_path, audio_concat, l_sample_rate)
b_save_file()
g_max_json_index = len(g_data_json) - 1
# return gr.Slider(value=g_index, maximum=g_max_json_index), *b_change_index(g_index, g_batch)
return {"value": g_index, "maximum": g_max_json_index, "__type__": "update"}, *b_change_index(g_index, g_batch)
def b_save_json():
with open(g_load_file, "w", encoding="utf-8") as file:
for data in g_data_json:
file.write(f"{json.dumps(data, ensure_ascii=False)}\n")
def b_save_list():
with open(g_load_file, "w", encoding="utf-8") as file:
for data in g_data_json:
wav_path = data["wav_path"]
speaker_name = data["speaker_name"]
language = data["language"]
text = data["text"]
file.write(f"{wav_path}|{speaker_name}|{language}|{text}".strip() + "\n")
def b_load_json():
global g_data_json, g_max_json_index
with open(g_load_file, "r", encoding="utf-8") as file:
g_data_json = file.readlines()
g_data_json = [json.loads(line) for line in g_data_json]
g_max_json_index = len(g_data_json) - 1
def b_load_list():
global g_data_json, g_max_json_index
with open(g_load_file, "r", encoding="utf-8") as source:
data_list = source.readlines()
for _ in data_list:
data = _.split("|")
if len(data) == 4:
wav_path, speaker_name, language, text = data
g_data_json.append(
{"wav_path": wav_path, "speaker_name": speaker_name, "language": language, "text": text.strip()}
)
else:
print("error line:", data)
g_max_json_index = len(g_data_json) - 1
def b_save_file():
if g_load_format == "json":
b_save_json()
elif g_load_format == "list":
b_save_list()
def b_load_file():
if g_load_format == "json":
b_load_json()
elif g_load_format == "list":
b_load_list()
def set_global(load_json, load_list, json_key_text, json_key_path, batch):
global g_json_key_text, g_json_key_path, g_load_file, g_load_format, g_batch
g_batch = int(batch)
if load_json != "None":
g_load_format = "json"
g_load_file = load_json
elif load_list != "None":
g_load_format = "list"
g_load_file = load_list
else:
g_load_format = "list"
g_load_file = "demo.list"
g_json_key_text = json_key_text
g_json_key_path = json_key_path
b_load_file()
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Process some integers.")
parser.add_argument("--load_json", default="None", help="source file, like demo.json")
parser.add_argument("--is_share", default="False", help="whether webui is_share=True")
parser.add_argument("--load_list", default="None", help="source file, like demo.list")
parser.add_argument("--webui_port_subfix", default=9871, help="source file, like demo.list")
parser.add_argument("--json_key_text", default="text", help="the text key name in json, Default: text")
parser.add_argument("--json_key_path", default="wav_path", help="the path key name in json, Default: wav_path")
parser.add_argument("--g_batch", default=10, help="max number g_batch wav to display, Default: 10")
args = parser.parse_args()
set_global(args.load_json, args.load_list, args.json_key_text, args.json_key_path, args.g_batch)
with gr.Blocks(analytics_enabled=False) as demo:
gr.Markdown(
value=i18n("Submit Text: 将当前页所有文本框内容手工保存到内存和文件(翻页前后或者退出标注页面前如果没点这个按钮，你再翻回来就回滚了，白忙活。)")
)
with gr.Row():
btn_change_index = gr.Button("Change Index")
btn_submit_change = gr.Button("Submit Text")
btn_merge_audio = gr.Button("Merge Audio")
btn_delete_audio = gr.Button("Delete Audio")
btn_previous_index = gr.Button("Previous Index")
btn_next_index = gr.Button("Next Index")
with gr.Row():
index_slider = gr.Slider(minimum=0, maximum=g_max_json_index, value=g_index, step=1, label="Index", scale=3)
splitpoint_slider = gr.Slider(
minimum=0, maximum=120.0, value=0, step=0.1, label="Audio Split Point(s)", scale=3
)
btn_audio_split = gr.Button("Split Audio", scale=1)
btn_save_json = gr.Button("Save File", visible=True, scale=1)
btn_invert_selection = gr.Button("Invert Selection", scale=1)
with gr.Row():
with gr.Column():
for _ in range(0, g_batch):
with gr.Row():
text = gr.Textbox(label="Text", visible=True, scale=5)
audio_output = gr.Audio(label="Output Audio", visible=True, scale=5)
audio_check = gr.Checkbox(label="Yes", show_label=True, info="Choose Audio", scale=1)
g_text_list.append(text)
g_audio_list.append(audio_output)
g_checkbox_list.append(audio_check)
with gr.Row():
batchsize_slider = gr.Slider(
minimum=1, maximum=g_batch, value=g_batch, step=1, label="Batch Size", scale=3, interactive=False
)
interval_slider = gr.Slider(minimum=0, maximum=2, value=0, step=0.01, label="Interval", scale=3)
btn_theme_dark = gr.Button("Light Theme", link="?__theme=light", scale=1)
btn_theme_light = gr.Button("Dark Theme", link="?__theme=dark", scale=1)
btn_change_index.click(
b_change_index,
inputs=[
index_slider,
batchsize_slider,
],
outputs=[*g_text_list, *g_audio_list, *g_checkbox_list],
)
btn_submit_change.click(
b_submit_change,
inputs=[
*g_text_list,
],
outputs=[index_slider, *g_text_list, *g_audio_list, *g_checkbox_list],
)
btn_previous_index.click(
b_previous_index,
inputs=[
index_slider,
batchsize_slider,
],
outputs=[index_slider, *g_text_list, *g_audio_list, *g_checkbox_list],
)
btn_next_index.click(
b_next_index,
inputs=[
index_slider,
batchsize_slider,
],
outputs=[index_slider, *g_text_list, *g_audio_list, *g_checkbox_list],
)
btn_delete_audio.click(
b_delete_audio,
inputs=[*g_checkbox_list],
outputs=[index_slider, *g_text_list, *g_audio_list, *g_checkbox_list],
)
btn_merge_audio.click(
b_merge_audio,
inputs=[interval_slider, *g_checkbox_list],
outputs=[index_slider, *g_text_list, *g_audio_list, *g_checkbox_list],
)
btn_audio_split.click(
b_audio_split,
inputs=[splitpoint_slider, *g_checkbox_list],
outputs=[index_slider, *g_text_list, *g_audio_list, *g_checkbox_list],
)
btn_invert_selection.click(b_invert_selection, inputs=[*g_checkbox_list], outputs=[*g_checkbox_list])
btn_save_json.click(b_save_file)
demo.load(
b_change_index,
inputs=[
index_slider,
batchsize_slider,
],
outputs=[*g_text_list, *g_audio_list, *g_checkbox_list],
)
demo.launch(
server_name="0.0.0.0",
inbrowser=True,
# quiet=True,
share=eval(args.is_share),
server_port=int(args.webui_port_subfix),
)

4
tools/uvr5/mdxnet.py
View File

@ -190,14 +190,14 @@ class Predictor:
opt_path_vocal = path_vocal[:-4] + ".%s" % format
opt_path_other = path_other[:-4] + ".%s" % format
if os.path.exists(path_vocal):
os.system("ffmpeg -i \"%s\" -vn \"%s\" -q:a 2 -y" % (path_vocal, opt_path_vocal))
os.system('ffmpeg -i "%s" -vn "%s" -q:a 2 -y' % (path_vocal, opt_path_vocal))
if os.path.exists(opt_path_vocal):
try:
os.remove(path_vocal)
except:
pass
if os.path.exists(path_other):
os.system("ffmpeg -i \"%s\" -vn \"%s\" -q:a 2 -y" % (path_other, opt_path_other))
os.system('ffmpeg -i "%s" -vn "%s" -q:a 2 -y' % (path_other, opt_path_other))
if os.path.exists(opt_path_other):
try:
os.remove(path_other)

8
tools/uvr5/vr.py
View File

@ -140,7 +140,7 @@ class AudioPre:
)
if os.path.exists(path):
opt_format_path = path[:-4] + ".%s" % format
cmd="ffmpeg -i \"%s\" -vn \"%s\" -q:a 2 -y" % (path, opt_format_path)
cmd = 'ffmpeg -i "%s" -vn "%s" -q:a 2 -y' % (path, opt_format_path)
print(cmd)
os.system(cmd)
if os.path.exists(opt_format_path):
@ -177,7 +177,7 @@ class AudioPre:
)
if os.path.exists(path):
opt_format_path = path[:-4] + ".%s" % format
cmd="ffmpeg -i \"%s\" -vn \"%s\" -q:a 2 -y" % (path, opt_format_path)
cmd = 'ffmpeg -i "%s" -vn "%s" -q:a 2 -y' % (path, opt_format_path)
print(cmd)
os.system(cmd)
if os.path.exists(opt_format_path):
@ -307,7 +307,7 @@ class AudioPreDeEcho:
)
if os.path.exists(path):
opt_format_path = path[:-4] + ".%s" % format
cmd="ffmpeg -i \"%s\" -vn \"%s\" -q:a 2 -y" % (path, opt_format_path)
cmd = 'ffmpeg -i "%s" -vn "%s" -q:a 2 -y' % (path, opt_format_path)
print(cmd)
os.system(cmd)
if os.path.exists(opt_format_path):
@ -340,7 +340,7 @@ class AudioPreDeEcho:
)
if os.path.exists(path):
opt_format_path = path[:-4] + ".%s" % format
cmd="ffmpeg -i \"%s\" -vn \"%s\" -q:a 2 -y" % (path, opt_format_path)
cmd = 'ffmpeg -i "%s" -vn "%s" -q:a 2 -y' % (path, opt_format_path)
print(cmd)
os.system(cmd)
if os.path.exists(opt_format_path):

21
tools/uvr5/webui.py
View File

@ -1,23 +1,22 @@
import logging
import os
import sys
import traceback
import gradio as gr
from tools.i18n.i18n import I18nAuto
from tools.my_utils import clean_path
i18n = I18nAuto()
logger = logging.getLogger(__name__)
import sys
import ffmpeg
import gradio as gr
import torch
from bsroformer import Roformer_Loader
from mdxnet import MDXNetDereverb
from vr import AudioPre, AudioPreDeEcho
from tools.i18n.i18n import I18nAuto
from tools.my_utils import clean_path, load_cudnn
i18n = I18nAuto()
logger = logging.getLogger(__name__)
weight_uvr5_root = "tools/uvr5/uvr5_weights"
uvr5_names = []
for name in os.listdir(weight_uvr5_root):
@ -44,6 +43,7 @@ def html_center(text, label="p"):
def uvr(model_name, inp_root, save_root_vocal, paths, save_root_ins, agg, format0):
infos = []
load_cudnn()
try:
inp_root = clean_path(inp_root)
save_root_vocal = clean_path(save_root_vocal)
@ -220,5 +220,6 @@ app.queue().launch( # concurrency_count=511, max_size=1022
inbrowser=True,
share=is_share,
server_port=webui_port_uvr5,
show_api=False,
# quiet=True,
)

26
webui.py
View File

@ -12,6 +12,7 @@ import platform
import shutil
import signal
import gradio as gr
import psutil
import torch
import yaml
@ -58,6 +59,7 @@ for site_packages_root in site_packages_roots:
traceback.print_exc()
import shutil
import subprocess
from multiprocessing import cpu_count
from subprocess import Popen
from tools.assets import css, js, top_html
@ -66,7 +68,6 @@ from tools.i18n.i18n import I18nAuto, scan_language_list
language = sys.argv[-1] if sys.argv[-1] in scan_language_list() else "Auto"
os.environ["language"] = language
i18n = I18nAuto(language=language)
from multiprocessing import cpu_count
from config import (
GPU_INDEX,
@ -86,14 +87,9 @@ from config import (
from tools import my_utils
from tools.my_utils import check_details, check_for_existance
# os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' # 当遇到mps不支持的步骤时使用cpu
try:
import gradio.analytics as analytics
analytics.version_check = lambda: None
except:
...
import gradio as gr
language = sys.argv[-1] if sys.argv[-1] in scan_language_list() else "Auto"
os.environ["language"] = language
i18n = I18nAuto(language=language)
n_cpu = cpu_count()
@ -276,12 +272,7 @@ def change_label(path_list):
if p_label is None:
check_for_existance([path_list])
path_list = my_utils.clean_path(path_list)
cmd = '"%s" -s tools/subfix_webui.py --load_list "%s" --webui_port %s --is_share %s' % (
python_exec,
path_list,
webui_port_subfix,
is_share,
)
cmd = f'"{python_exec}" -s tools/subfix.py --i18n-lang {language} --port {webui_port_subfix} --share {is_share} "{path_list}"'
yield (
process_info(process_name_subfix, "opened"),
{"__type__": "update", "visible": False},
@ -507,7 +498,7 @@ def open1Ba(
):
global p_train_SoVITS
if p_train_SoVITS == None:
exp_name=exp_name.rstrip(" ")
exp_name = exp_name.rstrip(" ")
config_file = (
"GPT_SoVITS/configs/s2.json"
if version not in {"v2Pro", "v2ProPlus"}
@ -604,7 +595,7 @@ def open1Bb(
):
global p_train_GPT
if p_train_GPT == None:
exp_name=exp_name.rstrip(" ")
exp_name = exp_name.rstrip(" ")
with open(
"GPT_SoVITS/configs/s1longer.yaml" if version == "v1" else "GPT_SoVITS/configs/s1longer-v2.yaml"
) as f:
@ -1981,5 +1972,6 @@ with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False, js=js, css=css
inbrowser=True,
share=is_share,
server_port=webui_port_main,
show_api=False,
# quiet=True,
)