blackteay/GPT-SoVITS
1
0
Fork 0
mirror of https://github.com/RVC-Boss/GPT-SoVITS.git synced 2026-01-11 04:46:57 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'RVC-Boss:main' into main

Browse Source
This commit is contained in:
符玄 2024-09-15 14:41:45 +08:00 committed by GitHub
commit bafd1bf632
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
13 changed files with 90 additions and 54 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
GPT_SoVITS/AR/models/t2s_model.py
View File

@ -854,6 +854,7 @@ class Text2SemanticDecoder(nn.Module):
if idx == 0: if idx == 0:
xy_attn_mask = None xy_attn_mask = None
if(idx<11):###至少预测出10个token不然不给停止（0.4s）
logits = logits[:, :-1] logits = logits[:, :-1]
samples = sample( samples = sample(

11
GPT_SoVITS/TTS_infer_pack/TTS.py
View File

@ -213,6 +213,10 @@ class TTS_Config:
"cnhuhbert_base_path": self.cnhuhbert_base_path, "cnhuhbert_base_path": self.cnhuhbert_base_path,
} }
return self.config return self.config
def update_version(self, version:str)->None:
self.version = version
self.languages = self.v2_languages if self.version=="v2" else self.v1_languages
def __str__(self): def __str__(self):
self.configs = self.update_configs() self.configs = self.update_configs()
@ -300,13 +304,14 @@ class TTS:
def init_vits_weights(self, weights_path: str): def init_vits_weights(self, weights_path: str):
print(f"Loading VITS weights from {weights_path}") print(f"Loading VITS weights from {weights_path}")
self.configs.vits_weights_path = weights_path self.configs.vits_weights_path = weights_path
self.configs.save_configs()
dict_s2 = torch.load(weights_path, map_location=self.configs.device) dict_s2 = torch.load(weights_path, map_location=self.configs.device)
hps = dict_s2["config"] hps = dict_s2["config"]
if dict_s2['weight']['enc_p.text_embedding.weight'].shape[0] == 322: if dict_s2['weight']['enc_p.text_embedding.weight'].shape[0] == 322:
self.configs.version = "v1" self.configs.update_version("v1")
else: else:
self.configs.version = "v2" self.configs.update_version("v2")
self.configs.save_configs()
hps["model"]["version"] = self.configs.version hps["model"]["version"] = self.configs.version
self.configs.filter_length = hps["data"]["filter_length"] self.configs.filter_length = hps["data"]["filter_length"]
self.configs.segment_size = hps["train"]["segment_size"] self.configs.segment_size = hps["train"]["segment_size"]

27
GPT_SoVITS/module/models_onnx.py
View File

@ -13,7 +13,9 @@ from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
from module.commons import init_weights, get_padding from module.commons import init_weights, get_padding
from module.mrte_model import MRTE from module.mrte_model import MRTE
from module.quantize import ResidualVectorQuantizer from module.quantize import ResidualVectorQuantizer
from text import symbols # from text import symbols
from text import symbols as symbols_v1
from text import symbols2 as symbols_v2
from torch.cuda.amp import autocast from torch.cuda.amp import autocast
@ -182,6 +184,7 @@ class TextEncoder(nn.Module):
kernel_size, kernel_size,
p_dropout, p_dropout,
latent_channels=192, latent_channels=192,
version="v2",
): ):
super().__init__() super().__init__()
self.out_channels = out_channels self.out_channels = out_channels
@ -192,6 +195,7 @@ class TextEncoder(nn.Module):
self.kernel_size = kernel_size self.kernel_size = kernel_size
self.p_dropout = p_dropout self.p_dropout = p_dropout
self.latent_channels = latent_channels self.latent_channels = latent_channels
self.version = version
self.ssl_proj = nn.Conv1d(768, hidden_channels, 1) self.ssl_proj = nn.Conv1d(768, hidden_channels, 1)
@ -207,6 +211,11 @@ class TextEncoder(nn.Module):
self.encoder_text = attentions.Encoder( self.encoder_text = attentions.Encoder(
hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
) )
if self.version == "v1":
symbols = symbols_v1.symbols
else:
symbols = symbols_v2.symbols
self.text_embedding = nn.Embedding(len(symbols), hidden_channels) self.text_embedding = nn.Embedding(len(symbols), hidden_channels)
self.mrte = MRTE() self.mrte = MRTE()
@ -817,6 +826,7 @@ class SynthesizerTrn(nn.Module):
use_sdp=True, use_sdp=True,
semantic_frame_rate=None, semantic_frame_rate=None,
freeze_quantizer=None, freeze_quantizer=None,
version="v2",
**kwargs **kwargs
): ):
super().__init__() super().__init__()
@ -837,6 +847,7 @@ class SynthesizerTrn(nn.Module):
self.segment_size = segment_size self.segment_size = segment_size
self.n_speakers = n_speakers self.n_speakers = n_speakers
self.gin_channels = gin_channels self.gin_channels = gin_channels
self.version = version
self.use_sdp = use_sdp self.use_sdp = use_sdp
self.enc_p = TextEncoder( self.enc_p = TextEncoder(
@ -847,6 +858,7 @@ class SynthesizerTrn(nn.Module):
n_layers, n_layers,
kernel_size, kernel_size,
p_dropout, p_dropout,
version=version,
) )
self.dec = Generator( self.dec = Generator(
inter_channels, inter_channels,
@ -871,9 +883,11 @@ class SynthesizerTrn(nn.Module):
inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels
) )
self.ref_enc = modules.MelStyleEncoder( # self.version=os.environ.get("version","v1")
spec_channels, style_vector_dim=gin_channels if self.version == "v1":
) self.ref_enc = modules.MelStyleEncoder(spec_channels, style_vector_dim=gin_channels)
else:
self.ref_enc = modules.MelStyleEncoder(704, style_vector_dim=gin_channels)
ssl_dim = 768 ssl_dim = 768
self.ssl_dim = ssl_dim self.ssl_dim = ssl_dim
@ -894,7 +908,10 @@ class SynthesizerTrn(nn.Module):
def forward(self, codes, text, refer): def forward(self, codes, text, refer):
refer_mask = torch.ones_like(refer[:1,:1,:]) refer_mask = torch.ones_like(refer[:1,:1,:])
ge = self.ref_enc(refer * refer_mask, refer_mask) if (self.version == "v1"):
ge = self.ref_enc(refer * refer_mask, refer_mask)
else:
ge = self.ref_enc(refer[:, :704] * refer_mask, refer_mask)
quantized = self.quantizer.decode(codes) quantized = self.quantizer.decode(codes)
if self.semantic_frame_rate == "25hz": if self.semantic_frame_rate == "25hz":

62
GPT_SoVITS/onnx_export.py
View File

@ -1,11 +1,12 @@
from module.models_onnx import SynthesizerTrn, symbols from module.models_onnx import SynthesizerTrn, symbols_v1, symbols_v2
from AR.models.t2s_lightning_module_onnx import Text2SemanticLightningModule from AR.models.t2s_lightning_module_onnx import Text2SemanticLightningModule
import torch import torch
import torchaudio import torchaudio
from torch import nn from torch import nn
from feature_extractor import cnhubert from feature_extractor import cnhubert
cnhubert_base_path = "pretrained_models/chinese-hubert-base"
cnhubert.cnhubert_base_path=cnhubert_base_path cnhubert_base_path = "GPT_SoVITS/pretrained_models/chinese-hubert-base"
cnhubert.cnhubert_base_path = cnhubert_base_path
ssl_model = cnhubert.get_model() ssl_model = cnhubert.get_model()
from text import cleaned_text_to_sequence from text import cleaned_text_to_sequence
import soundfile import soundfile
@ -196,6 +197,11 @@ class VitsModel(nn.Module):
super().__init__() super().__init__()
dict_s2 = torch.load(vits_path,map_location="cpu") dict_s2 = torch.load(vits_path,map_location="cpu")
self.hps = dict_s2["config"] self.hps = dict_s2["config"]
if dict_s2['weight']['enc_p.text_embedding.weight'].shape[0] == 322:
self.hps["model"]["version"] = "v1"
else:
self.hps["model"]["version"] = "v2"
self.hps = DictToAttrRecursive(self.hps) self.hps = DictToAttrRecursive(self.hps)
self.hps.model.semantic_frame_rate = "25hz" self.hps.model.semantic_frame_rate = "25hz"
self.vq_model = SynthesizerTrn( self.vq_model = SynthesizerTrn(
@ -267,13 +273,13 @@ class SSLModel(nn.Module):
return self.ssl.model(ref_audio_16k)["last_hidden_state"].transpose(1, 2) return self.ssl.model(ref_audio_16k)["last_hidden_state"].transpose(1, 2)
def export(vits_path, gpt_path, project_name): def export(vits_path, gpt_path, project_name, vits_model="v2"):
vits = VitsModel(vits_path) vits = VitsModel(vits_path)
gpt = T2SModel(gpt_path, vits) gpt = T2SModel(gpt_path, vits)
gpt_sovits = GptSoVits(vits, gpt) gpt_sovits = GptSoVits(vits, gpt)
ssl = SSLModel() ssl = SSLModel()
ref_seq = torch.LongTensor([cleaned_text_to_sequence(["n", "i2", "h", "ao3", ",", "w", "o3", "sh", "i4", "b", "ai2", "y", "e4"])]) ref_seq = torch.LongTensor([cleaned_text_to_sequence(["n", "i2", "h", "ao3", ",", "w", "o3", "sh", "i4", "b", "ai2", "y", "e4"],version=vits_model)])
text_seq = torch.LongTensor([cleaned_text_to_sequence(["w", "o3", "sh", "i4", "b", "ai2", "y", "e4", "w", "o3", "sh", "i4", "b", "ai2", "y", "e4", "w", "o3", "sh", "i4", "b", "ai2", "y", "e4"])]) text_seq = torch.LongTensor([cleaned_text_to_sequence(["w", "o3", "sh", "i4", "b", "ai2", "y", "e4", "w", "o3", "sh", "i4", "b", "ai2", "y", "e4", "w", "o3", "sh", "i4", "b", "ai2", "y", "e4"],version=vits_model)])
ref_bert = torch.randn((ref_seq.shape[1], 1024)).float() ref_bert = torch.randn((ref_seq.shape[1], 1024)).float()
text_bert = torch.randn((text_seq.shape[1], 1024)).float() text_bert = torch.randn((text_seq.shape[1], 1024)).float()
ref_audio = torch.randn((1, 48000 * 5)).float() ref_audio = torch.randn((1, 48000 * 5)).float()
@ -287,34 +293,38 @@ def export(vits_path, gpt_path, project_name):
pass pass
ssl_content = ssl(ref_audio_16k).float() ssl_content = ssl(ref_audio_16k).float()
debug = False # debug = False
debug = True
# gpt_sovits.export(ref_seq, text_seq, ref_bert, text_bert, ref_audio_sr, ssl_content, project_name)
if debug: if debug:
a, b = gpt_sovits(ref_seq, text_seq, ref_bert, text_bert, ref_audio_sr, ssl_content, debug=debug) a, b = gpt_sovits(ref_seq, text_seq, ref_bert, text_bert, ref_audio_sr, ssl_content, debug=debug)
soundfile.write("out1.wav", a.cpu().detach().numpy(), vits.hps.data.sampling_rate) soundfile.write("out1.wav", a.cpu().detach().numpy(), vits.hps.data.sampling_rate)
soundfile.write("out2.wav", b[0], vits.hps.data.sampling_rate) soundfile.write("out2.wav", b[0], vits.hps.data.sampling_rate)
return else:
a = gpt_sovits(ref_seq, text_seq, ref_bert, text_bert, ref_audio_sr, ssl_content).detach().cpu().numpy()
a = gpt_sovits(ref_seq, text_seq, ref_bert, text_bert, ref_audio_sr, ssl_content).detach().cpu().numpy() soundfile.write("out.wav", a, vits.hps.data.sampling_rate)
soundfile.write("out.wav", a, vits.hps.data.sampling_rate) if vits_model == "v1":
symbols = symbols_v1
gpt_sovits.export(ref_seq, text_seq, ref_bert, text_bert, ref_audio_sr, ssl_content, project_name) else:
symbols = symbols_v2
MoeVSConf = { MoeVSConf = {
"Folder" : f"{project_name}", "Folder": f"{project_name}",
"Name" : f"{project_name}", "Name": f"{project_name}",
"Type" : "GPT-SoVits", "Type": "GPT-SoVits",
"Rate" : vits.hps.data.sampling_rate, "Rate": vits.hps.data.sampling_rate,
"NumLayers": gpt.t2s_model.num_layers, "NumLayers": gpt.t2s_model.num_layers,
"EmbeddingDim": gpt.t2s_model.embedding_dim, "EmbeddingDim": gpt.t2s_model.embedding_dim,
"Dict": "BasicDict", "Dict": "BasicDict",
"BertPath": "chinese-roberta-wwm-ext-large", "BertPath": "chinese-roberta-wwm-ext-large",
"Symbol": symbols, # "Symbol": symbols,
"AddBlank": False "AddBlank": False,
} }
MoeVSConfJson = json.dumps(MoeVSConf) MoeVSConfJson = json.dumps(MoeVSConf)
with open(f"onnx/{project_name}.json", 'w') as MoeVsConfFile: with open(f"onnx/{project_name}.json", 'w') as MoeVsConfFile:
json.dump(MoeVSConf, MoeVsConfFile, indent = 4) json.dump(MoeVSConf, MoeVsConfFile, indent = 4)

4
GPT_SoVITS/text/g2pw/polyphonic-fix.rep
View File

@ -45021,4 +45021,6 @@
黄冠野服: ['huang2', 'guan4', 'ye3', 'fu2'] 黄冠野服: ['huang2', 'guan4', 'ye3', 'fu2']
黄发台背: ['huang2', 'fa1', 'tai2', 'bei4'] 黄发台背: ['huang2', 'fa1', 'tai2', 'bei4']
鼎铛玉石: ['ding3', 'cheng1', 'yu4', 'shi2'] 鼎铛玉石: ['ding3', 'cheng1', 'yu4', 'shi2']
齿豁头童: ['chi3', 'huo1', 'tou2', 'tong2'] 齿豁头童: ['chi3', 'huo1', 'tou2', 'tong2']
牦牛: ['mao2', 'niu2']
牦: ['mao2']

BIN
GPT_SoVITS/text/g2pw/polyphonic.pickle
View File

Binary file not shown.

6
README.md
View File

@ -184,8 +184,8 @@ D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
#### Integrated Package Users #### Integrated Package Users
Double-click `go-webui.bat`or use `go-webui.ps` 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.ps` if you want to switch to V1,then double-click`go-webui-v1.bat` or use `go-webui-v1.ps1`
#### Others #### Others
@ -220,7 +220,7 @@ Or maunally switch version in WebUI
#### Integrated Package Users #### Integrated Package Users
Double-click `go-webui-v2.bat` or use `go-webui-v2.ps` ,then open the inference webui at `1-GPT-SoVITS-TTS/1C-inference` Double-click `go-webui-v2.bat` or use `go-webui-v2.ps1` ,then open the inference webui at `1-GPT-SoVITS-TTS/1C-inference`
#### Others #### Others

6
api_v2.py
View File

@ -253,13 +253,13 @@ def check_params(req:dict):
if (text_lang in [None, ""]) : if (text_lang in [None, ""]) :
return JSONResponse(status_code=400, content={"message": "text_lang is required"}) return JSONResponse(status_code=400, content={"message": "text_lang is required"})
elif text_lang.lower() not in tts_config.languages: elif text_lang.lower() not in tts_config.languages:
return JSONResponse(status_code=400, content={"message": "text_lang is not supported"}) return JSONResponse(status_code=400, content={"message": f"text_lang: {text_lang} is not supported in version {tts_config.version}"})
if (prompt_lang in [None, ""]) : if (prompt_lang in [None, ""]) :
return JSONResponse(status_code=400, content={"message": "prompt_lang is required"}) return JSONResponse(status_code=400, content={"message": "prompt_lang is required"})
elif prompt_lang.lower() not in tts_config.languages: elif prompt_lang.lower() not in tts_config.languages:
return JSONResponse(status_code=400, content={"message": "prompt_lang is not supported"}) return JSONResponse(status_code=400, content={"message": f"prompt_lang: {prompt_lang} is not supported in version {tts_config.version}"})
if media_type not in ["wav", "raw", "ogg", "aac"]: if media_type not in ["wav", "raw", "ogg", "aac"]:
return JSONResponse(status_code=400, content={"message": "media_type is not supported"}) return JSONResponse(status_code=400, content={"message": f"media_type: {media_type} is not supported"})
elif media_type == "ogg" and not streaming_mode: elif media_type == "ogg" and not streaming_mode:
return JSONResponse(status_code=400, content={"message": "ogg format is not supported in non-streaming mode"}) return JSONResponse(status_code=400, content={"message": "ogg format is not supported in non-streaming mode"})

6
docs/cn/README.md
View File

@ -181,8 +181,8 @@ D:\GPT-SoVITS\xxx/xxx.wav|xxx|zh|我爱玩原神。
#### 整合包用户 #### 整合包用户
双击`go-webui.bat`或者使用`go-webui.ps` 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.ps` 若想使用V1,则双击`go-webui-v1.bat`或者使用`go-webui-v1.ps1`
#### 其他 #### 其他
@ -217,7 +217,7 @@ python webui.py v1 <language(optional)>
#### 整合包用户 #### 整合包用户
双击 `go-webui.bat` 或者使用 `go-webui.ps` ,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 #### 其他

6
docs/ja/README.md
View File

@ -171,8 +171,8 @@ D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
#### 統合パッケージ利用者 #### 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.ps`を使用します。 `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.ps`を使用してください。 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 #### その他
@ -207,7 +207,7 @@ python webui.py v1 <言語(オプション)>
#### 統合パッケージ利用者 #### 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.ps`を使用して、`1-GPT-SoVITS-TTS/1C-inference`で推論webuiを開きます。 `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。
#### その他 #### その他

6
docs/ko/README.md
View File

@ -175,8 +175,8 @@ D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
#### 통합 패키지 사용자 #### 통합 패키지 사용자
`go-webui.bat`을 더블 클릭하거나 `go-webui.ps`를 사용하십시오. `go-webui.bat`을 더블 클릭하거나 `go-webui.ps1`를 사용하십시오.
V1으로 전환하려면, `go-webui-v1.bat`을 더블 클릭하거나 `go-webui-v1.ps`를 사용하십시오. V1으로 전환하려면, `go-webui-v1.bat`을 더블 클릭하거나 `go-webui-v1.ps1`를 사용하십시오.
#### 기타 #### 기타
@ -211,7 +211,7 @@ python webui.py v1 <언어(옵션)>
#### 통합 패키지 사용자 #### 통합 패키지 사용자
`go-webui-v2.bat`을 더블 클릭하거나 `go-webui-v2.ps`를 사용한 다음 `1-GPT-SoVITS-TTS/1C-inference`에서 추론 webui를 엽니다. `go-webui-v2.bat`을 더블 클릭하거나 `go-webui-v2.ps1`를 사용한 다음 `1-GPT-SoVITS-TTS/1C-inference`에서 추론 webui를 엽니다.
#### 기타 #### 기타

8
docs/tr/README.md
View File

@ -172,8 +172,8 @@ D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
#### Entegre Paket Kullanıcıları #### Entegre Paket Kullanıcıları
`go-webui.bat` dosyasına çift tıklayın veya `go-webui.ps` kullanın. `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.ps` kullanın. V1'e geçmek istiyorsanız, `go-webui-v1.bat` dosyasına çift tıklayın veya `go-webui-v1.ps1` kullanın.
#### Diğerleri #### Diğerleri
@ -208,7 +208,7 @@ veya WebUI'de manuel olarak sürüm değiştirin.
#### Entegre Paket Kullanıcıları #### Entegre Paket Kullanıcıları
`go-webui-v2.bat` dosyasına çift tıklayın veya `go-webui-v2.ps` kullanın, ardından çıkarım webui'sini `1-GPT-SoVITS-TTS/1C-inference` adresinde açın. `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.
#### Diğerleri #### Diğerleri
@ -330,4 +330,4 @@ python ./tools/asr/fasterwhisper_asr.py -i <girdi> -o <çıktı> -l <dil>
<a href="https://github.com/RVC-Boss/GPT-SoVITS/graphs/contributors" target="_blank"> <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" />
</a> </a>

1
requirements.txt
View File

@ -33,3 +33,4 @@ ko_pron
opencc; sys_platform != 'linux' opencc; sys_platform != 'linux'
opencc==1.1.1; sys_platform == 'linux' opencc==1.1.1; sys_platform == 'linux'
python_mecab_ko; sys_platform != 'win32' python_mecab_ko; sys_platform != 'win32'
fastapi<0.112.2