mirror of
https://github.com/RVC-Boss/GPT-SoVITS.git
synced 2025-04-06 03:57:44 +08:00
Merge branch 'main' into main
This commit is contained in:
Harry C
GitHub
commit
0c8516422a
@ -854,6 +854,7 @@ class Text2SemanticDecoder(nn.Module):
|
||||
|
||||
if idx == 0:
|
||||
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]
|
||||
|
||||
samples = sample(
|
||||
|
||||
@ -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.mrte_model import MRTE
|
||||
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
|
||||
|
||||
|
||||
@ -182,6 +184,7 @@ class TextEncoder(nn.Module):
|
||||
kernel_size,
|
||||
p_dropout,
|
||||
latent_channels=192,
|
||||
version="v2",
|
||||
):
|
||||
super().__init__()
|
||||
self.out_channels = out_channels
|
||||
@ -192,6 +195,7 @@ class TextEncoder(nn.Module):
|
||||
self.kernel_size = kernel_size
|
||||
self.p_dropout = p_dropout
|
||||
self.latent_channels = latent_channels
|
||||
self.version = version
|
||||
|
||||
self.ssl_proj = nn.Conv1d(768, hidden_channels, 1)
|
||||
|
||||
@ -207,6 +211,11 @@ class TextEncoder(nn.Module):
|
||||
self.encoder_text = attentions.Encoder(
|
||||
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.mrte = MRTE()
|
||||
@ -817,6 +826,7 @@ class SynthesizerTrn(nn.Module):
|
||||
use_sdp=True,
|
||||
semantic_frame_rate=None,
|
||||
freeze_quantizer=None,
|
||||
version="v2",
|
||||
**kwargs
|
||||
):
|
||||
super().__init__()
|
||||
@ -837,6 +847,7 @@ class SynthesizerTrn(nn.Module):
|
||||
self.segment_size = segment_size
|
||||
self.n_speakers = n_speakers
|
||||
self.gin_channels = gin_channels
|
||||
self.version = version
|
||||
|
||||
self.use_sdp = use_sdp
|
||||
self.enc_p = TextEncoder(
|
||||
@ -847,6 +858,7 @@ class SynthesizerTrn(nn.Module):
|
||||
n_layers,
|
||||
kernel_size,
|
||||
p_dropout,
|
||||
version=version,
|
||||
)
|
||||
self.dec = Generator(
|
||||
inter_channels,
|
||||
@ -871,9 +883,11 @@ class SynthesizerTrn(nn.Module):
|
||||
inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels
|
||||
)
|
||||
|
||||
self.ref_enc = modules.MelStyleEncoder(
|
||||
spec_channels, style_vector_dim=gin_channels
|
||||
)
|
||||
# self.version=os.environ.get("version","v1")
|
||||
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
|
||||
self.ssl_dim = ssl_dim
|
||||
@ -894,7 +908,10 @@ class SynthesizerTrn(nn.Module):
|
||||
|
||||
def forward(self, codes, text, refer):
|
||||
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)
|
||||
if self.semantic_frame_rate == "25hz":
|
||||
|
||||
@ -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
|
||||
import torch
|
||||
import torchaudio
|
||||
from torch import nn
|
||||
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()
|
||||
from text import cleaned_text_to_sequence
|
||||
import soundfile
|
||||
@ -196,6 +197,11 @@ class VitsModel(nn.Module):
|
||||
super().__init__()
|
||||
dict_s2 = torch.load(vits_path,map_location="cpu")
|
||||
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.model.semantic_frame_rate = "25hz"
|
||||
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)
|
||||
|
||||
|
||||
def export(vits_path, gpt_path, project_name):
|
||||
def export(vits_path, gpt_path, project_name, vits_model="v2"):
|
||||
vits = VitsModel(vits_path)
|
||||
gpt = T2SModel(gpt_path, vits)
|
||||
gpt_sovits = GptSoVits(vits, gpt)
|
||||
ssl = SSLModel()
|
||||
ref_seq = torch.LongTensor([cleaned_text_to_sequence(["n", "i2", "h", "ao3", ",", "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"])])
|
||||
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"],version=vits_model)])
|
||||
ref_bert = torch.randn((ref_seq.shape[1], 1024)).float()
|
||||
text_bert = torch.randn((text_seq.shape[1], 1024)).float()
|
||||
ref_audio = torch.randn((1, 48000 * 5)).float()
|
||||
@ -287,34 +293,38 @@ def export(vits_path, gpt_path, project_name):
|
||||
pass
|
||||
|
||||
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:
|
||||
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("out2.wav", b[0], vits.hps.data.sampling_rate)
|
||||
return
|
||||
|
||||
a = gpt_sovits(ref_seq, text_seq, ref_bert, text_bert, ref_audio_sr, ssl_content).detach().cpu().numpy()
|
||||
else:
|
||||
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)
|
||||
|
||||
gpt_sovits.export(ref_seq, text_seq, ref_bert, text_bert, ref_audio_sr, ssl_content, project_name)
|
||||
if vits_model == "v1":
|
||||
symbols = symbols_v1
|
||||
else:
|
||||
symbols = symbols_v2
|
||||
|
||||
MoeVSConf = {
|
||||
"Folder" : f"{project_name}",
|
||||
"Name" : f"{project_name}",
|
||||
"Type" : "GPT-SoVits",
|
||||
"Rate" : vits.hps.data.sampling_rate,
|
||||
"NumLayers": gpt.t2s_model.num_layers,
|
||||
"EmbeddingDim": gpt.t2s_model.embedding_dim,
|
||||
"Dict": "BasicDict",
|
||||
"BertPath": "chinese-roberta-wwm-ext-large",
|
||||
"Symbol": symbols,
|
||||
"AddBlank": False
|
||||
}
|
||||
|
||||
"Folder": f"{project_name}",
|
||||
"Name": f"{project_name}",
|
||||
"Type": "GPT-SoVits",
|
||||
"Rate": vits.hps.data.sampling_rate,
|
||||
"NumLayers": gpt.t2s_model.num_layers,
|
||||
"EmbeddingDim": gpt.t2s_model.embedding_dim,
|
||||
"Dict": "BasicDict",
|
||||
"BertPath": "chinese-roberta-wwm-ext-large",
|
||||
# "Symbol": symbols,
|
||||
"AddBlank": False,
|
||||
}
|
||||
|
||||
MoeVSConfJson = json.dumps(MoeVSConf)
|
||||
with open(f"onnx/{project_name}.json", 'w') as MoeVsConfFile:
|
||||
json.dump(MoeVSConf, MoeVsConfFile, indent = 4)
|
||||
|
||||
@ -27,7 +27,7 @@ if is_g2pw:
|
||||
print("当前使用g2pw进行拼音推理")
|
||||
from text.g2pw import G2PWPinyin, correct_pronunciation
|
||||
parent_directory = os.path.dirname(current_file_path)
|
||||
g2pw = G2PWPinyin(model_dir="GPT_SoVITS/text/G2PWModel",model_source="GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large",v_to_u=False, neutral_tone_with_five=True)
|
||||
g2pw = G2PWPinyin(model_dir="GPT_SoVITS/text/G2PWModel",model_source=os.environ.get("bert_path","GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large"),v_to_u=False, neutral_tone_with_five=True)
|
||||
|
||||
rep_map = {
|
||||
"：": ",",
|
||||
|
||||
@ -45021,4 +45021,6 @@
|
||||
黄冠野服: ['huang2', 'guan4', 'ye3', 'fu2']
|
||||
黄发台背: ['huang2', 'fa1', 'tai2', 'bei4']
|
||||
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']
|
||||
Binary file not shown.
@ -186,6 +186,7 @@ def replace_positive_quantifier(match) -> str:
|
||||
match_2: str = match_2 if match_2 else ""
|
||||
quantifiers: str = match.group(3)
|
||||
number: str = num2str(number)
|
||||
number = "两" if number == "二" else number
|
||||
result = f"{number}{match_2}{quantifiers}"
|
||||
return result
|
||||
|
||||
|
||||
@ -184,8 +184,8 @@ D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
|
||||
|
||||
#### Integrated Package Users
|
||||
|
||||
Double-click `go-webui.bat`or use `go-webui.ps`
|
||||
if you want to switch to V1,then double-click`go-webui-v1.bat` or use `go-webui-v1.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.ps1`
|
||||
|
||||
#### Others
|
||||
|
||||
@ -220,7 +220,7 @@ Or maunally switch version in WebUI
|
||||
|
||||
#### 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
|
||||
|
||||
|
||||
@ -181,8 +181,8 @@ 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.ps`
|
||||
若想使用V1,则双击`go-webui-v1.bat`或者使用`go-webui-v1.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.ps1`
|
||||
|
||||
#### 其他
|
||||
|
||||
@ -217,7 +217,7 @@ python webui.py v1 <language(optional)>
|
||||
|
||||
#### 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
|
||||
|
||||
|
||||
@ -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
|
||||
|
||||
`go-webui.bat`をダブルクリックするか、`go-webui.ps`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。
|
||||
`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
|
||||
|
||||
@ -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
|
||||
|
||||
`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
|
||||
|
||||
|
||||
@ -175,8 +175,8 @@ D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
|
||||
|
||||
#### 통합 패키지 사용자
|
||||
|
||||
`go-webui.bat`을 더블 클릭하거나 `go-webui.ps`를 사용하십시오.
|
||||
V1으로 전환하려면, `go-webui-v1.bat`을 더블 클릭하거나 `go-webui-v1.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.ps1`를 사용하십시오.
|
||||
|
||||
#### 기타
|
||||
|
||||
@ -211,7 +211,7 @@ python webui.py v1 <언어(옵션)>
|
||||
|
||||
#### 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
|
||||
|
||||
|
||||
@ -172,8 +172,8 @@ D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
|
||||
|
||||
#### Entegre Paket Kullanıcıları
|
||||
|
||||
`go-webui.bat` dosyasına çift tıklayın veya `go-webui.ps` kullanın.
|
||||
V1'e geçmek istiyorsanız, `go-webui-v1.bat` dosyasına çift tıklayın veya `go-webui-v1.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.ps1` kullanın.
|
||||
|
||||
#### Diğerleri
|
||||
|
||||
@ -208,7 +208,7 @@ veya WebUI'de manuel olarak sürüm değiştirin.
|
||||
|
||||
#### 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
|
||||
|
||||
@ -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">
|
||||
<img src="https://contrib.rocks/image?repo=RVC-Boss/GPT-SoVITS" />
|
||||
</a>
|
||||
</a>
|
||||
|
||||
@ -34,3 +34,4 @@ opencc; sys_platform != 'linux'
|
||||
opencc==1.1.1; sys_platform == 'linux'
|
||||
python_mecab_ko; sys_platform != 'win32'
|
||||
gradio-i18n
|
||||
fastapi<0.112.2
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user