Merge branch 'main' into inference_optin

.gitignore

@@ -10,3 +10,5 @@ reference
 GPT_weights
 SoVITS_weights
 TEMP
+
+

Dockerfile

@@ -40,15 +40,6 @@ COPY . /workspace
 # Copy the rest of the application
 COPY . /workspace
 
-
+EXPOSE 9871 9872 9873 9874 9880
-EXPOSE 9870
-EXPOSE 9871
-EXPOSE 9872
-EXPOSE 9873
-EXPOSE 9874
-
-VOLUME /workspace/output
-VOLUME /workspace/logs
-VOLUME /workspace/SoVITS_weights
 
 CMD ["python", "webui.py"]

GPT_SoVITS/AR/data/data_module.py

@@ -41,7 +41,8 @@ class Text2SemanticDataModule(LightningDataModule):
         #     pad_val=self.config['data']['pad_val'])
 
     def train_dataloader(self):
-        batch_size = max(min(self.config["train"]["batch_size"],len(self._train_dataset)//4),1)#防止不保存
+        batch_size=self.config["train"]["batch_size"]//2 if self.config["train"].get("if_dpo",False)==True else self.config["train"]["batch_size"]
+        batch_size = max(min(batch_size,len(self._train_dataset)//4),1)#防止不保存
         sampler = DistributedBucketSampler(self._train_dataset, batch_size=batch_size)
         return DataLoader(
             self._train_dataset,

GPT_SoVITS/AR/models/t2s_lightning_module.py

@@ -11,7 +11,6 @@ from AR.models.t2s_model import Text2SemanticDecoder
 from AR.modules.lr_schedulers import WarmupCosineLRSchedule
 from AR.modules.optim import ScaledAdam
 
-
 class Text2SemanticLightningModule(LightningModule):
     def __init__(self, config, output_dir, is_train=True):
         super().__init__()
@@ -35,7 +34,8 @@ class Text2SemanticLightningModule(LightningModule):
     def training_step(self, batch: Dict, batch_idx: int):
         opt = self.optimizers()
         scheduler = self.lr_schedulers()
-        loss, acc = self.model.forward(
+        forward=self.model.forward if self.config["train"].get("if_dpo",False)==True else self.model.forward_old
+        loss, acc = forward(
             batch["phoneme_ids"],
             batch["phoneme_ids_len"],
             batch["semantic_ids"],

GPT_SoVITS/AR/models/t2s_model.py

@@ -8,6 +8,9 @@ from AR.models.utils import (
     sample,
     logits_to_probs,
     multinomial_sample_one_no_sync,
+    dpo_loss,
+    make_reject_y, 
+    get_batch_logps
 )
 from AR.modules.embedding import SinePositionalEmbedding
 from AR.modules.embedding import TokenEmbedding
@@ -85,11 +88,104 @@ class Text2SemanticDecoder(nn.Module):
             ignore_index=self.EOS,
         )
 
+    def make_input_data(self, x, x_lens, y, y_lens, bert_feature):
+        x = self.ar_text_embedding(x)
+        x = x + self.bert_proj(bert_feature.transpose(1, 2))
+        x = self.ar_text_position(x)
+        x_mask = make_pad_mask(x_lens)
+
+        y_mask = make_pad_mask(y_lens)
+        y_mask_int = y_mask.type(torch.int64)
+        codes = y.type(torch.int64) * (1 - y_mask_int)
+
+        # Training
+        # AR Decoder
+        y, targets = self.pad_y_eos(codes, y_mask_int, eos_id=self.EOS)
+        x_len = x_lens.max()
+        y_len = y_lens.max()
+        y_emb = self.ar_audio_embedding(y)
+        y_pos = self.ar_audio_position(y_emb)
+
+        xy_padding_mask = torch.concat([x_mask, y_mask], dim=1)
+
+        ar_xy_padding_mask = xy_padding_mask
+
+        x_attn_mask = F.pad(
+            torch.zeros((x_len, x_len), dtype=torch.bool, device=x.device),
+            (0, y_len),
+            value=True,
+        )
+        
+        y_attn_mask = F.pad(
+            torch.triu(
+                torch.ones(y_len, y_len, dtype=torch.bool, device=x.device),
+                diagonal=1,
+            ),
+            (x_len, 0),
+            value=False,
+        )
+
+        xy_attn_mask = torch.concat([x_attn_mask, y_attn_mask], dim=0)
+        bsz, src_len = x.shape[0], x_len + y_len
+        _xy_padding_mask = (
+            ar_xy_padding_mask.view(bsz, 1, 1, src_len)
+            .expand(-1, self.num_head, -1, -1)
+            .reshape(bsz * self.num_head, 1, src_len)
+        )
+        xy_attn_mask = xy_attn_mask.logical_or(_xy_padding_mask)
+        new_attn_mask = torch.zeros_like(xy_attn_mask, dtype=x.dtype)
+        new_attn_mask.masked_fill_(xy_attn_mask, float("-inf"))
+        xy_attn_mask = new_attn_mask
+        # x 和完整的 y 一次性输入模型
+        xy_pos = torch.concat([x, y_pos], dim=1)
+
+        return xy_pos, xy_attn_mask, targets
+
     def forward(self, x, x_lens, y, y_lens, bert_feature):
         """
         x: phoneme_ids
         y: semantic_ids
         """
+
+        reject_y, reject_y_lens = make_reject_y(y, y_lens)
+
+        xy_pos, xy_attn_mask, targets = self.make_input_data(x, x_lens, y, y_lens, bert_feature)
+
+        xy_dec, _ = self.h(
+            (xy_pos, None),
+            mask=xy_attn_mask,
+        )
+        x_len = x_lens.max()
+        logits = self.ar_predict_layer(xy_dec[:, x_len:])
+
+        ###### DPO #############
+        reject_xy_pos, reject_xy_attn_mask, reject_targets = self.make_input_data(x, x_lens, reject_y, reject_y_lens, bert_feature)
+
+        reject_xy_dec, _ = self.h(
+            (reject_xy_pos, None),
+            mask=reject_xy_attn_mask,
+        )
+        x_len = x_lens.max()
+        reject_logits = self.ar_predict_layer(reject_xy_dec[:, x_len:])
+
+        # loss
+        # from feiteng: 每次 duration 越多, 梯度更新也应该更多, 所以用 sum
+
+        loss_1 = F.cross_entropy(logits.permute(0, 2, 1), targets, reduction="sum")
+        acc = self.ar_accuracy_metric(logits.permute(0, 2, 1).detach(), targets).item()
+
+        A_logits, R_logits = get_batch_logps(logits, reject_logits, targets, reject_targets)
+        loss_2, _, _ = dpo_loss(A_logits, R_logits, 0, 0, 0.2, reference_free=True)
+        
+        loss = loss_1 + loss_2
+
+        return loss, acc
+
+    def forward_old(self, x, x_lens, y, y_lens, bert_feature):
+        """
+        x: phoneme_ids
+        y: semantic_ids
+        """
         x = self.ar_text_embedding(x)
         x = x + self.bert_proj(bert_feature.transpose(1, 2))
         x = self.ar_text_position(x)
@@ -231,6 +327,7 @@ class Text2SemanticDecoder(nn.Module):
         prompts,  ####参考音频token
         bert_feature,
         top_k: int = -100,
+        top_p: int = 100,
         early_stop_num: int = -1,
         temperature: float = 1.0,
     ):
@@ -299,7 +396,7 @@ class Text2SemanticDecoder(nn.Module):
             if(idx==0):###第一次跑不能EOS否则没有了
                 logits = logits[:, :-1]  ###刨除1024终止符号的概率
             samples = sample(
-                logits[0], y, top_k=top_k, top_p=1.0, repetition_penalty=1.35
+                logits[0], y, top_k=top_k, top_p=top_p, repetition_penalty=1.35, temperature=temperature
             )[0].unsqueeze(0)
             # 本次生成的 semantic_ids 和之前的 y 构成新的 y
             # print(samples.shape)#[1,1]#第一个1是bs

GPT_SoVITS/AR/models/utils.py

@@ -1,7 +1,7 @@
 # modified from https://github.com/feng-yufei/shared_debugging_code/blob/main/model/utils.py\
 import torch
 import torch.nn.functional as F
-
+from typing import Tuple
 
 def sequence_mask(length, max_length=None):
     if max_length is None:
@@ -159,3 +159,70 @@ def sample(
     )
     idx_next = multinomial_sample_one_no_sync(probs)
     return idx_next, probs
+
+def dpo_loss(policy_chosen_logps: torch.FloatTensor,
+             policy_rejected_logps: torch.FloatTensor,
+             reference_chosen_logps: torch.FloatTensor,
+             reference_rejected_logps: torch.FloatTensor,
+             beta: float,
+             reference_free: bool = False) -> Tuple[torch.FloatTensor, torch.FloatTensor, torch.FloatTensor]:
+    pi_logratios = policy_chosen_logps - policy_rejected_logps
+    ref_logratios = reference_chosen_logps - reference_rejected_logps
+
+    if reference_free:
+        ref_logratios = 0
+
+    logits = pi_logratios - ref_logratios
+
+    losses = -F.logsigmoid(beta * logits)
+    chosen_rewards = beta * (policy_chosen_logps - reference_chosen_logps).detach()
+    rejected_rewards = beta * (policy_rejected_logps - reference_rejected_logps).detach()
+
+    return losses.mean(), chosen_rewards, rejected_rewards
+
+def get_batch_logps(logits_target: torch.FloatTensor, logits_reject: torch.FloatTensor, labels_target: torch.LongTensor, labels_reject: torch.LongTensor, average_log_prob: bool = False) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
+
+    # dummy token; we'll ignore the losses on these tokens later
+
+    per_token_logps_target = torch.gather(logits_target.log_softmax(-1), dim=2, index=labels_target.unsqueeze(2)).squeeze(2)
+    per_token_logps_reject = torch.gather(logits_reject.log_softmax(-1), dim=2, index=labels_reject.unsqueeze(2)).squeeze(2)
+
+    return per_token_logps_target.sum(-1), per_token_logps_reject.sum(-1)
+
+def make_reject_y(y_o, y_lens):
+    def repeat_P(y):
+        range_idx, _ = torch.randint(0, len(y), size=(2,)).sort()
+        pre = y[:range_idx[0]]
+        shf = y[range_idx[1]:]
+        range_text = y[range_idx[0]:range_idx[1]]
+        new_y = torch.cat([pre, range_text, range_text, shf])
+        return new_y
+    def lost_P(y):
+        range_idx, _ = torch.randint(0, len(y), size=(2,)).sort()
+        pre = y[:range_idx[0]]
+        shf = y[range_idx[1]:]
+        range_text = y[range_idx[0]:range_idx[1]]
+        new_y = torch.cat([pre, shf])
+        return new_y
+    bs = len(y_lens)
+    reject_y = []
+    reject_y_lens = []
+    for b in range(bs):
+        process_item_idx = torch.randint(0, 1, size=(1, ))[0]
+        if process_item_idx == 0:
+            new_y = repeat_P(y_o[b])
+            reject_y.append(new_y)
+            reject_y_lens.append(len(new_y))
+        elif process_item_idx==1:
+            new_y = lost_P(y_o[b])
+            reject_y.append(new_y)
+            reject_y_lens.append(len(new_y))
+    max_length = max(reject_y_lens)
+    for b in range(bs):
+        pad_length = max_length - reject_y_lens[b]
+        reject_y[b] = torch.cat([reject_y[b], torch.zeros(pad_length, dtype=y_o.dtype, device=y_o.device)], dim=0)
+
+    reject_y = torch.stack(reject_y, dim = 0)
+    reject_y_lens = torch.tensor(reject_y_lens, device=y_lens.device)
+
+    return reject_y, reject_y_lens

GPT_SoVITS/inference_webui.py

@@ -245,7 +245,14 @@ def splite_en_inf(sentence, language):
 
 
 def clean_text_inf(text, language):
-    phones, word2ph, norm_text = clean_text(text, language.replace("all_",""))
+    formattext = ""
+    language = language.replace("all_","")
+    for tmp in LangSegment.getTexts(text):
+        if tmp["lang"] == language:
+            formattext += tmp["text"] + " "
+    while "  " in formattext:
+        formattext = formattext.replace("  ", " ")
+    phones, word2ph, norm_text = clean_text(formattext, language)
     phones = cleaned_text_to_sequence(phones)
     return phones, word2ph, norm_text
 
@@ -305,9 +312,8 @@ def nonen_get_bert_inf(text, language):
     print(langlist)
     bert_list = []
     for i in range(len(textlist)):
-        text = textlist[i]
         lang = langlist[i]
-        phones, word2ph, norm_text = clean_text_inf(text, lang)
+        phones, word2ph, norm_text = clean_text_inf(textlist[i], lang)
         bert = get_bert_inf(phones, word2ph, norm_text, lang)
         bert_list.append(bert)
     bert = torch.cat(bert_list, dim=1)
@@ -359,7 +365,7 @@ def merge_short_text_in_array(texts, threshold):
             result[len(result) - 1] += text
     return result
 
-def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, how_to_cut=i18n("不切")):
+def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, how_to_cut=i18n("不切"), top_k=20, top_p=0.6, temperature=0.6):
     t0 = ttime()
     prompt_language = dict_language[prompt_language]
     text_language = dict_language[text_language]
@@ -439,7 +445,9 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language,
                 None,
                 bert,
                 # prompt_phone_len=ph_offset,
-                top_k=config["inference"]["top_k"],
+                top_k=top_k,
+                top_p=top_p,
+                temperature=temperature,
                 early_stop_num=hz * max_sec,
             )
         t3 = ttime()
@@ -616,6 +624,10 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
                 value=i18n("凑四句一切"),
                 interactive=True,
             )
+            with gr.Row():
+                top_k = gr.Slider(minimum=1,maximum=100,step=1,label=i18n("top_k"),value=5,interactive=True)
+                top_p = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("top_p"),value=1,interactive=True)
+                temperature = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("temperature"),value=1,interactive=True)
             inference_button = gr.Button(i18n("合成语音"), variant="primary")
             output = gr.Audio(label=i18n("输出的语音"))
 

GPT_SoVITS/prepare_datasets/1-get-text.py

@@ -33,13 +33,13 @@ from time import time as ttime
 import shutil
 
 
-def my_save(fea, path):  #####fix issue: torch.save doesn't support chinese path
+def my_save(fea,path):#####fix issue: torch.save doesn't support chinese path
-    dir = os.path.dirname(path)
+    dir=os.path.dirname(path)
-    name = os.path.basename(path)
+    name=os.path.basename(path)
-    tmp_path = "%s/%s%s.pth" % (dir, ttime(), i_part)
+    # tmp_path="%s/%s%s.pth"%(dir,ttime(),i_part)
-    torch.save(fea, tmp_path)
+    tmp_path="%s%s.pth"%(ttime(),i_part)
-    shutil.move(tmp_path, "%s/%s" % (dir, name))
+    torch.save(fea,tmp_path)
-
+    shutil.move(tmp_path,"%s/%s"%(dir,name))
 
 
 txt_path = "%s/2-name2text-%s.txt" % (opt_dir, i_part)

GPT_SoVITS/prepare_datasets/2-get-hubert-wav32k.py

@@ -35,7 +35,8 @@ import shutil
 def my_save(fea,path):#####fix issue: torch.save doesn't support chinese path
     dir=os.path.dirname(path)
     name=os.path.basename(path)
-    tmp_path="%s/%s%s.pth"%(dir,ttime(),i_part)
+    # tmp_path="%s/%s%s.pth"%(dir,ttime(),i_part)
+    tmp_path="%s%s.pth"%(ttime(),i_part)
     torch.save(fea,tmp_path)
     shutil.move(tmp_path,"%s/%s"%(dir,name))
 

GPT_SoVITS/text/tone_sandhi.py

@@ -672,6 +672,7 @@ class ToneSandhi:
                 and i + 1 < len(seg)
                 and seg[i - 1][0] == seg[i + 1][0]
                 and seg[i - 1][1] == "v"
+                and seg[i + 1][1] == "v"
             ):
                 new_seg[i - 1][0] = new_seg[i - 1][0] + "一" + new_seg[i - 1][0]
             else:

README.md

@@ -154,7 +154,7 @@ docker compose -f "docker-compose.yaml" up -d
 As above, modify the corresponding parameters based on your actual situation, then run the following command:
 
 ```
-docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-DockerTest\output:/workspace/output --volume=G:\GPT-SoVITS-DockerTest\logs:/workspace/logs --volume=G:\GPT-SoVITS-DockerTest\SoVITS_weights:/workspace/SoVITS_weights --workdir=/workspace -p 9870:9870 -p 9871:9871 -p 9872:9872 -p 9873:9873 -p 9874:9874 --shm-size="16G" -d breakstring/gpt-sovits:xxxxx
+docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-DockerTest\output:/workspace/output --volume=G:\GPT-SoVITS-DockerTest\logs:/workspace/logs --volume=G:\GPT-SoVITS-DockerTest\SoVITS_weights:/workspace/SoVITS_weights --workdir=/workspace -p 9880:9880 -p 9871:9871 -p 9872:9872 -p 9873:9873 -p 9874:9874 --shm-size="16G" -d breakstring/gpt-sovits:xxxxx
 ```
 
 ## Dataset Format

config.py

@@ -5,8 +5,10 @@ import torch
 # 推理用的指定模型
 sovits_path = ""
 gpt_path = ""
-is_half = eval(os.environ.get("is_half","True"))
+is_half_str = os.environ.get("is_half", "True")
-is_share=False
+is_half = True if is_half_str.lower() == 'true' else False
+is_share_str = os.environ.get("is_share","False")
+is_share= True if is_share_str.lower() == 'true' else False
 
 cnhubert_path = "GPT_SoVITS/pretrained_models/chinese-hubert-base"
 bert_path = "GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large"

docker-compose.yaml

@@ -2,10 +2,11 @@ version: '3.8'
 
 services:
   gpt-sovits:
-    image: breakstring/gpt-sovits:xxxxx   # please change the image name and tag base your environment
+    image: breakstring/gpt-sovits:latest   # please change the image name and tag base your environment. If the tag contains the word 'elite', such as "latest-elite", it indicates that the image does not include the necessary models such as GPT-SoVITS, UVR5, Damo ASR, etc. You will need to download them yourself and map them into the container.
     container_name: gpt-sovits-container
     environment:
       - is_half=False
+      - is_share=False
     volumes:
       - ./output:/workspace/output
       - ./logs:/workspace/logs
@@ -13,7 +14,7 @@ services:
       - ./reference:/workspace/reference
     working_dir: /workspace
     ports:
-      - "9870:9870"
+      - "9880:9880"
       - "9871:9871"
       - "9872:9872"
       - "9873:9873"

docs/cn/Changelog_CN.md

@@ -107,8 +107,21 @@
 
 1-GPT训练卡死（win10 1909）和https://github.com/RVC-Boss/GPT-SoVITS/issues/232 （系统语言繁体）GPT训练报错，[尝试修复](https://github.com/RVC-Boss/GPT-SoVITS/commit/59f35adad85815df27e9c6b33d420f5ebfd8376b)。
 
+### 20240212更新
+
+1-faster whisper和funasr逻辑优化。faster whisper转镜像站下载，规避huggingface连不上的问题。
+
+2-DPO Loss实验性训练选项开启，通过构造负样本训练缓解GPT重复漏字问题。推理界面公开几个推理参数。 https://github.com/RVC-Boss/GPT-SoVITS/pull/457
+
+### 20240214更新
+
+1-训练支持中文实验名（原来会报错）
+
+2-DPO训练改为可勾选选项而非必须。如勾选batch size自动减半。修复推理界面新参数不传参的问题。
+
 todolist：
 
 1-中文多音字推理优化
 
 
+

docs/cn/README.md

@@ -138,7 +138,7 @@ docker compose -f "docker-compose.yaml" up -d
 同上，根据您自己的实际情况修改对应的参数，然后运行如下命令：
 
 ```
-docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-DockerTest\output:/workspace/output --volume=G:\GPT-SoVITS-DockerTest\logs:/workspace/logs --volume=G:\GPT-SoVITS-DockerTest\SoVITS_weights:/workspace/SoVITS_weights --workdir=/workspace -p 9870:9870 -p 9871:9871 -p 9872:9872 -p 9873:9873 -p 9874:9874 --shm-size="16G" -d breakstring/gpt-sovits:xxxxx
+docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-DockerTest\output:/workspace/output --volume=G:\GPT-SoVITS-DockerTest\logs:/workspace/logs --volume=G:\GPT-SoVITS-DockerTest\SoVITS_weights:/workspace/SoVITS_weights --workdir=/workspace -p 9880:9880 -p 9871:9871 -p 9872:9872 -p 9873:9873 -p 9874:9874 --shm-size="16G" -d breakstring/gpt-sovits:xxxxx
 ```
 
 ### 预训练模型

docs/ja/README.md

@@ -138,7 +138,7 @@ docker compose -f "docker-compose.yaml" up -d
 上記と同様に、実際の状況に基づいて対応するパラメータを変更し、次のコマンドを実行します：
 
 ```markdown
-docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-DockerTest\output:/workspace/output --volume=G:\GPT-SoVITS-DockerTest\logs:/workspace/logs --volume=G:\GPT-SoVITS-DockerTest\SoVITS_weights:/workspace/SoVITS_weights --workdir=/workspace -p 9870:9870 -p 9871:9871 -p 9872:9872 -p 9873:9873 -p 9874:9874 --shm-size="16G" -d breakstring/gpt-sovits:xxxxx
+docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-DockerTest\output:/workspace/output --volume=G:\GPT-SoVITS-DockerTest\logs:/workspace/logs --volume=G:\GPT-SoVITS-DockerTest\SoVITS_weights:/workspace/SoVITS_weights --workdir=/workspace -p 9880:9880 -p 9871:9871 -p 9872:9872 -p 9873:9873 -p 9874:9874 --shm-size="16G" -d breakstring/gpt-sovits:xxxxx
 ```
 
 ### 事前訓練済みモデル

docs/ko/README.md

@@ -141,7 +141,7 @@ docker compose -f "docker-compose.yaml" up -d
 위와 동일하게 실제 상황에 맞게 매개변수를 수정한 다음 다음 명령을 실행합니다:
 
 ```
-docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-DockerTest\output:/workspace/output --volume=G:\GPT-SoVITS-DockerTest\logs:/workspace/logs --volume=G:\GPT-SoVITS-DockerTest\SoVITS_weights:/workspace/SoVITS_weights --workdir=/workspace -p 9870:9870 -p 9871:9871 -p 9872:9872 -p 9873:9873 -p 9874:9874 --shm-size="16G" -d breakstring/gpt-sovits:xxxxx
+docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-DockerTest\output:/workspace/output --volume=G:\GPT-SoVITS-DockerTest\logs:/workspace/logs --volume=G:\GPT-SoVITS-DockerTest\SoVITS_weights:/workspace/SoVITS_weights --workdir=/workspace -p 9880:9880 -p 9871:9871 -p 9872:9872 -p 9873:9873 -p 9874:9874 --shm-size="16G" -d breakstring/gpt-sovits:xxxxx
 ```
 
 ### 사전 훈련된 모델

i18n/locale/fr_FR.json

@@ -8,8 +8,16 @@
     "是否开启UVR5-WebUI": "Activer UVR5-WebUI",
     "UVR5进程输出信息": "Informations de processus UVR5",
     "0b-语音切分工具": "0b-Outil de découpage vocal",
+    ".list标注文件的路径": "Chemin du fichier d'annotation .list",
+    "GPT模型列表": "Liste des modèles GPT",
+    "SoVITS模型列表": "Liste des modèles SoVITS",
+    "填切割后音频所在目录！读取的音频文件完整路径=该目录-拼接-list文件里波形对应的文件名（不是全路径）。": "Répertoire où sont enregistrés les fichiers audio après la découpe ! Chemin complet du fichier audio à lire = ce répertoire - nom du fichier correspondant à la forme d'onde dans le fichier liste (pas le chemin complet).",
     "音频自动切分输入路径，可文件可文件夹": "Chemin d'entrée automatique de découpage audio, peut être un fichier ou un dossier",
     "切分后的子音频的输出根目录": "Répertoire racine de sortie des sous-audios après découpage",
+    "怎么切": "Comment découper",
+    "不切": "Pas de découpe",
+    "凑四句一切": "Composez quatre phrases pour tout remplir",
+    "按英文句号.切": "Découpez par des points en anglais",
     "threshold:音量小于这个值视作静音的备选切割点": "seuil: le volume inférieur à cette valeur est considéré comme un point de coupe silencieux alternatif",
     "min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值": "min_length: longueur minimale de chaque segment, si le premier segment est trop court, il est continué avec le segment suivant jusqu'à dépasser cette valeur",
     "min_interval:最短切割间隔": "min_interval: intervalle de coupe minimum",

tools/asr/config.py

@@ -13,7 +13,7 @@ def check_fw_local_models():
         "large-v2", "large-v3"]
     for i, size in enumerate(model_size_list):
         if os.path.exists(f'tools/asr/models/faster-whisper-{size}'):
-            model_size_list[i] = size + '(local)'
+            model_size_list[i] = size + '-local'
     return model_size_list
 
 asr_dict = {

tools/asr/fasterwhisper_asr.py

@@ -1,5 +1,6 @@
 import argparse
 import os
+os.environ["HF_ENDPOINT"]="https://hf-mirror.com"
 import traceback
 import requests
 from glob import glob
@@ -8,6 +9,7 @@ from faster_whisper import WhisperModel
 from tqdm import tqdm
 
 from tools.asr.config import check_fw_local_models
+from tools.asr.funasr_asr import only_asr
 
 os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE"
 
@@ -35,8 +37,8 @@ language_code_list = [
     "auto"]
 
 def execute_asr(input_folder, output_folder, model_size, language,precision):
-    if 'local' in model_size:
+    if '-local' in model_size:
-        model_size = model_size.split('(')[0]
+        model_size = model_size[:-6]
         model_path = f'tools/asr/models/faster-whisper-{model_size}'
     else:
         model_path = model_size
@@ -63,8 +65,14 @@ def execute_asr(input_folder, output_folder, model_size, language,precision):
                 vad_parameters = dict(min_silence_duration_ms=700),
                 language       = language)
             text = ''
-            for segment in segments:
+
-                text += segment.text
+            if info.language == "zh":
+                print("检测为中文文本,转funasr处理")
+                text = only_asr(file)
+
+            if text == '':
+                for segment in segments:
+                    text += segment.text
             output.append(f"{file}|{output_file_name}|{info.language.upper()}|{text}")
         except:
             return print(traceback.format_exc())
@@ -96,4 +104,4 @@ if __name__ == '__main__':
         model_size    = cmd.model_size,
         language      = cmd.language,
         precision     = cmd.precision,
-    )
+    )

tools/asr/funasr_asr.py

@@ -23,6 +23,14 @@ model = AutoModel(
     punc_model_revision = "v2.0.4",
 )
 
+def only_asr(input_file):
+    try:
+        text = model.generate(input=input_file)[0]["text"]
+    except:
+        text = ''
+        print(traceback.format_exc())
+    return text
+
 def execute_asr(input_folder, output_folder, model_size, language):
     input_file_names = os.listdir(input_folder)
     input_file_names.sort()
@@ -35,7 +43,7 @@ def execute_asr(input_folder, output_folder, model_size, language):
             text = model.generate(input="%s/%s"%(input_folder, name))[0]["text"]
             output.append(f"{input_folder}/{name}|{output_file_name}|{language.upper()}|{text}")
         except:
-            return print(traceback.format_exc())
+            print(traceback.format_exc())
 
     output_folder = output_folder or "output/asr_opt"
     os.makedirs(output_folder, exist_ok=True)
@@ -65,4 +73,4 @@ if __name__ == '__main__':
         output_folder = cmd.output_folder,
         model_size    = cmd.model_size,
         language      = cmd.language,
-    )
+    )

webui.py

@@ -266,7 +266,7 @@ def close1Ba():
     return "已终止SoVITS训练",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
 
 p_train_GPT=None
-def open1Bb(batch_size,total_epoch,exp_name,if_save_latest,if_save_every_weights,save_every_epoch,gpu_numbers,pretrained_s1):
+def open1Bb(batch_size,total_epoch,exp_name,if_dpo,if_save_latest,if_save_every_weights,save_every_epoch,gpu_numbers,pretrained_s1):
     global p_train_GPT
     if(p_train_GPT==None):
         with open("GPT_SoVITS/configs/s1longer.yaml")as f:
@@ -283,6 +283,7 @@ def open1Bb(batch_size,total_epoch,exp_name,if_save_latest,if_save_every_weights
         data["train"]["save_every_n_epoch"]=save_every_epoch
         data["train"]["if_save_every_weights"]=if_save_every_weights
         data["train"]["if_save_latest"]=if_save_latest
+        data["train"]["if_dpo"]=if_dpo
         data["train"]["half_weights_save_dir"]=GPT_weight_root
         data["train"]["exp_name"]=exp_name
         data["train_semantic_path"]="%s/6-name2semantic.tsv"%s1_dir
@@ -807,6 +808,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
                 with gr.Row():
                     batch_size1Bb = gr.Slider(minimum=1,maximum=40,step=1,label=i18n("每张显卡的batch_size"),value=default_batch_size,interactive=True)
                     total_epoch1Bb = gr.Slider(minimum=2,maximum=50,step=1,label=i18n("总训练轮数total_epoch"),value=15,interactive=True)
+                    if_dpo = gr.Checkbox(label=i18n("是否开启dpo训练选项(实验性)"), value=False, interactive=True, show_label=True)
                     if_save_latest1Bb = gr.Checkbox(label=i18n("是否仅保存最新的ckpt文件以节省硬盘空间"), value=True, interactive=True, show_label=True)
                     if_save_every_weights1Bb = gr.Checkbox(label=i18n("是否在每次保存时间点将最终小模型保存至weights文件夹"), value=True, interactive=True, show_label=True)
                     save_every_epoch1Bb = gr.Slider(minimum=1,maximum=50,step=1,label=i18n("保存频率save_every_epoch"),value=5,interactive=True)
@@ -817,7 +819,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
                     info1Bb=gr.Textbox(label=i18n("GPT训练进程输出信息"))
             button1Ba_open.click(open1Ba, [batch_size,total_epoch,exp_name,text_low_lr_rate,if_save_latest,if_save_every_weights,save_every_epoch,gpu_numbers1Ba,pretrained_s2G,pretrained_s2D], [info1Ba,button1Ba_open,button1Ba_close])
             button1Ba_close.click(close1Ba, [], [info1Ba,button1Ba_open,button1Ba_close])
-            button1Bb_open.click(open1Bb, [batch_size1Bb,total_epoch1Bb,exp_name,if_save_latest1Bb,if_save_every_weights1Bb,save_every_epoch1Bb,gpu_numbers1Bb,pretrained_s1],   [info1Bb,button1Bb_open,button1Bb_close])
+            button1Bb_open.click(open1Bb, [batch_size1Bb,total_epoch1Bb,exp_name,if_dpo,if_save_latest1Bb,if_save_every_weights1Bb,save_every_epoch1Bb,gpu_numbers1Bb,pretrained_s1],   [info1Bb,button1Bb_open,button1Bb_close])
             button1Bb_close.click(close1Bb, [], [info1Bb,button1Bb_open,button1Bb_close])
             with gr.TabItem(i18n("1C-推理")):
                 gr.Markdown(value=i18n("选择训练完存放在SoVITS_weights和GPT_weights下的模型。默认的一个是底模，体验5秒Zero Shot TTS用。"))