Sync with main branch (#1134)

* api接口，修复文本切分符号设定中，中文分号错写为英文分号的问题 (#1001)

* 一些小问题修复 (#1021)

* fix import error. It may happen when calling
 api.py

* Update README.md

* Update gpt-sovits_kaggle.ipynb

* Update gpt-sovits_kaggle.ipynb

* fix path error
delete useless line wraps

* 删除重复的 COPY 指令 (#1073)

* [优化] 1Aa-文本获取 (#1102)

* Filter unsupported languages

* add feedback

* simplify modification

* fix detail

* Update english.py (#1106)

copy but not ref the phones list becoz it will be extend later, if not do so,it will affect the self.cmu dict values.

* Update models.py

* modify freeze_quantizer mode, avoid quantizer's codebook updating (#953)

---------

Co-authored-by: FengQingYunDan <pingdengjia0liu@163.com>
Co-authored-by: Kenn Zhang <breakstring@hotmail.com>
Co-authored-by: 蓝梦实 <36986837+SapphireLab@users.noreply.github.com>
Co-authored-by: lyris <lyris@users.noreply.github.com>
Co-authored-by: hcwu1993 <15855138469@163.com>

.gitignore

@@ -10,8 +10,3 @@ reference
 GPT_weights
 SoVITS_weights
 TEMP
-PortableGit
-ffmpeg.exe
-ffprobe.exe
-tmp_audio
-trained

Dockerfile

@@ -34,9 +34,6 @@ RUN if [ "$IMAGE_TYPE" != "elite" ]; then \
     fi
 
 
-# Copy the rest of the application
-COPY . /workspace
-
 # Copy the rest of the application
 COPY . /workspace
 

GPT_SoVITS/module/models.py

@@ -16,6 +16,7 @@ from module.mrte_model import MRTE
 from module.quantize import ResidualVectorQuantizer
 from text import symbols
 from torch.cuda.amp import autocast
+import contextlib
 
 
 class StochasticDurationPredictor(nn.Module):
@@ -891,9 +892,10 @@ class SynthesizerTrn(nn.Module):
             self.ssl_proj = nn.Conv1d(ssl_dim, ssl_dim, 1, stride=1)
 
         self.quantizer = ResidualVectorQuantizer(dimension=ssl_dim, n_q=1, bins=1024)
-        if freeze_quantizer:
-            self.ssl_proj.requires_grad_(False)
-            self.quantizer.requires_grad_(False)
+        self.freeze_quantizer = freeze_quantizer
+        # if freeze_quantizer:
+        #     self.ssl_proj.requires_grad_(False)
+        #     self.quantizer.requires_grad_(False)
             #self.quantizer.eval()
             # self.enc_p.text_embedding.requires_grad_(False)
             # self.enc_p.encoder_text.requires_grad_(False)
@@ -906,6 +908,11 @@ class SynthesizerTrn(nn.Module):
         ge = self.ref_enc(y * y_mask, y_mask)
 
         with autocast(enabled=False):
+            maybe_no_grad = torch.no_grad() if self.freeze_quantizer else contextlib.nullcontext()
+            with maybe_no_grad:
+                if self.freeze_quantizer:
+                    self.ssl_proj.eval()
+                    self.quantizer.eval()
             ssl = self.ssl_proj(ssl)
             quantized, codes, commit_loss, quantized_list = self.quantizer(
                 ssl, layers=[0]

GPT_SoVITS/prepare_datasets/1-get-text.py

@@ -117,9 +117,12 @@ if os.path.exists(txt_path) == False:
         try:
             wav_name, spk_name, language, text = line.split("|")
             # todo.append([name,text,"zh"])
-            todo.append(
-                [wav_name, text, language_v1_to_language_v2.get(language, language)]
-            )
+            if language in language_v1_to_language_v2.keys():
+                todo.append(
+                    [wav_name, text, language_v1_to_language_v2.get(language, language)]
+                )
+            else:
+                print(f"\033[33m[Waring] The {language = } of {wav_name} is not supported for training.\033[0m")
         except:
             print(line, traceback.format_exc())
 

GPT_SoVITS/text/english.py

@@ -320,7 +320,7 @@ class en_G2p(G2p):
 
         # 尝试分离所有格
         if re.match(r"^([a-z]+)('s)$", word):
-            phones = self.qryword(word[:-2])
+            phones = self.qryword(word[:-2])[:]
             # P T K F TH HH 无声辅音结尾 's 发 ['S']
             if phones[-1] in ['P', 'T', 'K', 'F', 'TH', 'HH']:
                 phones.extend(['S'])

api.py

@@ -120,6 +120,11 @@ RESP: 无
 import argparse
 import os,re
 import sys
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+sys.path.append("%s/GPT_SoVITS" % (now_dir))
+
 import signal
 import LangSegment
 from time import time as ttime
@@ -381,7 +386,7 @@ def read_clean_buffer(audio_bytes):
 
 
 def cut_text(text, punc):
-    punc_list = [p for p in punc if p in {",", ".", ";", "?", "!", "、", "，", "。", "？", "！", ";", "：", "…"}]
+    punc_list = [p for p in punc if p in {",", ".", ";", "?", "!", "、", "，", "。", "？", "！", "；", "：", "…"}]
     if len(punc_list) > 0:
         punds = r"[" + "".join(punc_list) + r"]"
         text = text.strip("\n")
@@ -536,10 +541,6 @@ def handle(refer_wav_path, prompt_text, prompt_language, text, text_language, cu
 # --------------------------------
 # 初始化部分
 # --------------------------------
-now_dir = os.getcwd()
-sys.path.append(now_dir)
-sys.path.append("%s/GPT_SoVITS" % (now_dir))
-
 dict_language = {
     "中文": "all_zh",
     "英文": "en",
@@ -579,7 +580,7 @@ parser.add_argument("-hp", "--half_precision", action="store_true", default=Fals
 # 此时 full_precision==True, half_precision==False
 parser.add_argument("-sm", "--stream_mode", type=str, default="close", help="流式返回模式, close / normal / keepalive")
 parser.add_argument("-mt", "--media_type", type=str, default="wav", help="音频编码格式, wav / ogg / aac")
-parser.add_argument("-cp", "--cut_punc", type=str, default="", help="文本切分符号设定, 符号范围,.;?!、，。？！;：…")
+parser.add_argument("-cp", "--cut_punc", type=str, default="", help="文本切分符号设定, 符号范围,.;?!、，。？！；：…")
 # 切割常用分句符为 `python ./api.py -cp ".?!。？！"`
 parser.add_argument("-hb", "--hubert_path", type=str, default=g_config.cnhubert_path, help="覆盖config.cnhubert_path")
 parser.add_argument("-b", "--bert_path", type=str, default=g_config.bert_path, help="覆盖config.bert_path")

docs/ja/README.md

@@ -159,7 +159,7 @@ D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
 - [ ] **優先度 高:**
 
   - [x] 日本語と英語でのローカライズ。
-  - [] ユーザーガイド。
+  - [ ] ユーザーガイド。
   - [x] 日本語データセットと英語データセットのファインチューニングトレーニング。
 
 - [ ] **機能:**

gpt-sovits_kaggle.ipynb

@@ -54,11 +54,11 @@
    "source": [
     "# @title Download pretrained models 下载预训练模型\n",
     "!mkdir -p /kaggle/working/GPT-SoVITS/GPT_SoVITS/pretrained_models\n",
-    "!mkdir -p /kaggle/working/GPT-SoVITS/tools/damo_asr/models\n",
+    "!mkdir -p /kaggle/working/GPT-SoVITS/tools/asr/models\n",
     "!mkdir -p /kaggle/working/GPT-SoVITS/tools/uvr5\n",
     "%cd /kaggle/working/GPT-SoVITS/GPT_SoVITS/pretrained_models\n",
     "!git clone https://huggingface.co/lj1995/GPT-SoVITS\n",
-    "%cd /kaggle/working/GPT-SoVITS/tools/damo_asr/models\n",
+    "%cd /kaggle/working/GPT-SoVITS/tools/asr/models\n",
     "!git clone https://www.modelscope.cn/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch.git\n",
     "!git clone https://www.modelscope.cn/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch.git\n",
     "!git clone https://www.modelscope.cn/damo/punc_ct-transformer_zh-cn-common-vocab272727-pytorch.git\n",

tools/my_utils.py

@@ -28,4 +28,4 @@ def load_audio(file, sr):
 def clean_path(path_str):
     if platform.system() == 'Windows':
         path_str = path_str.replace('/', '\\')
-    return path_str.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+    return path_str.strip(" ").strip('"').strip("\n").strip('"').strip(" ").strip("\u202a")

webui.py

@@ -418,7 +418,10 @@ def open1a(inp_text,inp_wav_dir,exp_name,gpu_numbers,bert_pretrained_dir):
         with open(path_text, "w", encoding="utf8") as f:
             f.write("\n".join(opt) + "\n")
         ps1a=[]
-        yield "文本进程结束",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
+        if len("".join(opt)) > 0:
+            yield "文本进程成功", {"__type__": "update", "visible": True}, {"__type__": "update", "visible": False}
+        else:
+            yield "文本进程失败", {"__type__": "update", "visible": True}, {"__type__": "update", "visible": False}
     else:
         yield "已有正在进行的文本任务，需先终止才能开启下一次任务", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
 
@@ -583,7 +586,7 @@ def open1abc(inp_text,inp_wav_dir,exp_name,gpu_numbers1a,gpu_numbers1Ba,gpu_numb
                     os.remove(txt_path)
                 with open(path_text, "w",encoding="utf8") as f:
                     f.write("\n".join(opt) + "\n")
-
+                assert len("".join(opt)) > 0, "1Aa-文本获取进程失败"
             yield "进度：1a-done", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
             ps1abc=[]
             #############################1b