Merge 50a88a596dea718c83e535136e9cb46b513cef6f into bf81cdb14a38b674b6e9996dabc97340bc9978d2

- Upgrade funasr from ==1.0.27 to >=1.3.7
- Add Fun-ASR-Nano (31 languages, Chinese dialects, recommended default)
- Add SenseVoice (ultra-fast 170x realtime, 5 languages)
- Keep original Paraformer as '达摩 ASR (中文经典)' for backward compat
- WebUI shows 3 FunASR options + Faster Whisper

Tested: routing logic verified for all backends (zh/en/ja/ko).
Resolves #2777

Co-authored-by: xiaoyunchong.xyc <xiaoyunchong.xyc@alibaba-inc.com>

GPT_SoVITS/text/g2pw/g2pw.py

@@ -1,5 +1,6 @@
 # This code is modified from https://github.com/mozillazg/pypinyin-g2pW
 
+import hashlib
 import pickle
 import os
 
@@ -14,6 +15,16 @@ current_file_path = os.path.dirname(__file__)
 CACHE_PATH = os.path.join(current_file_path, "polyphonic.pickle")
 PP_DICT_PATH = os.path.join(current_file_path, "polyphonic.rep")
 PP_FIX_DICT_PATH = os.path.join(current_file_path, "polyphonic-fix.rep")
+MD5_PATH = os.path.join(current_file_path, "polyphonic.md5")
+
+def get_file_md5(file_path):
+    if not os.path.exists(file_path):
+        return ""
+    hasher = hashlib.md5()
+    with open(file_path, "rb") as f:
+        for chunk in iter(lambda: f.read(4096), b""):
+            hasher.update(chunk)
+    return hasher.hexdigest()
 
 
 class G2PWPinyin(Pinyin):
@@ -115,13 +126,22 @@ def cache_dict(polyphonic_dict, file_path):
 
 
 def get_dict():
-    if os.path.exists(CACHE_PATH):
+    new_md5 = get_file_md5(PP_DICT_PATH) + get_file_md5(PP_FIX_DICT_PATH)
+    old_md5 = ""
+    if os.path.exists(MD5_PATH):
+        with open(MD5_PATH, "r", encoding="utf-8") as f:
+            old_md5 = f.read().strip()
+    need_rebuild = (not os.path.exists(CACHE_PATH)) or (new_md5 != old_md5)
+
+    if not need_rebuild:
         with open(CACHE_PATH, "rb") as pickle_file:
             polyphonic_dict = pickle.load(pickle_file)
     else:
+        print("Rebuilding Polyphonic Dictionary: " + f"{old_md5} -> {new_md5}")
         polyphonic_dict = read_dict()
         cache_dict(polyphonic_dict, CACHE_PATH)
-
+        with open(MD5_PATH, "w", encoding="utf-8") as f:
+            f.write(new_md5)
     return polyphonic_dict
 
 

GPT_SoVITS/text/g2pw/polyphonic-fix.rep

@@ -45023,4 +45023,5 @@
 鼎铛玉石: ['ding3', 'cheng1', 'yu4', 'shi2']
 齿豁头童: ['chi3', 'huo1', 'tou2', 'tong2']
 牦牛: ['mao2', 'niu2']
-牦: ['mao2']
+牦: ['mao2']
+唑: ['zuo4']

GPT_SoVITS/text/g2pw/polyphonic.md5

@@ -0,0 +1 @@
+13b2211c317c75794123ffdf7c2aea021c75b0c606ad61d7c8b05bb00b64fa21

requirements.txt

@@ -10,7 +10,7 @@ ffmpeg-python
 onnxruntime; platform_machine == "aarch64" or platform_machine == "arm64"
 onnxruntime-gpu; platform_machine == "x86_64" or platform_machine == "AMD64"
 tqdm
-funasr==1.0.27
+funasr>=1.3.7
 cn2an
 pypinyin
 pyopenjtalk>=0.4.1

tools/asr/config.py

@@ -5,15 +5,14 @@ def get_models():
         "large-v2",
         "large-v3",
         "large-v3-turbo",
-        #"distil-large-v2",
-        #"distil-large-v3",
-        #"distil-large-v3.5",
     ]
     return model_size_list
 
 
 asr_dict = {
-    "达摩 ASR (中文)": {"lang": ["zh", "yue"], "size": ["large"], "path": "funasr_asr.py", "precision": ["float32"]},
+    "Fun-ASR-Nano (31语种+方言, 推荐)": {"lang": ["zh", "en", "ja", "ko", "yue", "auto"], "size": ["large"], "path": "funasr_asr.py", "precision": ["float32"]},
+    "SenseVoice (极速, 5语种)": {"lang": ["zh", "en", "ja", "ko", "yue", "auto"], "size": ["large"], "path": "funasr_asr.py", "precision": ["float32"]},
+    "达摩 ASR (中文经典)": {"lang": ["zh", "yue"], "size": ["large"], "path": "funasr_asr.py", "precision": ["float32"]},
     "Faster Whisper (多语种)": {
         "lang": ["auto", "en", "ja", "ko"],
         "size": get_models(),

tools/asr/funasr_asr.py

@@ -11,9 +11,9 @@ from tqdm import tqdm
 funasr_models = {}  # 存储模型避免重复加载
 
 
-def only_asr(input_file, language):
+def only_asr(input_file, language, backend="fun-asr-nano"):
     try:
-        model = create_model(language)
+        model = create_model(language, backend=backend)
         text = model.generate(input=input_file)[0]["text"]
     except Exception:
         text = ""
@@ -21,7 +21,39 @@ def only_asr(input_file, language):
     return text
 
 
-def create_model(language="zh"):
+def create_model(language="zh", **kwargs):
+    backend = kwargs.get("backend", "fun-asr-nano")
+
+    # For non-classic backends, route to multilingual models regardless of language
+    if backend in ("fun-asr-nano", "sensevoice") and language != "yue":
+        import torch
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+        cache_key = f"{language}_{backend}"
+        if cache_key in funasr_models:
+            return funasr_models[cache_key]
+
+        if backend == "fun-asr-nano":
+            model = AutoModel(
+                model="FunAudioLLM/Fun-ASR-Nano-2512",
+                trust_remote_code=True,
+                hub="hf",
+                vad_model="fsmn-vad",
+                device=device,
+                disable_update=True,
+            )
+            print(f"FunASR Fun-ASR-Nano 模型加载完成: {language.upper()}")
+        else:
+            model = AutoModel(
+                model="iic/SenseVoiceSmall",
+                vad_model="fsmn-vad",
+                device=device,
+                disable_update=True,
+            )
+            print(f"FunASR SenseVoice 模型加载完成: {language.upper()}")
+
+        funasr_models[cache_key] = model
+        return model
+
     if language == "zh":
         path_vad = "tools/asr/models/speech_fsmn_vad_zh-cn-16k-common-pytorch"
         path_punc = "tools/asr/models/punc_ct-transformer_zh-cn-common-vocab272727-pytorch"
@@ -50,7 +82,7 @@ def create_model(language="zh"):
         vad_model_revision = punc_model_revision = ""
         model_revision = "master"
     else:
-        raise ValueError(f"{language} is not supported")
+        raise ValueError(f"{language} is not supported. Supported: zh, yue, ja, en, ko, auto")
 
     if language in funasr_models:
         return funasr_models[language]
@@ -69,14 +101,14 @@ def create_model(language="zh"):
         return model
 
 
-def execute_asr(input_folder, output_folder, model_size, language):
+def execute_asr(input_folder, output_folder, model_size, language, backend="fun-asr-nano"):
     input_file_names = os.listdir(input_folder)
     input_file_names.sort()
 
     output = []
     output_file_name = os.path.basename(input_folder)
 
-    model = create_model(language)
+    model = create_model(language, backend=backend)
 
     for file_name in tqdm(input_file_names):
         try:
@@ -105,7 +137,7 @@ if __name__ == "__main__":
     parser.add_argument("-o", "--output_folder", type=str, required=True, help="Output folder to store transcriptions.")
     parser.add_argument("-s", "--model_size", type=str, default="large", help="Model Size of FunASR is Large")
     parser.add_argument(
-        "-l", "--language", type=str, default="zh", choices=["zh", "yue", "auto"], help="Language of the audio files."
+        "-l", "--language", type=str, default="zh", choices=["zh", "yue", "ja", "en", "ko", "auto"], help="Language of the audio files."
     )
     parser.add_argument(
         "-p", "--precision", type=str, default="float16", choices=["float16", "float32"], help="fp16 or fp32"