Merge 539ad38d310d452aa1a08e659d5eac22f843a01d into 9ec3a60f30d228719e5ec6cd6796c5b2d888dd1a

* Add ModelScope Snapshot Download For ASR

* Typo Fix

* Remove YUE in whisper

* Remove HF ENDPOINT

* Add FunASR Download

api_v2.py

@@ -41,11 +41,9 @@ POST:
     "repetition_penalty": 1.35,   # float. repetition penalty for T2S model.
     "sample_steps": 32,           # int. number of sampling steps for VITS model V3.
     "super_sampling": False,      # bool. whether to use super-sampling for audio when using VITS model V3.
-    "return_fragment": False,     # bool. step by step return the audio fragment. (Best Quality, Slowest response speed. old version of streaming mode)
-    "streaming_mode": False,      # bool. return audio chunk by chunk. (Medium quality, Slow response speed)
+    "streaming_mode": False,      # bool or int. return audio chunk by chunk.T he available options are: 0,1,2,3 or True/False (0/False: Disabled | 1/True: Best Quality, Slowest response speed (old version streaming_mode) | 2: Medium Quality, Slow response speed | 3: Lower Quality, Faster response speed )
     "overlap_length": 2,          # int. overlap length of semantic tokens for streaming mode.
-    "min_chunk_length": 16,        # int. The minimum chunk length of semantic tokens for streaming mode. (affects audio chunk size)
-    "fixed_length_chunk": False,  # bool. When turned on, it can achieve faster streaming response, but with lower quality. (lower quality, faster response speed)
+    "min_chunk_length": 16,       # int. The minimum chunk length of semantic tokens for streaming mode. (affects audio chunk size)
 }
 ```
 
@@ -106,7 +104,7 @@ RESP:
 import os
 import sys
 import traceback
-from typing import Generator
+from typing import Generator, Union
 
 now_dir = os.getcwd()
 sys.path.append(now_dir)
@@ -171,15 +169,13 @@ class TTS_Request(BaseModel):
     fragment_interval: float = 0.3
     seed: int = -1
     media_type: str = "wav"
-    streaming_mode: bool = False
+    streaming_mode: Union[bool, int] = False
     parallel_infer: bool = True
     repetition_penalty: float = 1.35
     sample_steps: int = 32
     super_sampling: bool = False
     overlap_length: int = 2
     min_chunk_length: int = 16
-    return_fragment: bool = False
-    fixed_length_chunk: bool = False
 
 
 def pack_ogg(io_buffer: BytesIO, data: np.ndarray, rate: int):
@@ -373,11 +369,9 @@ async def tts_handle(req: dict):
                 "repetition_penalty": 1.35,   # float. repetition penalty for T2S model.
                 "sample_steps": 32,           # int. number of sampling steps for VITS model V3.
                 "super_sampling": False,      # bool. whether to use super-sampling for audio when using VITS model V3.
-                "return_fragment": False,     # bool. step by step return the audio fragment. (Best Quality, Slowest response speed. old version of streaming mode)
-                "streaming_mode": False,      # bool. return audio chunk by chunk. (Medium quality, Slow response speed)
+                "streaming_mode": False,      # bool or int. return audio chunk by chunk.T he available options are: 0,1,2,3 or True/False (0/False: Disabled | 1/True: Best Quality, Slowest response speed (old version streaming_mode) | 2: Medium Quality, Slow response speed | 3: Lower Quality, Faster response speed )
                 "overlap_length": 2,          # int. overlap length of semantic tokens for streaming mode.
-                "min_chunk_length": 16,        # int. The minimum chunk length of semantic tokens for streaming mode. (affects audio chunk size)
-                "fixed_length_chunk": False,  # bool. When turned on, it can achieve faster streaming response, but with lower quality. (lower quality, faster response speed)
+                "min_chunk_length": 16,       # int. The minimum chunk length of semantic tokens for streaming mode. (affects audio chunk size)
             }
     returns:
         StreamingResponse: audio stream response.
@@ -390,9 +384,33 @@ async def tts_handle(req: dict):
     check_res = check_params(req)
     if check_res is not None:
         return check_res
+    
+    if streaming_mode == 0:
+        streaming_mode = False
+        return_fragment = False
+        fixed_length_chunk = False
+    elif streaming_mode == 1:
+        streaming_mode = False
+        return_fragment = True
+        fixed_length_chunk = False
+    elif streaming_mode == 2:
+        streaming_mode = True
+        return_fragment = False
+        fixed_length_chunk = False
+    elif streaming_mode == 3:
+        streaming_mode = True
+        return_fragment = False
+        fixed_length_chunk = True
+
+    else:
+        return JSONResponse(status_code=400, content={"message": f"the value of streaming_mode must be 0, 1, 2, 3(int) or true/false(bool)"})
 
     req["streaming_mode"] = streaming_mode
     req["return_fragment"] = return_fragment
+    req["fixed_length_chunk"] = fixed_length_chunk
+
+    print(f"{streaming_mode} {return_fragment} {fixed_length_chunk}")
+
     streaming_mode = streaming_mode or return_fragment
 
 
@@ -457,11 +475,9 @@ async def tts_get_endpoint(
     repetition_penalty: float = 1.35,
     sample_steps: int = 32,
     super_sampling: bool = False,
-    return_fragment: bool = False,
-    streaming_mode: bool = False,
+    streaming_mode: Union[bool, int] = False,
     overlap_length: int = 2,
     min_chunk_length: int = 16,
-    fixed_length_chunk: bool = False,
 ):
     req = {
         "text": text,
@@ -488,8 +504,6 @@ async def tts_get_endpoint(
         "super_sampling": super_sampling,
         "overlap_length": int(overlap_length),
         "min_chunk_length": int(min_chunk_length),
-        "return_fragment": return_fragment,
-        "fixed_length_chunk": fixed_length_chunk
     }
     return await tts_handle(req)
 

go-webui-simple-mode.bat

@@ -0,0 +1,2 @@
+runtime\python.exe -I webui_simple.py zh_CN
+pause

requirements.txt

@@ -16,7 +16,7 @@ pypinyin
 pyopenjtalk>=0.4.1
 g2p_en
 torchaudio
-modelscope==1.10.0
+modelscope
 sentencepiece
 transformers>=4.43,<=4.50
 peft
@@ -39,7 +39,5 @@ x_transformers
 torchmetrics<=1.5
 pydantic<=2.10.6
 ctranslate2>=4.0,<5
-huggingface_hub>=0.13
-tokenizers>=0.13,<1
 av>=11
 tqdm

tools/asr/config.py

@@ -1,34 +1,13 @@
-import os
-
-
-def check_fw_local_models():
-    """
-    启动时检查本地是否有 Faster Whisper 模型.
-    """
-    model_size_list = [
-        "medium",
-        "medium.en",
-        "distil-large-v2",
-        "distil-large-v3",
-        "large-v1",
-        "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"
-    return model_size_list
-
-
 def get_models():
     model_size_list = [
         "medium",
         "medium.en",
-        "distil-large-v2",
-        "distil-large-v3",
-        "large-v1",
         "large-v2",
         "large-v3",
+        "large-v3-turbo",
+        #"distil-large-v2",
+        #"distil-large-v3",
+        #"distil-large-v3.5",
     ]
     return model_size_list
 
@@ -36,7 +15,7 @@ def get_models():
 asr_dict = {
     "达摩 ASR (中文)": {"lang": ["zh", "yue"], "size": ["large"], "path": "funasr_asr.py", "precision": ["float32"]},
     "Faster Whisper (多语种)": {
-        "lang": ["auto", "zh", "en", "ja", "ko", "yue"],
+        "lang": ["auto", "en", "ja", "ko"],
         "size": get_models(),
         "path": "fasterwhisper_asr.py",
         "precision": ["float32", "float16", "int8"],

tools/asr/fasterwhisper_asr.py

@@ -1,12 +1,12 @@
 import argparse
 import os
-import time
 import traceback
 
+import requests
 import torch
 from faster_whisper import WhisperModel
-from huggingface_hub import snapshot_download
-from huggingface_hub.errors import LocalEntryNotFoundError
+from huggingface_hub import snapshot_download as snapshot_download_hf
+from modelscope import snapshot_download as snapshot_download_ms
 from tqdm import tqdm
 
 from tools.asr.config import get_models
@@ -40,11 +40,32 @@ language_code_list = [
 
 
 def download_model(model_size: str):
-    if "distil" in model_size:
-        repo_id = "Systran/faster-{}-whisper-{}".format(*model_size.split("-", maxsplit=1))
+    url = "https://huggingface.co/api/models/gpt2"
+    try:
+        requests.get(url, timeout=3)
+        source = "HF"
+    except Exception:
+        source = "ModelScope"
+
+    model_path = ""
+    if source == "HF":
+        if "distil" in model_size:
+            if "3.5" in model_size:
+                repo_id = "distil-whisper/distil-large-v3.5-ct2"
+                model_path = "tools/asr/models/faster-distil-whisper-large-v3.5"
+            else:
+                repo_id = "Systran/faster-{}-whisper-{}".format(*model_size.split("-", maxsplit=1))
+        elif model_size == "large-v3-turbo":
+            repo_id = "mobiuslabsgmbh/faster-whisper-large-v3-turbo"
+            model_path = "tools/asr/models/faster-whisper-large-v3-turbo"
+        else:
+            repo_id = f"Systran/faster-whisper-{model_size}"
+        model_path = (
+            model_path or f"tools/asr/models/{repo_id.replace('Systran/', '').replace('distil-whisper/', '', 1)}"
+        )
     else:
-        repo_id = f"Systran/faster-whisper-{model_size}"
-    model_path = f"tools/asr/models/{repo_id.strip('Systran/')}"
+        repo_id = "XXXXRT/faster-whisper"
+        model_path = "tools/asr/models"
 
     files: list[str] = [
         "config.json",
@@ -52,32 +73,31 @@ def download_model(model_size: str):
         "tokenizer.json",
         "vocabulary.txt",
     ]
-    if model_size == "large-v3" or "distil" in model_size:
+    if "large-v3" in model_size or "distil" in model_size:
         files.append("preprocessor_config.json")
         files.append("vocabulary.json")
 
         files.remove("vocabulary.txt")
 
-    for attempt in range(2):
-        try:
-            snapshot_download(
-                repo_id=repo_id,
-                allow_patterns=files,
-                local_dir=model_path,
-            )
-            break
-        except LocalEntryNotFoundError:
-            if attempt < 1:
-                time.sleep(2)
-            else:
-                print("[ERROR] LocalEntryNotFoundError and no fallback.")
-                traceback.print_exc()
-                exit(1)
-        except Exception as e:
-            print(f"[ERROR] Unexpected error on attempt {attempt + 1}: {e}")
-            traceback.print_exc()
-            exit(1)
+    if source == "ModelScope":
+        files = [f"faster-whisper-{model_size}/{file}".replace("whisper-distil", "distil-whisper") for file in files]
 
+    if source == "HF":
+        print(f"Downloading model from HuggingFace: {repo_id} to {model_path}")
+        snapshot_download_hf(
+            repo_id,
+            local_dir=model_path,
+            local_dir_use_symlinks=False,
+            allow_patterns=files,
+        )
+    else:
+        print(f"Downloading model from ModelScope: {repo_id} to {model_path}")
+        snapshot_download_ms(
+            repo_id,
+            local_dir=model_path,
+            allow_patterns=files,
+        )
+        return model_path + f"/faster-whisper-{model_size}".replace("whisper-distil", "distil-whisper")
     return model_path
 
 
@@ -106,7 +126,7 @@ def execute_asr(input_folder, output_folder, model_path, language, precision):
             )
             text = ""
 
-            if info.language == "zh":
+            if info.language in ["zh", "yue"]:
                 print("检测为中文文本, 转 FunASR 处理")
                 text = only_asr(file_path, language=info.language.lower())
 

tools/asr/funasr_asr.py

@@ -4,9 +4,8 @@ import argparse
 import os
 import traceback
 
-# from funasr.utils import version_checker
-# version_checker.check_for_update = lambda: None
 from funasr import AutoModel
+from modelscope import snapshot_download
 from tqdm import tqdm
 
 funasr_models = {}  # 存储模型避免重复加载
@@ -16,40 +15,43 @@ def only_asr(input_file, language):
     try:
         model = create_model(language)
         text = model.generate(input=input_file)[0]["text"]
-    except:
+    except Exception:
         text = ""
         print(traceback.format_exc())
     return text
 
 
 def create_model(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"
-    path_vad = path_vad if os.path.exists(path_vad) else "iic/speech_fsmn_vad_zh-cn-16k-common-pytorch"
-    path_punc = path_punc if os.path.exists(path_punc) else "iic/punc_ct-transformer_zh-cn-common-vocab272727-pytorch"
-    vad_model_revision = punc_model_revision = "v2.0.4"
-
     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"
         path_asr = "tools/asr/models/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
-        path_asr = (
-            path_asr
-            if os.path.exists(path_asr)
-            else "iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
+        snapshot_download(
+            "iic/speech_fsmn_vad_zh-cn-16k-common-pytorch",
+            local_dir="tools/asr/models/speech_fsmn_vad_zh-cn-16k-common-pytorch",
+        )
+        snapshot_download(
+            "iic/punc_ct-transformer_zh-cn-common-vocab272727-pytorch",
+            local_dir="tools/asr/models/punc_ct-transformer_zh-cn-common-vocab272727-pytorch",
+        )
+        snapshot_download(
+            "iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch",
+            local_dir="tools/asr/models/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch",
         )
         model_revision = "v2.0.4"
     elif language == "yue":
         path_asr = "tools/asr/models/speech_UniASR_asr_2pass-cantonese-CHS-16k-common-vocab1468-tensorflow1-online"
-        path_asr = (
-            path_asr
-            if os.path.exists(path_asr)
-            else "iic/speech_UniASR_asr_2pass-cantonese-CHS-16k-common-vocab1468-tensorflow1-online"
+        snapshot_download(
+            "iic/speech_UniASR_asr_2pass-cantonese-CHS-16k-common-vocab1468-tensorflow1-online",
+            local_dir="tools/asr/models/speech_UniASR_asr_2pass-cantonese-CHS-16k-common-vocab1468-tensorflow1-online",
         )
-        model_revision = "master"
         path_vad = path_punc = None
-        vad_model_revision = punc_model_revision = None
-        ###友情提示：粤语带VAD识别可能会有少量shape不对报错的，但是不带VAD可以.不带vad只能分阶段单独加标点。不过标点模型对粤语效果真的不行…
+        vad_model_revision = punc_model_revision = ""
+        model_revision = "master"
     else:
-        raise ValueError("FunASR 不支持该语言" + ": " + language)
+        raise ValueError(f"{language} is not supported")
+
+    vad_model_revision = punc_model_revision = "v2.0.4"
 
     if language in funasr_models:
         return funasr_models[language]
@@ -83,7 +85,7 @@ def execute_asr(input_folder, output_folder, model_size, language):
             file_path = os.path.join(input_folder, file_name)
             text = model.generate(input=file_path)[0]["text"]
             output.append(f"{file_path}|{output_file_name}|{language.upper()}|{text}")
-        except:
+        except Exception:
             print(traceback.format_exc())
 
     output_folder = output_folder or "output/asr_opt"

tools/i18n/locale/en_US.json

@@ -38,7 +38,7 @@
     "hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）": "hop_size: FO hop size, the smaller the value, the higher the accuracy）",
     "max:归一化后最大值多少": "Loudness multiplier after normalized",
     "max_sil_kept:切完后静音最多留多长": "Maximum length for silence to be kept",
-    "min_interval:最短切割间隔": "Minumum interval for audio cutting",
+    "min_interval:最短切割间隔": "Minimum interval for audio cutting",
     "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: the minimum length of each segment. If the first segment is too short, it will be concatenated with the next segment until it exceeds this value",
     "temperature": "temperature",
     "threshold:音量小于这个值视作静音的备选切割点": "Noise gate threshold (loudness below this value will be treated as noise",
@@ -176,7 +176,7 @@
     "语音降噪": "Speech Denoising",
     "请上传3~10秒内参考音频，超过会报错！": "Please upload a reference audio within the 3-10 second range; if it exceeds this duration, it will raise errors.",
     "请上传参考音频": "Please Upload the Reference Audio",
-    "请填入推理文本": "Please Fill in the Terget Text",
+    "请填入推理文本": "Please Fill in the Target Text",
     "请填入正确的List路径": "Please Fill in the Correct List Path",
     "请填入正确的音频文件夹路径": "Please Fill in the Correct Audio Folder Path",
     "请输入有效文本": "Please enter valid text.",

webui.py

@@ -86,7 +86,6 @@ from config import (
 from tools import my_utils
 from tools.my_utils import check_details, check_for_existance
 
-os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"
 os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
 
 # os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' # 当遇到mps不支持的步骤时使用cpu
@@ -117,8 +116,8 @@ def set_default():
     gpu_info = "\n".join(gpu_infos)
     if is_gpu_ok:
         minmem = min(mem)
-        default_batch_size = minmem // 2 if version not in v3v4set else minmem // 8
-        default_batch_size_s1 = minmem // 2
+        default_batch_size = int(minmem // 2 if version not in v3v4set else minmem // 8)
+        default_batch_size_s1 = int(minmem // 2)
     else:
         default_batch_size = default_batch_size_s1 = int(psutil.virtual_memory().total / 1024 / 1024 / 1024 / 4)
     if version not in v3v4set:
@@ -1980,4 +1979,3 @@ with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False, js=js, css=css
         server_port=webui_port_main,
         # quiet=True,
     )
-