Merge 485cb85a552c5bd5b6387569da9b1727332e940e into c767f0b83b998e996a4d230d86da575a03f54a3f

* 修复bug

* fallbak and bug fix

GPT_SoVITS/AR/models/t2s_model.py

@@ -707,10 +707,12 @@ class Text2SemanticDecoder(nn.Module):
 
             if idx == 0:
                 attn_mask = F.pad(attn_mask[:, :, -1].unsqueeze(-2), (0, 1), value=False)
-                logits = logits[:, :-1]
             else:
                 attn_mask = F.pad(attn_mask, (0, 1), value=False)
 
+            if idx < 11:  ###至少预测出10个token不然不给停止（0.4s）
+                logits = logits[:, :-1] 
+
             samples = sample(
                 logits, y, top_k=top_k, top_p=top_p, repetition_penalty=repetition_penalty, temperature=temperature
             )[0]

GPT_SoVITS/TTS_infer_pack/TTS.py

@@ -1008,7 +1008,7 @@ class TTS:
                     "aux_ref_audio_paths": [],    # list.(optional) auxiliary reference audio paths for multi-speaker tone fusion
                     "prompt_text": "",            # str.(optional) prompt text for the reference audio
                     "prompt_lang": "",            # str.(required) language of the prompt text for the reference audio
-                    "top_k": 5,                   # int. top k sampling
+                    "top_k": 15,                  # int. top k sampling
                     "top_p": 1,                   # float. top p sampling
                     "temperature": 1,             # float. temperature for sampling
                     "text_split_method": "cut1",  # str. text split method, see text_segmentation_method.py for details.
@@ -1039,7 +1039,7 @@ class TTS:
         aux_ref_audio_paths: list = inputs.get("aux_ref_audio_paths", [])
         prompt_text: str = inputs.get("prompt_text", "")
         prompt_lang: str = inputs.get("prompt_lang", "")
-        top_k: int = inputs.get("top_k", 5)
+        top_k: int = inputs.get("top_k", 15)
         top_p: float = inputs.get("top_p", 1)
         temperature: float = inputs.get("temperature", 1)
         text_split_method: str = inputs.get("text_split_method", "cut1")

GPT_SoVITS/inference_cli.py

@@ -7,79 +7,163 @@ from GPT_SoVITS.inference_webui import change_gpt_weights, change_sovits_weights
 
 i18n = I18nAuto()
 
+LANGUAGE_CHOICES = ["中文", "英文", "日文", "韩文", "粤语"]
+MIXED_LANGUAGE_CHOICES = ["中英混合", "日英混合", "粤英混合", "韩英混合", "多语种混合"]
 
-def synthesize(
-    GPT_model_path,
-    SoVITS_model_path,
-    ref_audio_path,
-    ref_text_path,
-    ref_language,
-    target_text_path,
-    target_language,
-    output_path,
-):
-    # Read reference text
-    with open(ref_text_path, "r", encoding="utf-8") as file:
-        ref_text = file.read()
+SLICE_METHOD_CHOICES = ["凑四句一切", "凑50字一切", "按标点符号切", "按中文句号。切", "按英文句号.切"]
 
-    # Read target text
-    with open(target_text_path, "r", encoding="utf-8") as file:
-        target_text = file.read()
+def synthesize(args: argparse.Namespace):
 
     # Change model weights
-    change_gpt_weights(gpt_path=GPT_model_path)
-    change_sovits_weights(sovits_path=SoVITS_model_path)
+    change_gpt_weights(gpt_path=args.gpt_path)
+    change_sovits_weights(sovits_path=args.sovits_path)
+
+    params = {
+        "ref_wav_path": args.ref_audio,
+        "prompt_text": args.ref_text,
+        "prompt_language": i18n(args.ref_language),
+        "text": args.target_text,
+        "text_language": i18n(args.target_language),
+    }
+
+    # region - optional params
+    if args.slicer: params["how_to_cut"] = i18n(args.slicer)
+    if args.top_k: params["top_k"] = args.top_k
+    if args.top_p: params["top_p"] = args.top_p
+    if args.temperature: params["temperature"] = args.temperature
+    if args.ref_free: params["ref_free"] = args.ref_free
+    if args.speed: params["speed"] = args.speed
+    if args.if_freeze: params["if_freeze"] = args.if_freeze
+    if args.inp_refs: params["inp_refs"] = args.inp_refs
+    if args.sample_steps: params["sample_steps"] = args.sample_steps
+    if args.if_sr: params["if_sr"] = args.if_sr
+    if args.pause_second: params["pause_second"] = args.pause_second
+    # endregion - optional params
 
     # Synthesize audio
-    synthesis_result = get_tts_wav(
-        ref_wav_path=ref_audio_path,
-        prompt_text=ref_text,
-        prompt_language=i18n(ref_language),
-        text=target_text,
-        text_language=i18n(target_language),
-        top_p=1,
-        temperature=1,
-    )
+    synthesis_result = get_tts_wav(**params)
 
     result_list = list(synthesis_result)
 
     if result_list:
+        os.makedirs(args.output_path, exist_ok=True) # Create output directory if it doesn't exist
+        if args.output_path.endswith(".wav"):
+            output_wav_path = args.output_path
+        else:
+            output_wav_path = os.path.join(args.output_path, "output.wav")
         last_sampling_rate, last_audio_data = result_list[-1]
-        output_wav_path = os.path.join(output_path, "output.wav")
         sf.write(output_wav_path, last_audio_data, last_sampling_rate)
         print(f"Audio saved to {output_wav_path}")
 
 
-def main():
+def build_parser():
     parser = argparse.ArgumentParser(description="GPT-SoVITS Command Line Tool")
-    parser.add_argument("--gpt_model", required=True, help="Path to the GPT model file")
-    parser.add_argument("--sovits_model", required=True, help="Path to the SoVITS model file")
+
+    # reference settings
     parser.add_argument("--ref_audio", required=True, help="Path to the reference audio file")
-    parser.add_argument("--ref_text", required=True, help="Path to the reference text file")
-    parser.add_argument(
-        "--ref_language", required=True, choices=["中文", "英文", "日文"], help="Language of the reference audio"
-    )
-    parser.add_argument("--target_text", required=True, help="Path to the target text file")
-    parser.add_argument(
-        "--target_language",
-        required=True,
-        choices=["中文", "英文", "日文", "中英混合", "日英混合", "多语种混合"],
-        help="Language of the target text",
-    )
+    parser.add_argument("--ref_text", required=True, help="Transcript of the reference audio")
+    parser.add_argument("--ref_language", required=True,
+                       choices=LANGUAGE_CHOICES, help="Language of the reference audio")
+
+    # output settings
+    parser.add_argument("--target_text", required=True, help="Text to be synthesized")
+    parser.add_argument("--target_language", required=True,
+                       choices=LANGUAGE_CHOICES+MIXED_LANGUAGE_CHOICES,
+                       help="Language of the target text")
+    parser.add_argument("--slicer", required=False,
+                       choices=SLICE_METHOD_CHOICES, help="Slicer method")
     parser.add_argument("--output_path", required=True, help="Path to the output directory")
 
+    # region - inference settings
+    parser.add_argument("--top_k", required=False, type=int, help="Top-k value")
+    parser.add_argument("--top_p", required=False, type=float, help="Top-p value")
+    parser.add_argument("--temperature", required=False, type=float, help="Temperature value")
+    parser.add_argument("--ref_free", required=False, type=bool, help="Reference free value")
+    parser.add_argument("--speed", required=False, type=float, help="Speed value")
+    parser.add_argument("--if_freeze", required=False, type=bool, help="If freeze value")
+    parser.add_argument("--inp_refs", required=False, type=str, help="Input references")
+    parser.add_argument("--sample_steps", required=False, type=int, help="Sample steps value")
+    parser.add_argument("--if_sr", required=False, type=bool, help="If super resolution value")
+    parser.add_argument("--pause_second", required=False, type=float, help="Pause second value")
+    # endregion - inference settings
+
+    # region - model selection
+    sub = parser.add_subparsers(dest="mode", required=True)
+
+    # Mode 1: provide model paths directly
+    p_paths = sub.add_parser("paths", help="Use explicit model file paths")
+    p_paths.add_argument("--gpt_path", required=True, help="Path to the GPT model file")
+    p_paths.add_argument("--sovits_path", required=True, help="Path to the SoVITS model file")
+
+    # Mode 2: select by experiment/version
+    p_sel = sub.add_parser("select", help="Select models by experiment/version")
+    p_sel.add_argument("--exp_name", required=True, help="Experiment name")
+    available_gpt_versions = ["v1", "v2", "v2Pro", "v2ProPlus", "v3", "v4"]
+    p_sel.add_argument("--gpt_version", required=True, choices=available_gpt_versions, help="Version of the GPT model")
+    available_sovits_versions = ["v1", "v2", "v2Pro", "v2ProPlus", "v3", "v4"]
+    p_sel.add_argument("--sovits_version", required=True, choices=available_sovits_versions, help="Version of the SoVITS model")
+    p_sel.add_argument("--gpt_epoch", type=int, help="Epoch of the GPT model")
+    p_sel.add_argument("--sovits_epoch", type=int, help="Epoch of the SoVITS model")
+    # endregion - model selection
+
+    return parser
+
+
+def get_model_path(args)->argparse.Namespace:
+    """
+    Get the model path from exp_name, version and epoch
+
+    Args:
+        args: argparse.Namespace
+
+    Returns:
+        args: argparse.Namespace
+    """
+    exist_gpt_path = []
+    exist_sovits_path = []
+
+    def _get_model_dir(model_type, version):
+        if version == "v1":
+            return f"{model_type}_weights"
+        else:
+            return f"{model_type}_weights_{version}"
+
+    # get all the model paths with the same exp_name
+    for files in os.listdir(_get_model_dir("GPT", args.gpt_version)):
+        if args.exp_name in files and files.endswith(".ckpt"):
+            exist_gpt_path.append(os.path.join(_get_model_dir("GPT", args.gpt_version), files))
+    for files in os.listdir(_get_model_dir("SoVITS", args.sovits_version)):
+        if args.exp_name in files and files.endswith(".pth"):
+            exist_sovits_path.append(os.path.join(_get_model_dir("SoVITS", args.sovits_version), files))
+
+    # get the largest epoch if not specified
+    if args.gpt_epoch:
+        args.gpt_path = [i for i in exist_gpt_path if f"e{args.gpt_epoch}" in i]
+    else:
+        args.gpt_path = sorted(exist_gpt_path)[-1]
+    if args.sovits_epoch:
+        args.sovits_path = [i for i in exist_sovits_path if f"e{args.sovits_epoch}" in i]
+    else:
+        args.sovits_path = sorted(exist_sovits_path)[-1]
+
+    if not args.gpt_path or not args.sovits_path:
+        raise ValueError("No model found")
+    
+    return args
+
+
+def main():
+    parser = build_parser()
     args = parser.parse_args()
 
-    synthesize(
-        args.gpt_model,
-        args.sovits_model,
-        args.ref_audio,
-        args.ref_text,
-        args.ref_language,
-        args.target_text,
-        args.target_language,
-        args.output_path,
-    )
+    print(args)
+    if args.mode == "select":
+        args = get_model_path(args)
+    
+    args.target_text = args.target_text.replace("'", "").replace('"', "")
+
+
+    synthesize(args)
 
 
 if __name__ == "__main__":

GPT_SoVITS/inference_webui.py

@@ -180,6 +180,8 @@ def get_bert_feature(text, word2ph):
     for i in range(len(word2ph)):
         repeat_feature = res[i].repeat(word2ph[i], 1)
         phone_level_feature.append(repeat_feature)
+    if len(phone_level_feature) == 0:
+        return torch.empty((res.shape[1], 0), dtype=res.dtype, device=res.device)
     phone_level_feature = torch.cat(phone_level_feature, dim=0)
     return phone_level_feature.T
 

GPT_SoVITS/inference_webui_fast.py

@@ -385,7 +385,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False, js=js, css=css
                         minimum=0.6, maximum=1.65, step=0.05, label="语速", value=1.0, interactive=True
                     )
                 with gr.Row():
-                    top_k = gr.Slider(minimum=1, maximum=100, step=1, label=i18n("top_k"), value=5, interactive=True)
+                    top_k = gr.Slider(minimum=1, maximum=100, step=1, label=i18n("top_k"), value=15, interactive=True)
                     top_p = gr.Slider(minimum=0, maximum=1, step=0.05, label=i18n("top_p"), value=1, interactive=True)
                 with gr.Row():
                     temperature = gr.Slider(

README.md

@@ -347,7 +347,7 @@ Use v4 from v1/v2/v3 environment:
 
 2. Clone the latest codes from github.
 
-3. Download v4 pretrained models (gsv-v4-pretrained/s2v4.ckpt, and gsv-v4-pretrained/vocoder.pth) from [huggingface](https://huggingface.co/lj1995/GPT-SoVITS/tree/main) and put them into `GPT_SoVITS/pretrained_models`.
+3. Download v4 pretrained models (gsv-v4-pretrained/s2v4.pth, and gsv-v4-pretrained/vocoder.pth) from [huggingface](https://huggingface.co/lj1995/GPT-SoVITS/tree/main) and put them into `GPT_SoVITS/pretrained_models`.
 
 ## V2Pro Release Notes
 

api_v2.py

@@ -27,7 +27,7 @@ POST:
     "aux_ref_audio_paths": [],    # list.(optional) auxiliary reference audio paths for multi-speaker tone fusion
     "prompt_text": "",            # str.(optional) prompt text for the reference audio
     "prompt_lang": "",            # str.(required) language of the prompt text for the reference audio
-    "top_k": 5,                   # int. top k sampling
+    "top_k": 15,                  # int. top k sampling
     "top_p": 1,                   # float. top p sampling
     "temperature": 1,             # float. temperature for sampling
     "text_split_method": "cut5",  # str. text split method, see text_segmentation_method.py for details.
@@ -158,7 +158,7 @@ class TTS_Request(BaseModel):
     aux_ref_audio_paths: list = None
     prompt_lang: str = None
     prompt_text: str = ""
-    top_k: int = 5
+    top_k: int = 15
     top_p: float = 1
     temperature: float = 1
     text_split_method: str = "cut5"
@@ -355,7 +355,7 @@ async def tts_handle(req: dict):
                 "aux_ref_audio_paths": [],    # list.(optional) auxiliary reference audio paths for multi-speaker tone fusion
                 "prompt_text": "",            # str.(optional) prompt text for the reference audio
                 "prompt_lang": "",            # str.(required) language of the prompt text for the reference audio
-                "top_k": 5,                   # int. top k sampling
+                "top_k": 15,                  # int. top k sampling
                 "top_p": 1,                   # float. top p sampling
                 "temperature": 1,             # float. temperature for sampling
                 "text_split_method": "cut5",  # str. text split method, see text_segmentation_method.py for details.
@@ -460,7 +460,7 @@ async def tts_get_endpoint(
     aux_ref_audio_paths: list = None,
     prompt_lang: str = None,
     prompt_text: str = "",
-    top_k: int = 5,
+    top_k: int = 15,
     top_p: float = 1,
     temperature: float = 1,
     text_split_method: str = "cut5",

requirements.txt

@@ -19,7 +19,7 @@ torchaudio
 modelscope
 sentencepiece
 transformers>=4.43,<=4.50
-peft
+peft<0.18.0
 chardet
 PyYAML
 psutil