Merge c93bf48785aba24ee47a268f1eb5dcee1b8866dd into ea2d2a81667239d37615697e8f0056e35bab2db6

* 调整日志格式

* docs: Update other languages' changelogs

GPT_SoVITS/AR/data/dataset.py

@@ -67,8 +67,10 @@ class Text2SemanticDataset(Dataset):
             )
         )  # "%s/3-bert"%exp_dir#bert_dir
         self.path6 = semantic_path  # "%s/6-name2semantic.tsv"%exp_dir#semantic_path
-        assert os.path.exists(self.path2)
+        if not os.path.exists(self.path2):
-        assert os.path.exists(self.path6)
+            raise FileNotFoundError(f"Phoneme data file not found: {self.path2}")
+        if not os.path.exists(self.path6):
+            raise FileNotFoundError(f"Semantic data file not found: {self.path6}")
         self.phoneme_data = {}
         with open(self.path2, "r", encoding="utf8") as f:
             lines = f.read().strip("\n").split("\n")
@@ -131,7 +133,7 @@ class Text2SemanticDataset(Dataset):
                 phoneme, word2ph, text = self.phoneme_data[item_name]
             except Exception:
                 traceback.print_exc()
-                # print(f"{item_name} not in self.phoneme_data !")
+                print(f"Warning: File \"{item_name}\" not in self.phoneme_data! Skipped. ")
                 num_not_in += 1
                 continue
 
@@ -152,7 +154,7 @@ class Text2SemanticDataset(Dataset):
                 phoneme_ids = cleaned_text_to_sequence(phoneme, version)
             except:
                 traceback.print_exc()
-                # print(f"{item_name} not in self.phoneme_data !")
+                print(f"Warning: Failed to convert phonemes to sequence for file \"{item_name}\"! Skipped. ")
                 num_not_in += 1
                 continue
             # if len(phoneme_ids) >400:###########2：改为恒定限制为semantic/2.5就行
@@ -228,7 +230,11 @@ class Text2SemanticDataset(Dataset):
             # bert_feature=torch.zeros_like(phoneme_ids,dtype=torch.float32)
             bert_feature = None
         else:
-            assert bert_feature.shape[-1] == len(phoneme_ids)
+            try:
+                assert bert_feature.shape[-1] == len(phoneme_ids)
+            except AssertionError:
+                print(f"AssertionError: The BERT feature dimension ({bert_feature.shape[-1]}) of the file '{item_name}' does not match the length of the phoneme sequence ({len(phoneme_ids)}).")
+                raise
         return {
             "idx": idx,
             "phoneme_ids": phoneme_ids,

GPT_SoVITS/AR/models/utils.py

@@ -262,7 +262,7 @@ def make_reject_y(y_o, y_lens):
     reject_y = []
     reject_y_lens = []
     for b in range(bs):
-        process_item_idx = torch.randint(0, 1, size=(1,))[0]
+        process_item_idx = torch.randint(0, 2, size=(1,))[0]
         if process_item_idx == 0:
             new_y = repeat_P(y_o[b])
             reject_y.append(new_y)

GPT_SoVITS/AR/modules/patched_mha_with_cache_onnx.py

@@ -8,30 +8,30 @@ def multi_head_attention_forward_patched(
     query,
     key,
     value,
-    embed_dim_to_check: int,
+    embed_dim_to_check,
-    num_heads: int,
+    num_heads,
     in_proj_weight,
-    in_proj_bias: Optional[Tensor],
+    in_proj_bias,
-    bias_k: Optional[Tensor],
+    bias_k,
-    bias_v: Optional[Tensor],
+    bias_v,
-    add_zero_attn: bool,
+    add_zero_attn,
-    dropout_p: float,
+    dropout_p,
-    out_proj_weight: Tensor,
+    out_proj_weight,
-    out_proj_bias: Optional[Tensor],
+    out_proj_bias,
-    training: bool = True,
+    training=True,
-    key_padding_mask: Optional[Tensor] = None,
+    key_padding_mask=None,
-    need_weights: bool = True,
+    need_weights=True,
-    attn_mask: Optional[Tensor] = None,
+    attn_mask=None,
-    use_separate_proj_weight: bool = False,
+    use_separate_proj_weight=False,
-    q_proj_weight: Optional[Tensor] = None,
+    q_proj_weight=None,
-    k_proj_weight: Optional[Tensor] = None,
+    k_proj_weight=None,
-    v_proj_weight: Optional[Tensor] = None,
+    v_proj_weight=None,
-    static_k: Optional[Tensor] = None,
+    static_k=None,
-    static_v: Optional[Tensor] = None,
+    static_v=None,
-    average_attn_weights: bool = True,
+    average_attn_weights=True,
-    is_causal: bool = False,
+    is_causal=False,
     cache=None,
-) -> Tuple[Tensor, Optional[Tensor]]:
+):
     # set up shape vars
     _, _, embed_dim = query.shape
     attn_mask = _canonical_mask(

GPT_SoVITS/TTS_infer_pack/TTS.py

@@ -499,7 +499,7 @@ class TTS:
 
         if if_lora_v3 == True and os.path.exists(path_sovits) == False:
             info = path_sovits + i18n("SoVITS %s 底模缺失，无法加载相应 LoRA 权重" % model_version)
-            raise FileExistsError(info)
+            raise FileNotFoundError(info)
 
         # dict_s2 = torch.load(weights_path, map_location=self.configs.device,weights_only=False)
         dict_s2 = load_sovits_new(weights_path)
@@ -1578,16 +1578,15 @@ class TTS:
                 max_audio = np.abs(audio).max()
                 if max_audio > 1:
                     audio /= max_audio
-            audio = (audio * 32768).astype(np.int16)
+                audio = (audio * 32768).astype(np.int16)
+            else:
+                audio = audio.cpu().numpy()
+                audio = (audio * 32768).astype(np.int16)
             t2 = time.perf_counter()
             print(f"超采样用时：{t2 - t1:.3f}s")
         else:
-            # audio = audio.float() * 32768
-            # audio = audio.to(dtype=torch.int16).clamp(-32768, 32767).cpu().numpy()
-
             audio = audio.cpu().numpy()
-
+            audio = (audio * 32768).astype(np.int16)
-        audio = (audio * 32768).astype(np.int16)
 
 
         # try:
@@ -1768,7 +1767,10 @@ class TTS:
             pos += chunk_len * upsample_rate
 
         audio = self.sola_algorithm(audio_fragments, overlapped_len * upsample_rate)
-        audio = audio[overlapped_len * upsample_rate : -padding_len * upsample_rate]
+        if padding_len > 0:
+            audio = audio[overlapped_len * upsample_rate : -padding_len * upsample_rate]
+        else:
+            audio = audio[overlapped_len * upsample_rate :]
 
         audio_fragments = []
         for feat_len in feat_lens:

GPT_SoVITS/TTS_infer_pack/text_segmentation_method.py

@@ -92,7 +92,7 @@ def cut0(inp):
     if not set(inp).issubset(punctuation):
         return inp
     else:
-        return "/n"
+        return "\n"
 
 
 # 凑四句一切

GPT_SoVITS/module/distrib.py

@@ -87,7 +87,7 @@ def sync_buffer(buffers, average=True):
     for buffer, handle in handles:
         handle.wait()
         if average:
-            buffer.data /= world_size
+            buffer.data /= world_size()
 
 
 def sync_grad(params):

GPT_SoVITS/s2_train_v3_lora.py

@@ -55,6 +55,10 @@ def main():
         n_gpus = torch.cuda.device_count()
     else:
         n_gpus = 1
+    if n_gpus <= 1:
+        run(0, n_gpus, hps)
+        return
+
     os.environ["MASTER_ADDR"] = "localhost"
     os.environ["MASTER_PORT"] = str(randint(20000, 55555))
 
@@ -77,12 +81,14 @@ def run(rank, n_gpus, hps):
         writer = SummaryWriter(log_dir=hps.s2_ckpt_dir)
         writer_eval = SummaryWriter(log_dir=os.path.join(hps.s2_ckpt_dir, "eval"))
 
-    dist.init_process_group(
+    use_ddp = n_gpus > 1
-        backend="gloo" if os.name == "nt" or not torch.cuda.is_available() else "nccl",
+    if use_ddp:
-        init_method="env://?use_libuv=False",
+        dist.init_process_group(
-        world_size=n_gpus,
+            backend="gloo" if os.name == "nt" or not torch.cuda.is_available() else "nccl",
-        rank=rank,
+            init_method="env://?use_libuv=False",
-    )
+            world_size=n_gpus,
+            rank=rank,
+        )
     torch.manual_seed(hps.train.seed)
     if torch.cuda.is_available():
         torch.cuda.set_device(rank)
@@ -118,15 +124,20 @@ def run(rank, n_gpus, hps):
         shuffle=True,
     )
     collate_fn = TextAudioSpeakerCollate()
-    train_loader = DataLoader(
+    worker_count = 0 if os.name == "nt" and n_gpus <= 1 else min(2 if os.name == "nt" else 5, os.cpu_count() or 1)
-        train_dataset,
+    loader_kwargs = dict(
-        num_workers=5,
+        num_workers=worker_count,
         shuffle=False,
-        pin_memory=True,
+        pin_memory=torch.cuda.is_available(),
         collate_fn=collate_fn,
         batch_sampler=train_sampler,
-        persistent_workers=True,
+    )
-        prefetch_factor=3,
+    if worker_count > 0:
+        loader_kwargs["persistent_workers"] = True
+        loader_kwargs["prefetch_factor"] = 2 if os.name == "nt" else 3
+    train_loader = DataLoader(
+        train_dataset,
+        **loader_kwargs,
     )
     save_root = "%s/logs_s2_%s_lora_%s" % (hps.data.exp_dir, hps.model.version, hps.train.lora_rank)
     os.makedirs(save_root, exist_ok=True)
@@ -156,7 +167,9 @@ def run(rank, n_gpus, hps):
 
     def model2cuda(net_g, rank):
         if torch.cuda.is_available():
-            net_g = DDP(net_g.cuda(rank), device_ids=[rank], find_unused_parameters=True)
+            net_g = net_g.cuda(rank)
+            if use_ddp:
+                net_g = DDP(net_g, device_ids=[rank], find_unused_parameters=True)
         else:
             net_g = net_g.to(device)
         return net_g
@@ -242,6 +255,8 @@ def run(rank, n_gpus, hps):
                 None,
             )
         scheduler_g.step()
+    if use_ddp and dist.is_initialized():
+        dist.destroy_process_group()
     print("training done")
 
 

GPT_SoVITS/text/chinese2.py

@@ -180,10 +180,15 @@ def _merge_erhua(initials: list[str], finals: list[str], word: str, pos: str) ->
 def _g2p(segments):
     phones_list = []
     word2ph = []
-    for seg in segments:
+    g2pw_batch_results = []
+    g2pw_batch_cursor = 0
+    processed_segments = [re.sub("[a-zA-Z]+", "", seg) for seg in segments]
+    if is_g2pw:
+        batch_inputs = [seg for seg in processed_segments if seg]
+        g2pw_batch_results = g2pw._g2pw(batch_inputs) if batch_inputs else []
+
+    for seg in processed_segments:
         pinyins = []
-        # Replace all English words in the sentence
-        seg = re.sub("[a-zA-Z]+", "", seg)
         seg_cut = psg.lcut(seg)
         seg_cut = tone_modifier.pre_merge_for_modify(seg_cut)
         initials = []
@@ -204,8 +209,10 @@ def _g2p(segments):
             finals = sum(finals, [])
             print("pypinyin结果", initials, finals)
         else:
-            # g2pw采用整句推理
+            # g2pw采用整句推理（批量推理，逐句取结果）
-            pinyins = g2pw.lazy_pinyin(seg, neutral_tone_with_five=True, style=Style.TONE3)
+            if seg:
+                pinyins = g2pw_batch_results[g2pw_batch_cursor]
+                g2pw_batch_cursor += 1
 
             pre_word_length = 0
             for word, pos in seg_cut:

GPT_SoVITS/text/g2pw/dataset.py

@@ -18,6 +18,7 @@ Credits
 
 from typing import Dict
 from typing import List
+from typing import Optional
 from typing import Tuple
 
 import numpy as np
@@ -37,6 +38,8 @@ def prepare_onnx_input(
     use_mask: bool = False,
     window_size: int = None,
     max_len: int = 512,
+    char2id: Optional[Dict[str, int]] = None,
+    char_phoneme_masks: Optional[Dict[str, List[int]]] = None,
 ) -> Dict[str, np.array]:
     if window_size is not None:
         truncated_texts, truncated_query_ids = _truncate_texts(
@@ -48,33 +51,88 @@ def prepare_onnx_input(
     phoneme_masks = []
     char_ids = []
     position_ids = []
+    tokenized_cache = {}
+
+    if char2id is None:
+        char2id = {char: idx for idx, char in enumerate(chars)}
+    if use_mask:
+        if char_phoneme_masks is None:
+            char_phoneme_masks = {
+                char: [1 if i in char2phonemes[char] else 0 for i in range(len(labels))]
+                for char in char2phonemes
+            }
+    else:
+        full_phoneme_mask = [1] * len(labels)
 
     for idx in range(len(texts)):
         text = (truncated_texts if window_size else texts)[idx].lower()
         query_id = (truncated_query_ids if window_size else query_ids)[idx]
 
-        try:
+        cached = tokenized_cache.get(text)
-            tokens, text2token, token2text = tokenize_and_map(tokenizer=tokenizer, text=text)
+        if cached is None:
-        except Exception:
+            try:
-            print(f'warning: text "{text}" is invalid')
+                tokens, text2token, token2text = tokenize_and_map(tokenizer=tokenizer, text=text)
-            return {}
+            except Exception:
+                print(f'warning: text "{text}" is invalid')
+                return {}
 
-        text, query_id, tokens, text2token, token2text = _truncate(
+            if len(tokens) <= max_len - 2:
-            max_len=max_len, text=text, query_id=query_id, tokens=tokens, text2token=text2token, token2text=token2text
+                processed_tokens = ["[CLS]"] + tokens + ["[SEP]"]
-        )
+                shared_input_id = list(np.array(tokenizer.convert_tokens_to_ids(processed_tokens)))
+                shared_token_type_id = list(np.zeros((len(processed_tokens),), dtype=int))
+                shared_attention_mask = list(np.ones((len(processed_tokens),), dtype=int))
+                cached = {
+                    "is_short": True,
+                    "tokens": tokens,
+                    "text2token": text2token,
+                    "token2text": token2text,
+                    "input_id": shared_input_id,
+                    "token_type_id": shared_token_type_id,
+                    "attention_mask": shared_attention_mask,
+                }
+            else:
+                cached = {
+                    "is_short": False,
+                    "tokens": tokens,
+                    "text2token": text2token,
+                    "token2text": token2text,
+                }
+            tokenized_cache[text] = cached
 
-        processed_tokens = ["[CLS]"] + tokens + ["[SEP]"]
+        if cached["is_short"]:
+            text_for_query = text
+            query_id_for_query = query_id
+            text2token_for_query = cached["text2token"]
+            input_id = cached["input_id"]
+            token_type_id = cached["token_type_id"]
+            attention_mask = cached["attention_mask"]
+        else:
+            (
+                text_for_query,
+                query_id_for_query,
+                tokens_for_query,
+                text2token_for_query,
+                _token2text_for_query,
+            ) = _truncate(
+                max_len=max_len,
+                text=text,
+                query_id=query_id,
+                tokens=cached["tokens"],
+                text2token=cached["text2token"],
+                token2text=cached["token2text"],
+            )
+            processed_tokens = ["[CLS]"] + tokens_for_query + ["[SEP]"]
+            input_id = list(np.array(tokenizer.convert_tokens_to_ids(processed_tokens)))
+            token_type_id = list(np.zeros((len(processed_tokens),), dtype=int))
+            attention_mask = list(np.ones((len(processed_tokens),), dtype=int))
 
-        input_id = list(np.array(tokenizer.convert_tokens_to_ids(processed_tokens)))
+        query_char = text_for_query[query_id_for_query]
-        token_type_id = list(np.zeros((len(processed_tokens),), dtype=int))
+        if use_mask:
-        attention_mask = list(np.ones((len(processed_tokens),), dtype=int))
+            phoneme_mask = char_phoneme_masks[query_char]
-
+        else:
-        query_char = text[query_id]
+            phoneme_mask = full_phoneme_mask
-        phoneme_mask = (
+        char_id = char2id[query_char]
-            [1 if i in char2phonemes[query_char] else 0 for i in range(len(labels))] if use_mask else [1] * len(labels)
+        position_id = text2token_for_query[query_id_for_query] + 1  # [CLS] token locate at first place
-        )
-        char_id = chars.index(query_char)
-        position_id = text2token[query_id] + 1  # [CLS] token locate at first place
 
         input_ids.append(input_id)
         token_type_ids.append(token_type_id)
@@ -83,10 +141,15 @@ def prepare_onnx_input(
         char_ids.append(char_id)
         position_ids.append(position_id)
 
+    max_token_length = max(len(seq) for seq in input_ids)
+
+    def _pad_sequences(sequences, pad_value=0):
+        return [seq + [pad_value] * (max_token_length - len(seq)) for seq in sequences]
+
     outputs = {
-        "input_ids": np.array(input_ids).astype(np.int64),
+        "input_ids": np.array(_pad_sequences(input_ids, pad_value=0)).astype(np.int64),
-        "token_type_ids": np.array(token_type_ids).astype(np.int64),
+        "token_type_ids": np.array(_pad_sequences(token_type_ids, pad_value=0)).astype(np.int64),
-        "attention_masks": np.array(attention_masks).astype(np.int64),
+        "attention_masks": np.array(_pad_sequences(attention_masks, pad_value=0)).astype(np.int64),
         "phoneme_masks": np.array(phoneme_masks).astype(np.float32),
         "char_ids": np.array(char_ids).astype(np.int64),
         "position_ids": np.array(position_ids).astype(np.int64),

GPT_SoVITS/text/g2pw/onnx_api.py

@@ -10,7 +10,6 @@ from typing import Any, Dict, List, Tuple
 import numpy as np
 import onnxruntime
 import requests
-import torch
 from opencc import OpenCC
 from pypinyin import Style, pinyin
 from transformers.models.auto.tokenization_auto import AutoTokenizer
@@ -22,9 +21,8 @@ from .utils import load_config
 onnxruntime.set_default_logger_severity(3)
 try:
     onnxruntime.preload_dlls()
-except:
+except Exception:
     pass
-    # traceback.print_exc()
 warnings.filterwarnings("ignore")
 
 model_version = "1.1"
@@ -55,6 +53,24 @@ def predict(session, onnx_input: Dict[str, Any], labels: List[str]) -> Tuple[Lis
     return all_preds, all_confidences
 
 
+def _load_json_from_candidates(filename: str, candidate_dirs: List[str]) -> Dict[str, Any]:
+    for candidate_dir in candidate_dirs:
+        if not candidate_dir:
+            continue
+        json_path = os.path.join(candidate_dir, filename)
+        if os.path.exists(json_path):
+            with open(json_path, "r", encoding="utf-8") as fr:
+                return json.load(fr)
+    raise FileNotFoundError(f"Cannot locate {filename} in candidate dirs: {candidate_dirs}")
+
+
+def _find_first_existing_file(*paths: str) -> str:
+    for path in paths:
+        if path and os.path.exists(path):
+            return path
+    raise FileNotFoundError(f"Files not found: {paths}")
+
+
 def download_and_decompress(model_dir: str = "G2PWModel/"):
     if not os.path.exists(model_dir):
         parent_directory = os.path.dirname(model_dir)
@@ -62,7 +78,7 @@ def download_and_decompress(model_dir: str = "G2PWModel/"):
         extract_dir = os.path.join(parent_directory, "G2PWModel_1.1")
         extract_dir_new = os.path.join(parent_directory, "G2PWModel")
         print("Downloading g2pw model...")
-        modelscope_url = "https://www.modelscope.cn/models/kamiorinn/g2pw/resolve/master/G2PWModel_1.1.zip"  # "https://paddlespeech.cdn.bcebos.com/Parakeet/released_models/g2p/G2PWModel_1.1.zip"
+        modelscope_url = "https://www.modelscope.cn/models/kamiorinn/g2pw/resolve/master/G2PWModel_1.1.zip"
         with requests.get(modelscope_url, stream=True) as r:
             r.raise_for_status()
             with open(zip_dir, "wb") as f:
@@ -79,7 +95,7 @@ def download_and_decompress(model_dir: str = "G2PWModel/"):
     return model_dir
 
 
-class G2PWOnnxConverter:
+class _G2PWBaseOnnxConverter:
     def __init__(
         self,
         model_dir: str = "G2PWModel/",
@@ -87,33 +103,16 @@ class G2PWOnnxConverter:
         model_source: str = None,
         enable_non_tradional_chinese: bool = False,
     ):
-        uncompress_path = download_and_decompress(model_dir)
+        self.model_dir = download_and_decompress(model_dir)
-
+        self.config = load_config(config_path=os.path.join(self.model_dir, "config.py"), use_default=True)
-        sess_options = onnxruntime.SessionOptions()
-        sess_options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
-        sess_options.execution_mode = onnxruntime.ExecutionMode.ORT_SEQUENTIAL
-        sess_options.intra_op_num_threads = 2 if torch.cuda.is_available() else 0
-        if "CUDAExecutionProvider" in onnxruntime.get_available_providers():
-            self.session_g2pW = onnxruntime.InferenceSession(
-                os.path.join(uncompress_path, "g2pW.onnx"),
-                sess_options=sess_options,
-                providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
-            )
-        else:
-            self.session_g2pW = onnxruntime.InferenceSession(
-                os.path.join(uncompress_path, "g2pW.onnx"),
-                sess_options=sess_options,
-                providers=["CPUExecutionProvider"],
-            )
-        self.config = load_config(config_path=os.path.join(uncompress_path, "config.py"), use_default=True)
 
         self.model_source = model_source if model_source else self.config.model_source
         self.enable_opencc = enable_non_tradional_chinese
-
         self.tokenizer = AutoTokenizer.from_pretrained(self.model_source)
 
-        polyphonic_chars_path = os.path.join(uncompress_path, "POLYPHONIC_CHARS.txt")
+        polyphonic_chars_path = os.path.join(self.model_dir, "POLYPHONIC_CHARS.txt")
-        monophonic_chars_path = os.path.join(uncompress_path, "MONOPHONIC_CHARS.txt")
+        monophonic_chars_path = os.path.join(self.model_dir, "MONOPHONIC_CHARS.txt")
+
         self.polyphonic_chars = [
             line.split("\t") for line in open(polyphonic_chars_path, encoding="utf-8").read().strip().split("\n")
         ]
@@ -149,31 +148,47 @@ class G2PWOnnxConverter:
         )
 
         self.chars = sorted(list(self.char2phonemes.keys()))
+        self.char2id = {char: idx for idx, char in enumerate(self.chars)}
+        self.char_phoneme_masks = (
+            {
+                char: [1 if i in self.char2phonemes[char] else 0 for i in range(len(self.labels))]
+                for char in self.char2phonemes
+            }
+            if self.config.use_mask
+            else None
+        )
 
         self.polyphonic_chars_new = set(self.chars)
         for char in self.non_polyphonic:
-            if char in self.polyphonic_chars_new:
+            self.polyphonic_chars_new.discard(char)
-                self.polyphonic_chars_new.remove(char)
 
         self.monophonic_chars_dict = {char: phoneme for char, phoneme in self.monophonic_chars}
         for char in self.non_monophonic:
-            if char in self.monophonic_chars_dict:
+            self.monophonic_chars_dict.pop(char, None)
-                self.monophonic_chars_dict.pop(char)
 
-        self.pos_tags = ["UNK", "A", "C", "D", "I", "N", "P", "T", "V", "DE", "SHI"]
+        default_asset_dir = os.path.normpath(os.path.join(os.path.dirname(__file__), "..", "G2PWModel"))
+        candidate_asset_dirs = [self.model_dir, default_asset_dir]
+        self.bopomofo_convert_dict = _load_json_from_candidates(
+            "bopomofo_to_pinyin_wo_tune_dict.json", candidate_asset_dirs
+        )
+        self.char_bopomofo_dict = _load_json_from_candidates("char_bopomofo_dict.json", candidate_asset_dirs)
 
-        with open(os.path.join(uncompress_path, "bopomofo_to_pinyin_wo_tune_dict.json"), "r", encoding="utf-8") as fr:
-            self.bopomofo_convert_dict = json.load(fr)
         self.style_convert_func = {
             "bopomofo": lambda x: x,
             "pinyin": self._convert_bopomofo_to_pinyin,
         }[style]
 
-        with open(os.path.join(uncompress_path, "char_bopomofo_dict.json"), "r", encoding="utf-8") as fr:
-            self.char_bopomofo_dict = json.load(fr)
-
         if self.enable_opencc:
             self.cc = OpenCC("s2tw")
+        self.enable_sentence_dedup = os.getenv("g2pw_sentence_dedup", "true").strip().lower() in {
+            "1",
+            "true",
+            "yes",
+            "y",
+            "on",
+        }
+        # 聚焦到多音字附近上下文，默认左右各16字；设为0表示关闭裁剪（整句）。
+        self.polyphonic_context_chars = max(0, int(os.getenv("g2pw_polyphonic_context_chars", "16")))
 
     def _convert_bopomofo_to_pinyin(self, bopomofo: str) -> str:
         tone = bopomofo[-1]
@@ -181,9 +196,8 @@ class G2PWOnnxConverter:
         component = self.bopomofo_convert_dict.get(bopomofo[:-1])
         if component:
             return component + tone
-        else:
+        print(f'Warning: "{bopomofo}" cannot convert to pinyin')
-            print(f'Warning: "{bopomofo}" cannot convert to pinyin')
+        return None
-            return None
 
     def __call__(self, sentences: List[str]) -> List[List[str]]:
         if isinstance(sentences, str):
@@ -197,51 +211,147 @@ class G2PWOnnxConverter:
                 translated_sentences.append(translated_sent)
             sentences = translated_sentences
 
-        texts, query_ids, sent_ids, partial_results = self._prepare_data(sentences=sentences)
+        texts, model_query_ids, result_query_ids, sent_ids, partial_results = self._prepare_data(sentences=sentences)
         if len(texts) == 0:
-            # sentences no polyphonic words
             return partial_results
 
-        onnx_input = prepare_onnx_input(
+        model_input = prepare_onnx_input(
             tokenizer=self.tokenizer,
             labels=self.labels,
             char2phonemes=self.char2phonemes,
             chars=self.chars,
             texts=texts,
-            query_ids=query_ids,
+            query_ids=model_query_ids,
             use_mask=self.config.use_mask,
             window_size=None,
+            char2id=self.char2id,
+            char_phoneme_masks=self.char_phoneme_masks,
         )
 
-        preds, confidences = predict(session=self.session_g2pW, onnx_input=onnx_input, labels=self.labels)
+        if not model_input:
+            return partial_results
+
+        if self.enable_sentence_dedup:
+            preds, _confidences = self._predict_with_sentence_dedup(model_input=model_input, texts=texts)
+        else:
+            preds, _confidences = self._predict(model_input=model_input)
+
         if self.config.use_char_phoneme:
             preds = [pred.split(" ")[1] for pred in preds]
 
         results = partial_results
-        for sent_id, query_id, pred in zip(sent_ids, query_ids, preds):
+        for sent_id, query_id, pred in zip(sent_ids, result_query_ids, preds):
             results[sent_id][query_id] = self.style_convert_func(pred)
 
         return results
 
-    def _prepare_data(self, sentences: List[str]) -> Tuple[List[str], List[int], List[int], List[List[str]]]:
+    def _prepare_data(
-        texts, query_ids, sent_ids, partial_results = [], [], [], []
+        self, sentences: List[str]
+    ) -> Tuple[List[str], List[int], List[int], List[int], List[List[str]]]:
+        texts, model_query_ids, result_query_ids, sent_ids, partial_results = [], [], [], [], []
         for sent_id, sent in enumerate(sentences):
-            # pypinyin works well for Simplified Chinese than Traditional Chinese
             sent_s = tranditional_to_simplified(sent)
             pypinyin_result = pinyin(sent_s, neutral_tone_with_five=True, style=Style.TONE3)
             partial_result = [None] * len(sent)
+            polyphonic_indices: List[int] = []
             for i, char in enumerate(sent):
                 if char in self.polyphonic_chars_new:
-                    texts.append(sent)
+                    polyphonic_indices.append(i)
-                    query_ids.append(i)
-                    sent_ids.append(sent_id)
                 elif char in self.monophonic_chars_dict:
                     partial_result[i] = self.style_convert_func(self.monophonic_chars_dict[char])
                 elif char in self.char_bopomofo_dict:
                     partial_result[i] = pypinyin_result[i][0]
-                    # partial_result[i] =  self.style_convert_func(self.char_bopomofo_dict[char][0])
                 else:
                     partial_result[i] = pypinyin_result[i][0]
 
+            if polyphonic_indices:
+                if self.polyphonic_context_chars > 0:
+                    left = max(0, polyphonic_indices[0] - self.polyphonic_context_chars)
+                    right = min(len(sent), polyphonic_indices[-1] + self.polyphonic_context_chars + 1)
+                    sent_for_predict = sent[left:right]
+                    query_offset = left
+                else:
+                    sent_for_predict = sent
+                    query_offset = 0
+
+                for index in polyphonic_indices:
+                    texts.append(sent_for_predict)
+                    model_query_ids.append(index - query_offset)
+                    result_query_ids.append(index)
+                    sent_ids.append(sent_id)
+
             partial_results.append(partial_result)
-        return texts, query_ids, sent_ids, partial_results
+        return texts, model_query_ids, result_query_ids, sent_ids, partial_results
+
+    def _predict(self, model_input: Dict[str, Any]) -> Tuple[List[str], List[float]]:
+        raise NotImplementedError
+
+    def _predict_with_sentence_dedup(
+        self, model_input: Dict[str, Any], texts: List[str]
+    ) -> Tuple[List[str], List[float]]:
+        if len(texts) <= 1:
+            return self._predict(model_input=model_input)
+
+        grouped_indices: Dict[str, List[int]] = {}
+        for idx, text in enumerate(texts):
+            grouped_indices.setdefault(text, []).append(idx)
+
+        if all(len(indices) == 1 for indices in grouped_indices.values()):
+            return self._predict(model_input=model_input)
+
+        preds: List[str] = [""] * len(texts)
+        confidences: List[float] = [0.0] * len(texts)
+        for indices in grouped_indices.values():
+            group_input = {name: value[indices] for name, value in model_input.items()}
+            if len(indices) > 1:
+                for name in ("input_ids", "token_type_ids", "attention_masks"):
+                    group_input[name] = group_input[name][:1]
+
+            group_preds, group_confidences = self._predict(model_input=group_input)
+            for output_idx, pred, confidence in zip(indices, group_preds, group_confidences):
+                preds[output_idx] = pred
+                confidences[output_idx] = confidence
+
+        return preds, confidences
+
+
+class G2PWOnnxConverter(_G2PWBaseOnnxConverter):
+    def __init__(
+        self,
+        model_dir: str = "G2PWModel/",
+        style: str = "bopomofo",
+        model_source: str = None,
+        enable_non_tradional_chinese: bool = False,
+    ):
+        super().__init__(
+            model_dir=model_dir,
+            style=style,
+            model_source=model_source,
+            enable_non_tradional_chinese=enable_non_tradional_chinese,
+        )
+
+        sess_options = onnxruntime.SessionOptions()
+        sess_options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
+        sess_options.execution_mode = onnxruntime.ExecutionMode.ORT_SEQUENTIAL
+        sess_options.intra_op_num_threads = 2
+
+        onnx_path = _find_first_existing_file(
+            os.path.join(self.model_dir, "g2pW.onnx"),
+            os.path.join(self.model_dir, "g2pw.onnx"),
+        )
+
+        if "CUDAExecutionProvider" in onnxruntime.get_available_providers():
+            self.session_g2pw = onnxruntime.InferenceSession(
+                onnx_path,
+                sess_options=sess_options,
+                providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
+            )
+        else:
+            self.session_g2pw = onnxruntime.InferenceSession(
+                onnx_path,
+                sess_options=sess_options,
+                providers=["CPUExecutionProvider"],
+            )
+
+    def _predict(self, model_input: Dict[str, Any]) -> Tuple[List[str], List[float]]:
+        return predict(session=self.session_g2pw, onnx_input=model_input, labels=self.labels)

README.md

@@ -48,6 +48,8 @@ https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-
 
 请不要尬黑GPT-SoVITS推理速度慢，谢谢！
 
+CPU-Optimized Inference Version：https://github.com/baicai-1145/GPT-SoVITS-CPUFast
+
 **User guide: [简体中文](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e) | [English](https://rentry.co/GPT-SoVITS-guide#/)**
 
 ## Installation

api_role.py

@@ -0,0 +1,825 @@
+"""
+GPT-SoVITS API 实现
+
+### 完整请求示例 (/ttsrole POST)
+{
+    "text": "你好",                     # str, 必填, 要合成的文本内容
+    "role": "role1",                   # str, 必填, 角色名称，决定使用 roles/{role} 中的配置和音频
+    "emotion": "开心",                  # str, 可选, 情感标签，用于从 roles/{role}/reference_audios 中选择音频
+    "text_lang": "auto",               # str, 可选, 默认 "auto", 文本语言，"auto" 时根据 emotion 或角色目录动态选择
+    "ref_audio_path": "/path/to/ref.wav",  # str, 可选, 参考音频路径，若提供则优先使用，跳过自动选择
+    "aux_ref_audio_paths": ["/path1.wav", "/path2.wav"],  # List[str], 可选, 辅助参考音频路径，用于多说话人融合
+    "prompt_lang": "ja",               # str, 可选, 提示文本语言，若提供 ref_audio_path 则需指定，"auto" 模式下动态选择
+    "prompt_text": "こんにちは",       # str, 可选, 提示文本，与 ref_audio_path 配对使用，自动选择时从文件或文件名生成
+    "top_k": 10,                       # int, 可选, Top-K 采样值，覆盖 inference.top_k
+    "top_p": 0.8,                      # float, 可选, Top-P 采样值，覆盖 inference.top_p
+    "temperature": 1.0,                # float, 可选, 温度值，覆盖 inference.temperature
+    "text_split_method": "cut5",       # str, 可选, 文本分割方法，覆盖 inference.text_split_method, 具体见text_segmentation_method.py
+    "batch_size": 2,                   # int, 可选, 批处理大小，覆盖 inference.batch_size
+    "batch_threshold": 0.75,           # float, 可选, 批处理阈值，覆盖 inference.batch_threshold
+    "split_bucket": true,              # bool, 可选, 是否按桶分割，覆盖 inference.split_bucket
+    "speed_factor": 1.2,               # float, 可选, 语速因子，覆盖 inference.speed_factor
+    "fragment_interval": 0.3,          # float, 可选, 片段间隔（秒），覆盖 inference.fragment_interval
+    "seed": 42,                        # int, 可选, 随机种子，覆盖 seed
+    "media_type": "wav",               # str, 可选, 默认 "wav", 输出格式，支持 "wav", "raw", "ogg", "aac"
+    "streaming_mode": false,           # bool, 可选, 默认 false, 是否流式返回
+    "parallel_infer": true,            # bool, 可选, 默认 true, 是否并行推理
+    "repetition_penalty": 1.35,        # float, 可选, 重复惩罚值，覆盖 inference.repetition_penalty
+    "version": "v2",                   # str, 可选, 配置文件版本，覆盖 version
+    "languages": ["zh", "ja", "en"],   # List[str], 可选, 支持的语言列表，覆盖 languages
+    "bert_base_path": "/path/to/bert", # str, 可选, BERT 模型路径，覆盖 bert_base_path
+    "cnhuhbert_base_path": "/path/to/hubert",  # str, 可选, HuBERT 模型路径，覆盖 cnhuhbert_base_path
+    "device": "cpu",                   # str, 可选, 统一设备，覆盖 device
+    "is_half": true,                   # bool, 可选, 是否使用半精度，覆盖 is_half
+    "t2s_weights_path": "/path/to/gpt.ckpt",  # str, 可选, GPT 模型路径，覆盖 t2s_weights_path
+    "vits_weights_path": "/path/to/sovits.pth",  # str, 可选, SoVITS 模型路径，覆盖 vits_weights_path
+    "t2s_model_path": "/path/to/gpt.ckpt",  # str, 可选, GPT 模型路径（与 t2s_weights_path 同义）
+    "t2s_model_device": "cpu",         # str, 可选, GPT 模型设备，覆盖 t2s_model.device，默认检测显卡
+    "vits_model_path": "/path/to/sovits.pth",  # str, 可选, SoVITS 模型路径（与 vits_weights_path 同义）
+    "vits_model_device": "cpu"         # str, 可选, SoVITS 模型设备，覆盖 vits_model.device，默认检测显卡
+}
+
+### 参数必要性和优先级
+- 必填参数:
+  - /ttsrole: text, role
+  - /tts: text, ref_audio_path, prompt_lang
+- 可选参数: 其他均为可选，默认值从 roles/{role}/tts_infer.yaml 或 GPT_SoVITS/configs/tts_infer.yaml 获取
+- 优先级: POST 请求参数 > roles/{role}/tts_infer.yaml > 默认 GPT_SoVITS/configs/tts_infer.yaml
+
+### 目录结构
+GPT-SoVITS-roleapi/
+├── api_role.py                    # 本文件, API 主程序
+├── GPT_SoVITS/                    # GPT-SoVITS 核心库
+│   └── configs/
+│       └── tts_infer.yaml         # 默认配置文件
+├── roles/                         # 角色配置目录
+│   ├── role1/                     # 示例角色 role1
+│   │   ├── tts_infer.yaml         # 角色配置文件（可选）
+│   │   ├── model.ckpt             # GPT 模型（可选）
+│   │   ├── model.pth              # SoVITS 模型（可选）
+│   │   └── reference_audios/      # 角色参考音频目录
+│   │       ├── zh/
+│   │       │   ├── 【开心】voice1.wav
+│   │       │   ├── 【开心】voice1.txt
+│   │       ├── ja/
+│   │       │   ├── 【开心】voice2.wav
+│   │       │   ├── 【开心】voice2.txt
+│   ├── role2/
+│   │   ├── tts_infer.yaml
+│   │   ├── model.ckpt
+│   │   ├── model.pth
+│   │   └── reference_audios/
+│   │       ├── zh/
+│   │       │   ├── 【开心】voice1.wav
+│   │       │   ├── 【开心】voice1.txt
+│   │       │   ├── 【悲伤】asdafasdas.wav
+│   │       │   ├── 【悲伤】asdafasdas.txt
+│   │       ├── ja/
+│   │       │   ├── 【开心】voice2.wav
+│   │       │   ├── 【开心】voice2.txt
+
+### text_lang, prompt_lang, prompt_text 选择逻辑 (/ttsrole)
+1. text_lang 选择逻辑:
+   - 默认值: "auto"
+   - 如果请求未提供 text_lang，视为 "auto"
+   - 当 text_lang = "auto" 且存在 emotion 参数：
+     - 从 roles/{role}/reference_audios 下所有语言文件夹中查找以 "【emotion】" 开头的音频
+     - 随机选择一个匹配的音频，语言由音频所在文件夹确定
+   - 当 text_lang 指定具体语言（如 "zh"）：
+     - 从 roles/{role}/reference_audios/{text_lang} 中选择音频
+     - 如果指定语言无匹配音频，则尝试其他语言文件夹
+2. prompt_lang 选择逻辑:
+   - 如果提供了 ref_audio_path，则需显式指定 prompt_lang
+   - 如果未提供 ref_audio_path 且 text_lang = "auto" 且存在 emotion：
+     - prompt_lang = 随机选择的音频所在语言文件夹名（如 "zh" 或 "ja"）
+   - 如果未提供 ref_audio_path 且 text_lang 指定具体语言：
+     - prompt_lang = text_lang（如 "zh"）
+     - 如果 text_lang 无匹配音频，则为随机选择的音频所在语言
+3. prompt_text 选择逻辑:
+   - 如果提供了 ref_audio_path（如 "/path/to/ref.wav"）：
+     - 检查文件名是否包含 "【xxx】" 前缀：
+       - 如果有（如 "【开心】abc.wav"）：
+         - 若存在对应 .txt 文件（如 "【开心】abc.txt"），prompt_text = .txt 文件内容
+         - 若无对应 .txt 文件，prompt_text = "abc"（去掉 "【开心】" 和 ".wav" 的部分）
+       - 如果无 "【xxx】" 前缀：
+         - 若存在对应 .txt 文件（如 "ref.txt"），prompt_text = .txt 文件内容
+         - 若无对应 .txt 文件，prompt_text = "ref"（去掉 ".wav" 的部分）
+   - 如果未提供 ref_audio_path：
+     - 从 roles/{role}/reference_audios 中选择音频（基于 text_lang 和 emotion）：
+       - 优先匹配 "【emotion】" 前缀的音频（如 "【开心】voice1.wav"）
+       - 若存在对应 .txt 文件（如 "【开心】voice1.txt"），prompt_text = .txt 文件内容
+       - 若无对应 .txt 文件，prompt_text = "voice1"（去掉 "【开心】" 和 ".wav" 的部分）
+       - 未匹配 emotion 则随机选择一个音频，逻辑同上
+
+### 讲解
+1. 必填参数:
+   - /ttsrole: text, role
+   - /tts: text, ref_audio_path, prompt_lang
+2. 音频选择 (/ttsrole):
+   - 若提供 ref_audio_path，则使用它
+   - 否则根据 role、text_lang、emotion 从 roles/{role}/reference_audios 中选择
+   - text_lang = "auto" 时，若有 emotion，则跨语言匹配 "【emotion】" 前缀音频
+   - emotion 匹配 "【emotion】" 前缀音频，未匹配则随机选择
+3. 设备选择:
+   - 默认尝试检测显卡（torch.cuda.is_available()），若可用则用 "cuda"，否则 "cpu"
+   - 若缺少 torch 依赖或检测失败，回退到 "cpu"
+   - POST 参数 device, t2s_model_device, vits_model_device 可强制指定设备，优先级最高
+4. 配置文件:
+   - 默认加载 GPT_SoVITS/configs/tts_infer.yaml
+   - 若 roles/{role}/tts_infer.yaml 存在且未被请求参数覆盖，则使用它 (/ttsrole)
+   - 请求参数（如 top_k, bert_base_path）覆盖所有配置文件
+5. 返回格式:
+   - 成功时返回音频流 (Response 或 StreamingResponse)
+   - 失败时返回 JSON，包含错误消息和可能的异常详情
+6. 运行:
+   - python api_role.py -a 127.0.0.1 -p 9880
+   - 检查启动日志确认设备
+   
+### 调用示例 (/ttsrole)
+## 非流式调用，会一次性返回完整的音频数据，适用于需要完整音频文件的场景
+import requests
+url = "http://127.0.0.1:9880/ttsrole"
+payload = {
+    "text": "你好，这是一个测试",  # 要合成的文本
+    "role": "role1",               # 角色名称，必填
+    "emotion": "开心",              # 情感标签，可选
+    "text_lang": "zh",             # 文本语言，可选，默认为 "zh"
+    "media_type": "wav"            # 输出音频格式，默认 "wav"
+}
+response = requests.post(url, json=payload)
+if response.status_code == 200:
+    with open("output_non_stream.wav", "wb") as f:
+        f.write(response.content)
+    print("非流式音频已生成并保存为 output_non_stream.wav")
+else:
+    print(f"请求失败: {response.json()}")
+
+## 流式调用，会分块返回音频数据，适用于实时播放或处理大文件的场景
+import requests
+url = "http://127.0.0.1:9880/ttsrole"
+payload = {
+    "text": "你好，这是一个测试",  # 要合成的文本
+    "role": "role1",               # 角色名称，必填
+    "emotion": "开心",              # 情感标签，可选
+    "text_lang": "zh",             # 文本语言，可选，默认为 "zh"
+    "media_type": "wav",           # 输出音频格式，默认 "wav"
+    "streaming_mode": True         # 启用流式模式
+}
+with requests.post(url, json=payload, stream=True) as response:
+    if response.status_code == 200:
+        with open("output_stream.wav", "wb") as f:
+            for chunk in response.iter_content(chunk_size=1024):
+                if chunk:  # 确保 chunk 不为空
+                    f.write(chunk)
+        print("流式音频已生成并保存为 output_stream.wav")
+    else:
+        print(f"请求失败: {response.json()}")
+"""
+
+import os
+import sys
+import traceback
+from typing import Generator, Optional, List, Dict
+import random
+import glob
+from concurrent.futures import ThreadPoolExecutor
+import asyncio
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+sys.path.append("%s/GPT_SoVITS" % (now_dir))
+
+import argparse
+import subprocess
+import wave
+import signal
+import numpy as np
+import soundfile as sf
+from fastapi import FastAPI, HTTPException, Response
+from fastapi.responses import StreamingResponse, JSONResponse
+from pydantic import BaseModel
+import uvicorn
+from io import BytesIO
+from tools.i18n.i18n import I18nAuto
+from GPT_SoVITS.TTS_infer_pack.TTS import TTS, TTS_Config
+from GPT_SoVITS.TTS_infer_pack.text_segmentation_method import get_method_names as get_cut_method_names
+
+# 尝试导入 PyTorch，检测显卡支持
+try:
+    import torch
+    cuda_available = torch.cuda.is_available()
+except ImportError:
+    cuda_available = False
+    print("缺少 PyTorch 依赖，默认使用 CPU")
+except Exception as e:
+    cuda_available = False
+    print(f"检测显卡时出错: {str(e)}，默认使用 CPU")
+
+i18n = I18nAuto()
+cut_method_names = get_cut_method_names()
+
+parser = argparse.ArgumentParser(description="GPT-SoVITS api")
+parser.add_argument("-c", "--tts_config", type=str, default="GPT_SoVITS/configs/tts_infer.yaml", help="tts_infer路径")
+parser.add_argument("-a", "--bind_addr", type=str, default="127.0.0.1", help="default: 127.0.0.1")
+parser.add_argument("-p", "--port", type=int, default="9880", help="default: 9880")
+args = parser.parse_args()
+config_path = args.tts_config
+port = args.port
+host = args.bind_addr
+argv = sys.argv
+
+if config_path in [None, ""]:
+    config_path = "GPT_SoVITS/configs/tts_infer.yaml"
+
+default_device = "cuda" if cuda_available else "cpu"
+print(f"默认设备设置为: {default_device}")
+
+# 初始化 TTS 配置
+tts_config = TTS_Config(config_path)
+print(f"TTS_Config contents: {tts_config.__dict__}")
+if hasattr(tts_config, 'device'):
+    tts_config.device = default_device
+tts_pipeline = TTS(tts_config)
+
+# 创建线程池用于异步执行 TTS 任务
+executor = ThreadPoolExecutor(max_workers=1)
+
+APP = FastAPI()
+
+class TTS_Request(BaseModel):
+    text: str
+    ref_audio_path: str
+    prompt_lang: str
+    text_lang: str = "auto"
+    aux_ref_audio_paths: Optional[List[str]] = None
+    prompt_text: Optional[str] = ""
+    top_k: Optional[int] = 5
+    top_p: Optional[float] = 1
+    temperature: Optional[float] = 1
+    text_split_method: Optional[str] = "cut5"
+    batch_size: Optional[int] = 1
+    batch_threshold: Optional[float] = 0.75
+    split_bucket: Optional[bool] = True
+    speed_factor: Optional[float] = 1.0
+    fragment_interval: Optional[float] = 0.3
+    seed: Optional[int] = -1
+    media_type: Optional[str] = "wav"
+    streaming_mode: Optional[bool] = False
+    parallel_infer: Optional[bool] = True
+    repetition_penalty: Optional[float] = 1.35
+    device: Optional[str] = None
+
+class TTSRole_Request(BaseModel):
+    text: str
+    role: str
+    text_lang: Optional[str] = "auto"
+    ref_audio_path: Optional[str] = None
+    aux_ref_audio_paths: Optional[List[str]] = None
+    prompt_lang: Optional[str] = None
+    prompt_text: Optional[str] = None
+    emotion: Optional[str] = None
+    top_k: Optional[int] = 5
+    top_p: Optional[float] = 1
+    temperature: Optional[float] = 1
+    text_split_method: Optional[str] = "cut5"
+    batch_size: Optional[int] = 1
+    batch_threshold: Optional[float] = 0.75
+    split_bucket: Optional[bool] = True
+    speed_factor: Optional[float] = 1.0
+    fragment_interval: Optional[float] = 0.3
+    seed: Optional[int] = -1
+    media_type: Optional[str] = "wav"
+    streaming_mode: Optional[bool] = False
+    parallel_infer: Optional[bool] = True
+    repetition_penalty: Optional[float] = 1.35
+    bert_base_path: Optional[str] = None
+    cnhuhbert_base_path: Optional[str] = None
+    device: Optional[str] = None
+    is_half: Optional[bool] = None
+    t2s_weights_path: Optional[str] = None
+    version: Optional[str] = None
+    vits_weights_path: Optional[str] = None
+    t2s_model_path: Optional[str] = None
+    vits_model_path: Optional[str] = None
+    t2s_model_device: Optional[str] = None
+    vits_model_device: Optional[str] = None
+
+def pack_ogg(io_buffer: BytesIO, data: np.ndarray, rate: int):
+    with sf.SoundFile(io_buffer, mode='w', samplerate=rate, channels=1, format='ogg') as audio_file:
+        audio_file.write(data)
+    io_buffer.seek(0)
+    return io_buffer
+
+def pack_raw(io_buffer: BytesIO, data: np.ndarray, rate: int):
+    io_buffer.write(data.tobytes())
+    io_buffer.seek(0)
+    return io_buffer
+
+def pack_wav(io_buffer: BytesIO, data: np.ndarray, rate: int):
+    sf.write(io_buffer, data, rate, format='wav')
+    io_buffer.seek(0)
+    return io_buffer
+
+def pack_aac(io_buffer: BytesIO, data: np.ndarray, rate: int):
+    process = subprocess.Popen([
+        'ffmpeg', '-f', 's16le', '-ar', str(rate), '-ac', '1', '-i', 'pipe:0',
+        '-c:a', 'aac', '-b:a', '192k', '-vn', '-f', 'adts', 'pipe:1'
+    ], stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+    out, _ = process.communicate(input=data.tobytes())
+    io_buffer.write(out)
+    io_buffer.seek(0)
+    return io_buffer
+
+def pack_audio(data: np.ndarray, rate: int, media_type: str) -> BytesIO:
+    io_buffer = BytesIO()
+    if media_type == "ogg":
+        io_buffer = pack_ogg(io_buffer, data, rate)
+    elif media_type == "aac":
+        io_buffer = pack_aac(io_buffer, data, rate)
+    elif media_type == "wav":
+        io_buffer = pack_wav(io_buffer, data, rate)
+    else:
+        io_buffer = pack_raw(io_buffer, data, rate)
+    return io_buffer
+
+def wave_header_chunk(frame_input=b"", channels=1, sample_width=2, sample_rate=32000):
+    wav_buf = BytesIO()
+    with wave.open(wav_buf, "wb") as vfout:
+        vfout.setnchannels(channels)
+        vfout.setsampwidth(sample_width)
+        vfout.setframerate(sample_rate)
+        vfout.writeframes(frame_input)
+    wav_buf.seek(0)
+    return wav_buf.read()
+
+def handle_control(command: str):
+    if command == "restart":
+        os.execl(sys.executable, sys.executable, *argv)
+    elif command == "exit":
+        os.kill(os.getpid(), signal.SIGTERM)
+        exit(0)
+
+def check_params(req: dict, is_ttsrole: bool = False):
+    text = req.get("text")
+    text_lang = req.get("text_lang", "auto")
+    ref_audio_path = req.get("ref_audio_path")
+    prompt_lang = req.get("prompt_lang")
+    media_type = req.get("media_type", "wav")
+    streaming_mode = req.get("streaming_mode", False)
+    text_split_method = req.get("text_split_method", "cut5")
+
+    if not text:
+        return {"status": "error", "message": "text is required"}
+    
+    if is_ttsrole:
+        role = req.get("role")
+        if not role:
+            return {"status": "error", "message": "role is required for /ttsrole"}
+    else:
+        if not ref_audio_path:
+            return {"status": "error", "message": "ref_audio_path is required"}
+        if not prompt_lang:
+            return {"status": "error", "message": "prompt_lang is required"}
+    
+    languages = req.get("languages") or tts_config.languages
+    if text_lang != "auto" and text_lang.lower() not in languages:
+        return {"status": "error", "message": f"text_lang: {text_lang} is not supported"}
+    if prompt_lang and prompt_lang.lower() not in languages:
+        return {"status": "error", "message": f"prompt_lang: {prompt_lang} is not supported"}
+    
+    if media_type not in ["wav", "raw", "ogg", "aac"]:
+        return {"status": "error", "message": f"media_type: {media_type} is not supported"}
+    if media_type == "ogg" and not streaming_mode:
+        return {"status": "error", "message": "ogg format is not supported in non-streaming mode"}
+    if text_split_method not in cut_method_names:
+        return {"status": "error", "message": f"text_split_method: {text_split_method} is not supported"}
+    
+    return None
+
+def load_role_config(role: str, req: dict):
+    role_dir = os.path.join(now_dir, "roles", role)
+    if not os.path.exists(role_dir):
+        return False
+    
+    if not any(req.get(k) for k in ["version", "bert_base_path", "cnhuhbert_base_path", "device", "is_half", "t2s_weights_path", "vits_weights_path"]):
+        config_path_new = os.path.join(role_dir, "tts_infer.yaml")
+        if os.path.exists(config_path_new):
+            global tts_config, tts_pipeline
+            tts_config = TTS_Config(config_path_new)
+            if hasattr(tts_config, 'device'):
+                tts_config.device = default_device
+            tts_pipeline = TTS(tts_config)
+    
+    if not req.get("t2s_weights_path") and not req.get("t2s_model_path"):
+        gpt_path = glob.glob(os.path.join(role_dir, "*.ckpt"))
+        if gpt_path:
+            tts_pipeline.init_t2s_weights(gpt_path[0])
+    if not req.get("vits_weights_path") and not req.get("vits_model_path"):
+        sovits_path = glob.glob(os.path.join(role_dir, "*.pth"))
+        if sovits_path:
+            tts_pipeline.init_vits_weights(sovits_path[0])
+    
+    return True
+
+def select_ref_audio(role: str, text_lang: str, emotion: str = None):
+    audio_base_dir = os.path.join(now_dir, "roles", role, "reference_audios")
+    if not os.path.exists(audio_base_dir):
+        return None, None, None
+    
+    if text_lang.lower() == "auto" and emotion:
+        all_langs = [d for d in os.listdir(audio_base_dir) if os.path.isdir(os.path.join(audio_base_dir, d))]
+        emotion_files = []
+        for lang in all_langs:
+            lang_dir = os.path.join(audio_base_dir, lang)
+            emotion_files.extend(glob.glob(os.path.join(lang_dir, f"【{emotion}】*.*")))
+        
+        if emotion_files:
+            audio_path = random.choice(emotion_files)
+            txt_path = audio_path.rsplit(".", 1)[0] + ".txt"
+            if os.path.exists(txt_path):
+                with open(txt_path, "r", encoding="utf-8") as f:
+                    prompt_text = f.read().strip()
+            else:
+                basename = os.path.basename(audio_path)
+                start_idx = basename.find("】") + 1
+                end_idx = basename.rfind(".")
+                prompt_text = basename[start_idx:end_idx] if end_idx > start_idx else basename
+            
+            prompt_lang = os.path.basename(os.path.dirname(audio_path))
+            return audio_path, prompt_text, prompt_lang
+    
+    lang_dir = os.path.join(audio_base_dir, text_lang.lower())
+    all_langs = [d for d in os.listdir(audio_base_dir) if os.path.isdir(os.path.join(audio_base_dir, d))]
+    
+    def find_audio_in_dir(dir_path):
+        if not os.path.exists(dir_path):
+            return None, None
+        audio_files = glob.glob(os.path.join(dir_path, "【*】*.*"))
+        if not audio_files:
+            audio_files = glob.glob(os.path.join(dir_path, "*.*"))
+        if not audio_files:
+            return None, None
+        
+        if emotion:
+            emotion_files = [f for f in audio_files if f"【{emotion}】" in os.path.basename(f)]
+            if emotion_files:
+                audio_path = random.choice(emotion_files)
+            else:
+                audio_path = random.choice(audio_files)
+        else:
+            audio_path = random.choice(audio_files)
+        
+        txt_path = audio_path.rsplit(".", 1)[0] + ".txt"
+        prompt_text = None
+        if os.path.exists(txt_path):
+            with open(txt_path, "r", encoding="utf-8") as f:
+                prompt_text = f.read().strip()
+        else:
+            basename = os.path.basename(audio_path)
+            start_idx = basename.find("】") + 1
+            end_idx = basename.rfind(".")
+            if start_idx > 0 and end_idx > start_idx:
+                prompt_text = basename[start_idx:end_idx]
+            else:
+                prompt_text = basename[:end_idx] if end_idx > 0 else basename
+        
+        return audio_path, prompt_text
+    
+    audio_path, prompt_text = find_audio_in_dir(lang_dir)
+    if audio_path:
+        return audio_path, prompt_text, text_lang.lower()
+    
+    for lang in all_langs:
+        if lang != text_lang.lower():
+            audio_path, prompt_text = find_audio_in_dir(os.path.join(audio_base_dir, lang))
+            if audio_path:
+                return audio_path, prompt_text, lang
+    
+    return None, None, None
+
+def set_pipeline_device(pipeline: TTS, device: str):
+    """将 TTS 管道中的所有模型和相关组件迁移到指定设备，仅在设备变化时执行"""
+    if not torch.cuda.is_available() and device.startswith("cuda"):
+        print(f"警告: CUDA 不可用，强制使用 CPU")
+        device = "cpu"
+    
+    target_device = torch.device(device)
+    
+    # 检查当前设备是否需要切换
+    current_device = None
+    if hasattr(pipeline, 't2s_model') and pipeline.t2s_model is not None:
+        current_device = next(pipeline.t2s_model.parameters()).device
+    elif hasattr(pipeline, 'vits_model') and pipeline.vits_model is not None:
+        current_device = next(pipeline.vits_model.parameters()).device
+    
+    if current_device == target_device:
+        print(f"设备已是 {device}，无需切换")
+        return
+    
+    # 更新配置中的设备
+    if hasattr(pipeline, 'configs') and hasattr(pipeline.configs, 'device'):
+        pipeline.configs.device = device
+    
+    # 迁移所有可能的模型到指定设备
+    for attr in ['t2s_model', 'vits_model']:
+        if hasattr(pipeline, attr) and getattr(pipeline, attr) is not None:
+            getattr(pipeline, attr).to(target_device)
+    
+    for attr in dir(pipeline):
+        if attr.endswith('_model') and getattr(pipeline, attr) is not None:
+            try:
+                getattr(pipeline, attr).to(target_device)
+                print(f"迁移 {attr} 到 {device}")
+            except AttributeError:
+                pass
+    
+    # 清理 GPU 缓存
+    if torch.cuda.is_available() and not device.startswith("cuda"):
+        torch.cuda.empty_cache()
+    
+    print(f"TTS 管道设备已设置为: {device}")
+
+def run_tts_pipeline(req):
+    """在线程池中运行 TTS 任务"""
+    return tts_pipeline.run(req)
+
+async def tts_handle(req: dict, is_ttsrole: bool = False):
+    streaming_mode = req.get("streaming_mode", False)
+    media_type = req.get("media_type", "wav")
+
+    if "text_lang" not in req:
+        req["text_lang"] = "auto"
+
+    check_res = check_params(req, is_ttsrole)
+    if check_res is not None:
+        return JSONResponse(status_code=400, content=check_res)
+    
+    # 如果请求中指定了 device，则覆盖所有与设备相关的参数并更新管道设备
+    if "device" in req and req["device"] is not None:
+        device = req["device"]
+        req["t2s_model_device"] = device
+        req["vits_model_device"] = device
+        if hasattr(tts_config, 'device'):
+            tts_config.device = device
+        set_pipeline_device(tts_pipeline, device)
+    
+    if is_ttsrole:
+        role_exists = load_role_config(req["role"], req)
+        
+        for key in ["bert_base_path", "cnhuhbert_base_path", "device", "is_half", "t2s_weights_path", "version", "vits_weights_path"]:
+            if req.get(key) is not None:
+                setattr(tts_config, key, req[key])
+        
+        if req.get("t2s_model_path"):
+            tts_config.t2s_weights_path = req["t2s_model_path"]
+            tts_pipeline.init_t2s_weights(req["t2s_model_path"])
+        if req.get("vits_model_path"):
+            tts_config.vits_weights_path = req["vits_model_path"]
+            tts_pipeline.init_vits_weights(req["vits_model_path"])
+        
+        if not req.get("ref_audio_path"):
+            ref_audio_path, prompt_text, prompt_lang = select_ref_audio(req["role"], req["text_lang"], req.get("emotion"))
+            if ref_audio_path:
+                req["ref_audio_path"] = ref_audio_path
+                req["prompt_text"] = prompt_text or ""
+                req["prompt_lang"] = prompt_lang or req["text_lang"]
+            elif not role_exists:
+                return JSONResponse(status_code=400, content={"status": "error", "message": "Role directory not found and no suitable reference audio provided"})
+        else:
+            ref_audio_path = req["ref_audio_path"]
+            txt_path = ref_audio_path.rsplit(".", 1)[0] + ".txt"
+            if os.path.exists(txt_path):
+                with open(txt_path, "r", encoding="utf-8") as f:
+                    req["prompt_text"] = f.read().strip()
+            else:
+                basename = os.path.basename(ref_audio_path)
+                if "【" in basename and "】" in basename:
+                    start_idx = basename.find("】") + 1
+                    end_idx = basename.rfind(".")
+                    if start_idx > 0 and end_idx > start_idx:
+                        req["prompt_text"] = basename[start_idx:end_idx]
+                    else:
+                        req["prompt_text"] = basename[:end_idx] if end_idx > 0 else basename
+                else:
+                    end_idx = basename.rfind(".")
+                    req["prompt_text"] = basename[:end_idx] if end_idx > 0 else basename
+    
+    if streaming_mode:
+        req["return_fragment"] = True
+    
+    try:
+        print(f"当前请求设备: {req.get('device')}")
+        if hasattr(tts_pipeline, 't2s_model') and tts_pipeline.t2s_model is not None:
+            print(f"t2s_model 设备: {next(tts_pipeline.t2s_model.parameters()).device}")
+        if hasattr(tts_pipeline, 'vits_model') and tts_pipeline.vits_model is not None:
+            print(f"vits_model 设备: {next(tts_pipeline.vits_model.parameters()).device}")
+        
+        # 异步执行 TTS 任务
+        loop = asyncio.get_event_loop()
+        tts_generator = await loop.run_in_executor(executor, run_tts_pipeline, req)
+        
+        if streaming_mode:
+            def streaming_generator():
+                if media_type == "wav":
+                    yield wave_header_chunk()
+                    stream_type = "raw"
+                else:
+                    stream_type = media_type
+                for sr, chunk in tts_generator:
+                    buf = pack_audio(chunk, sr, stream_type)
+                    yield buf.getvalue()
+            return StreamingResponse(streaming_generator(), media_type=f"audio/{media_type}")
+        else:
+            sr, audio_data = next(tts_generator)
+            buf = pack_audio(audio_data, sr, media_type)
+            return Response(buf.getvalue(), media_type=f"audio/{media_type}")
+    except Exception as e:
+        return JSONResponse(status_code=400, content={"status": "error", "message": "tts failed", "exception": str(e)})
+
+@APP.get("/control")
+async def control(command: str = None):
+    if command is None:
+        return JSONResponse(status_code=400, content={"status": "error", "message": "command is required"})
+    handle_control(command)
+
+@APP.get("/tts")
+async def tts_get_endpoint(
+    text: str,
+    ref_audio_path: str,
+    prompt_lang: str,
+    text_lang: str = "auto",
+    aux_ref_audio_paths: Optional[List[str]] = None,
+    prompt_text: Optional[str] = "",
+    top_k: Optional[int] = 5,
+    top_p: Optional[float] = 1,
+    temperature: Optional[float] = 1,
+    text_split_method: Optional[str] = "cut0",
+    batch_size: Optional[int] = 1,
+    batch_threshold: Optional[float] = 0.75,
+    split_bucket: Optional[bool] = True,
+    speed_factor: Optional[float] = 1.0,
+    fragment_interval: Optional[float] = 0.3,
+    seed: Optional[int] = -1,
+    media_type: Optional[str] = "wav",
+    streaming_mode: Optional[bool] = False,
+    parallel_infer: Optional[bool] = True,
+    repetition_penalty: Optional[float] = 1.35,
+    device: Optional[str] = None
+):
+    req = {
+        "text": text,
+        "text_lang": text_lang.lower(),
+        "ref_audio_path": ref_audio_path,
+        "aux_ref_audio_paths": aux_ref_audio_paths,
+        "prompt_lang": prompt_lang.lower(),
+        "prompt_text": prompt_text,
+        "top_k": top_k,
+        "top_p": top_p,
+        "temperature": temperature,
+        "text_split_method": text_split_method,
+        "batch_size": batch_size,
+        "batch_threshold": batch_threshold,
+        "split_bucket": split_bucket,
+        "speed_factor": speed_factor,
+        "fragment_interval": fragment_interval,
+        "seed": seed,
+        "media_type": media_type,
+        "streaming_mode": streaming_mode,
+        "parallel_infer": parallel_infer,
+        "repetition_penalty": repetition_penalty,
+        "device": device
+    }
+    return await tts_handle(req)
+
+@APP.post("/tts")
+async def tts_post_endpoint(request: TTS_Request):
+    req = request.dict(exclude_unset=True)
+    if "text_lang" in req:
+        req["text_lang"] = req["text_lang"].lower()
+    if "prompt_lang" in req:
+        req["prompt_lang"] = req["prompt_lang"].lower()
+    return await tts_handle(req)
+
+@APP.get("/ttsrole")
+async def ttsrole_get_endpoint(
+    text: str,
+    role: str,
+    text_lang: str = "auto",
+    ref_audio_path: Optional[str] = None,
+    aux_ref_audio_paths: Optional[List[str]] = None,
+    prompt_lang: Optional[str] = None,
+    prompt_text: Optional[str] = None,
+    emotion: Optional[str] = None,
+    top_k: Optional[int] = 5,
+    top_p: Optional[float] = 1,
+    temperature: Optional[float] = 1,
+    text_split_method: Optional[str] = "cut5",
+    batch_size: Optional[int] = 1,
+    batch_threshold: Optional[float] = 0.75,
+    split_bucket: Optional[bool] = True,
+    speed_factor: Optional[float] = 1.0,
+    fragment_interval: Optional[float] = 0.3,
+    seed: Optional[int] = -1,
+    media_type: Optional[str] = "wav",
+    streaming_mode: Optional[bool] = False,
+    parallel_infer: Optional[bool] = True,
+    repetition_penalty: Optional[float] = 1.35,
+    bert_base_path: Optional[str] = None,
+    cnhuhbert_base_path: Optional[str] = None,
+    device: Optional[str] = None,
+    is_half: Optional[bool] = None,
+    t2s_weights_path: Optional[str] = None,
+    version: Optional[str] = None,
+    vits_weights_path: Optional[str] = None,
+    t2s_model_path: Optional[str] = None,
+    vits_model_path: Optional[str] = None,
+    t2s_model_device: Optional[str] = None,
+    vits_model_device: Optional[str] = None
+):
+    req = {
+        "text": text,
+        "role": role,
+        "text_lang": text_lang.lower(),
+        "ref_audio_path": ref_audio_path,
+        "aux_ref_audio_paths": aux_ref_audio_paths,
+        "prompt_lang": prompt_lang.lower() if prompt_lang else None,
+        "prompt_text": prompt_text,
+        "emotion": emotion,
+        "top_k": top_k,
+        "top_p": top_p,
+        "temperature": temperature,
+        "text_split_method": text_split_method,
+        "batch_size": batch_size,
+        "batch_threshold": batch_threshold,
+        "split_bucket": split_bucket,
+        "speed_factor": speed_factor,
+        "fragment_interval": fragment_interval,
+        "seed": seed,
+        "media_type": media_type,
+        "streaming_mode": streaming_mode,
+        "parallel_infer": parallel_infer,
+        "repetition_penalty": repetition_penalty,
+        "bert_base_path": bert_base_path,
+        "cnhuhbert_base_path": cnhuhbert_base_path,
+        "device": device,
+        "is_half": is_half,
+        "t2s_weights_path": t2s_weights_path,
+        "version": version,
+        "vits_weights_path": vits_weights_path,
+        "t2s_model_path": t2s_model_path,
+        "vits_model_path": vits_model_path,
+        "t2s_model_device": t2s_model_device,
+        "vits_model_device": vits_model_device
+    }
+    return await tts_handle(req, is_ttsrole=True)
+
+@APP.post("/ttsrole")
+async def ttsrole_post_endpoint(request: TTSRole_Request):
+    req = request.dict(exclude_unset=True)
+    if "text_lang" in req:
+        req["text_lang"] = req["text_lang"].lower()
+    if "prompt_lang" in req:
+        req["prompt_lang"] = req["prompt_lang"].lower()
+    return await tts_handle(req, is_ttsrole=True)
+
+@APP.get("/set_gpt_weights")
+async def set_gpt_weights(weights_path: str = None):
+    try:
+        if not weights_path:
+            return JSONResponse(status_code=400, content={"status": "error", "message": "gpt weight path is required"})
+        tts_pipeline.init_t2s_weights(weights_path)
+        tts_config.t2s_weights_path = weights_path
+        return JSONResponse(status_code=200, content={"status": "success", "message": "success"})
+    except Exception as e:
+        return JSONResponse(status_code=400, content={"status": "error", "message": f"change gpt weight failed", "exception": str(e)})
+
+@APP.get("/set_sovits_weights")
+async def set_sovits_weights(weights_path: str = None):
+    try:
+        if not weights_path:
+            return JSONResponse(status_code=400, content={"status": "error", "message": "sovits weight path is required"})
+        tts_pipeline.init_vits_weights(weights_path)
+        tts_config.vits_weights_path = weights_path
+        return JSONResponse(status_code=200, content={"status": "success", "message": "success"})
+    except Exception as e:
+        return JSONResponse(status_code=400, content={"status": "error", "message": f"change sovits weight failed", "exception": str(e)})
+
+@APP.get("/set_refer_audio")
+async def set_refer_audio(refer_audio_path: str = None):
+    try:
+        if not refer_audio_path:
+            return JSONResponse(status_code=400, content={"status": "error", "message": "refer audio path is required"})
+        tts_pipeline.set_ref_audio(refer_audio_path)
+        return JSONResponse(status_code=200, content={"status": "success", "message": "success"})
+    except Exception as e:
+        return JSONResponse(status_code=400, content={"status": "error", "message": f"set refer audio failed", "exception": str(e)})
+
+if __name__ == "__main__":
+    try:
+        if host == 'None':  # 在调用时使用 -a None 参数，可以让api监听双栈
+            host = None
+        uvicorn.run(app=APP, host=host, port=port, workers=1)
+    except Exception as e:
+        traceback.print_exc()
+        os.kill(os.getpid(), signal.SIGTERM)
+        exit(0)

docs/cn/Changelog_CN.md

@@ -594,11 +594,11 @@
   - 内容: 修复实验名结尾出现空格在win中路径不正确的问题
   - 类型: 修复
   - 提交: RVC-Boss
-- 2025.06.10 [Commit#746cb536](https://github.com/RVC-Boss/GPT-SoVITS/commit/746cb536c68b1fe6ce3ca7e882235375b8a8dd89)
+- 2025.06.10 [PR#2449](https://github.com/RVC-Boss/GPT-SoVITS/pull/2449)
   - 内容: 语种分割优化
   - 类型: 优化
   - 提交: KamioRinn
-- 2025.06.11 [Commit#dd2b9253](https://github.com/RVC-Boss/GPT-SoVITS/commit/dd2b9253aabb09db32db7a3344570ed9df043351)
+- 2025.06.11 [PR#2450](https://github.com/RVC-Boss/GPT-SoVITS/pull/2450)
   - 内容: 修复并行推理对v2pro支持bug
   - 类型: 修复
   - 提交: YYuX-1145
@@ -606,21 +606,132 @@
   - 内容: v2pro对ge提取时会出现数值溢出的问题修复
   - 类型: 修复
   - 提交: RVC-Boss
-- 2025.06.11 [Commit#37f5abfc](https://github.com/RVC-Boss/GPT-SoVITS/commit/6fdc67ca83418306f11e90b9139278313ac5c3e9)[Commit#6fdc67ca](https://github.com/RVC-Boss/GPT-SoVITS/commit/37f5abfcb4a6553652235909db2e124b6f8ff3a5)
+- 2025.06.17 [PR#2464](https://github.com/RVC-Boss/GPT-SoVITS/pull/2464) [PR#2482](https://github.com/RVC-Boss/GPT-SoVITS/pull/2482)
   - 内容: install.sh逻辑优化
   - 类型: 优化
   - 提交: XXXXRT666
-- 2025.06.27 [Commit#90ebefa7](https://github.com/RVC-Boss/GPT-SoVITS/commit/90ebefa78fd544da36eebe0b2003620879c921b0)
+- 2025.06.27 [PR#2489](https://github.com/RVC-Boss/GPT-SoVITS/pull/2489)
   - 内容: onnxruntime加载逻辑优化（对gpu/cpu的判断）
   - 类型: 优化
   - 提交: KamioRinn
-- 2025.06.27 [Commit#6df61f58](https://github.com/RVC-Boss/GPT-SoVITS/commit/6df61f58e4d18d4c2ad9d1eddd6a1bd690034c23)
+- 2025.06.27 [PR#2488](https://github.com/RVC-Boss/GPT-SoVITS/pull/2488)
   - 内容: 语言分割及格式化优化
   - 类型: 优化
   - 提交: KamioRinn
+
+## 202507
+
 - 2025.07.10 [Commit#426e1a2bb](https://github.com/RVC-Boss/GPT-SoVITS/commit/426e1a2bb43614af2479b877c37acfb0591e952f)
   - 内容: 提升推理进程优先级（修复win11下可能GPU利用率受限的问题）
-  - 类型: 修复
+  - 类型: 优化
   - 提交: XianYue0125
+- 2025.07.16 [PR#2490](https://github.com/RVC-Boss/GPT-SoVITS/pull/2490)
+  - 内容: 解决 TTS.py 无法识别真正支持版本 v2Pro、v2ProPlus 的问题, 同时更新一版默认配置。
+  - 类型: 修复
+  - 提交: jiangsier-xyz
+- 2025.07.16 [Commit#4d8ebf85](https://github.com/RVC-Boss/GPT-SoVITS/commit/4d8ebf85233d4f1166d7cc02fdc595602975ca8f)
+  - 内容: 修复并行推理模式下v2pro模型识别问题
+  - 类型: 修复
+  - 提交: RVC-Boss
+- 2025.07.17 [PR#2531](https://github.com/RVC-Boss/GPT-SoVITS/pull/2531)
+  - 内容: whisper asr支持性价比更高的distill模型
+  - 类型: 优化
+  - 提交: XXXXRT666
+- 2025.07.18 [PR#2536](https://github.com/RVC-Boss/GPT-SoVITS/pull/2536)
+  - 内容: 优化TTS_Config的代码逻辑
+  - 类型: 优化
+  - 提交: ChasonJiang
+- 2025.07.18 [PR#2537](https://github.com/RVC-Boss/GPT-SoVITS/pull/2537)
+  - 内容: 修复gpt的loss计算问题
+  - 类型: 修复
+  - 提交: ChasonJiang
 
+## 202508
 
+- 2025.08.02 [PR#2561](https://github.com/RVC-Boss/GPT-SoVITS/pull/2561)
+  - 内容: WSL Rocm
+  - 类型: 修复
+  - 提交: XXXXRT666
+
+## 202509
+
+- 2025.09.10 [Commit#11aa78bd](https://github.com/RVC-Boss/GPT-SoVITS/commit/11aa78bd9bda8b53047cfcae03abf7ca94d27391)
+  - 内容: 修复环境变量可能不为str的问题
+  - 类型: 修复
+  - 提交: RVC-Boss
+
+## 202511
+
+- 2025.11.28 [PR#2671](https://github.com/RVC-Boss/GPT-SoVITS/pull/2671) [PR#2678](https://github.com/RVC-Boss/GPT-SoVITS/pull/2678)
+  - 内容: 流式推理
+  - 类型: 新功能
+  - 提交: ChasonJiang
+- 2025.11.28 [PR#2636](https://github.com/RVC-Boss/GPT-SoVITS/pull/2636)
+  - 内容: 数学计算文本前端逻辑优化
+  - 类型: 优化
+  - 提交: KamioRinn
+- 2025.11.28 [PR#2469](https://github.com/RVC-Boss/GPT-SoVITS/pull/2469)
+  - 内容: 流式推理
+  - 类型: 新功能
+  - 提交: L-jasmine
+- 2025.11.28 [PR#2577](https://github.com/RVC-Boss/GPT-SoVITS/pull/2577)
+  - 内容: 支持vq分布式训练
+  - 类型: 优化
+  - 提交: wzy3650
+- 2025.11.28 [PR#2627](https://github.com/RVC-Boss/GPT-SoVITS/pull/2627) [PR#2679](https://github.com/RVC-Boss/GPT-SoVITS/pull/2679)
+  - 内容: ASR模型下载逻辑优化
+  - 类型: 优化
+  - 提交: XXXXRT666
+- 2025.11.28 [PR#2662](https://github.com/RVC-Boss/GPT-SoVITS/pull/2662)
+  - 内容: default batch size bug 修复
+  - 类型: 修复
+  - 提交: Spr-Aachen
+
+## 202512
+
+- 2025.12.30 [PR#2703](https://github.com/RVC-Boss/GPT-SoVITS/pull/2703) [PR#2704](https://github.com/RVC-Boss/GPT-SoVITS/pull/2704)
+  - 内容: 修复采样错误
+  - 类型: 修复
+  - 提交: ChasonJiang
+
+## 202602
+
+- 2026.02.08 [PR#2727](https://github.com/RVC-Boss/GPT-SoVITS/pull/2727)
+  - 内容: 修复 Conda 条款未同意导致的构建失败
+  - 类型: 修复
+  - 提交: Oarora
+- 2026.02.09 [PR#2732](https://github.com/RVC-Boss/GPT-SoVITS/pull/2732)
+  - 内容: 环境自动构建优化
+  - 类型: 优化
+  - 提交: XXXXRT666
+
+## 202604
+
+- 2026.04.18 [PR#2763](https://github.com/RVC-Boss/GPT-SoVITS/pull/2763)
+  - 内容: 优化 G2PW 的推理输入构造与多音字处理流程，减少重复计算，降低长句场景下的推理开销
+  - 类型: 优化
+  - 提交: baicai-1145
+- 2026.04.18 [PR#2767](https://github.com/RVC-Boss/GPT-SoVITS/pull/2767)
+  - 内容: 改进 Windows 单卡 v3 LoRA 训练流程
+  - 类型: 优化
+  - 提交: 2409324124
+- 2026.04.18 [PR#2755](https://github.com/RVC-Boss/GPT-SoVITS/pull/2755)
+  - 内容: 修复多个模块中的独立 bug
+  - 类型: 修复
+  - 提交: wishhyt
+- 2026.04.18 [PR#2758](https://github.com/RVC-Boss/GPT-SoVITS/pull/2758)
+  - 内容: 添加数据集的错误处理提示
+  - 类型: 优化
+  - 提交: mushroomcowisheggs
+- 2026.04.18 [PR#2753](https://github.com/RVC-Boss/GPT-SoVITS/pull/2753)
+  - 内容: 并行推理部分bug修复
+  - 类型: 修复
+  - 提交: wishhyt
+- 2026.04.18 [PR#2733](https://github.com/RVC-Boss/GPT-SoVITS/pull/2733)
+  - 内容: bug修复：DPO 训练不支持漏字模拟
+  - 类型: 修复
+  - 提交: Mr-Neutr0n
+- 2026.04.18 [Commit#02425ea](https://github.com/RVC-Boss/GPT-SoVITS/commit/02425ea25680c26c700be0bc158756c69103d827)
+  - 内容: 修复onnx脚本未导入Optional等的问题
+  - 类型: 修复
+  - 提交: RVC-Boss

docs/en/Changelog_EN.md

@@ -578,3 +578,160 @@
   - Content: Optimized automatic precision detection logic; added collapsible functionality to WebUI frontend modules.
   - Type: New Feature
   - Contributors: XXXXRT666, RVC-Boss
+- 2025.06.06 [PR#2427](https://github.com/RVC-Boss/GPT-SoVITS/pull/2427)
+  - Content: Fix polyphone detection for "X一X" pattern
+  - Type: Fix
+  - Contributor: wzy3650
+- 2025.06.05 [PR#2439](https://github.com/RVC-Boss/GPT-SoVITS/pull/2439)
+  - Content: Config fix; fix SoVITS model loading
+  - Type: Fix
+  - Contributor: wzy3650
+- 2025.06.09 [Commit#8056efe4](https://github.com/RVC-Boss/GPT-SoVITS/commit/8056efe4ab7bbc3610c72ae356a6f37518441f7d)
+  - Content: Fix possible numerical explosion of `ge.sum` causing silent inference
+  - Type: Fix
+  - Contributor: RVC-Boss
+- 2025.06.10 [Commit#2c0436b9](https://github.com/RVC-Boss/GPT-SoVITS/commit/2c0436b9ce397424ae03476c836fb64c6e5ebcc6)
+  - Content: Fix incorrect Windows path when experiment name ends with a space
+  - Type: Fix
+  - Contributor: RVC-Boss
+- 2025.06.10 [PR#2449](https://github.com/RVC-Boss/GPT-SoVITS/pull/2449)
+  - Content: Optimize language segmentation
+  - Type: Optimization
+  - Contributor: KamioRinn
+- 2025.06.11 [PR#2450](https://github.com/RVC-Boss/GPT-SoVITS/pull/2450)
+  - Content: Fix bug in parallel inference support for v2pro
+  - Type: Fix
+  - Contributor: YYuX-1145
+- 2025.06.11 [Commit#ed89a023](https://github.com/RVC-Boss/GPT-SoVITS/commit/ed89a023378dabba9d4b6580235bb9742245816d)
+  - Content: Fix numerical overflow issue when extracting `ge` for v2pro
+  - Type: Fix
+  - Contributor: RVC-Boss
+- 2025.06.17 [PR#2464](https://github.com/RVC-Boss/GPT-SoVITS/pull/2464) [PR#2482](https://github.com/RVC-Boss/GPT-SoVITS/pull/2482)
+  - Content: Optimize `install.sh` logic
+  - Type: Optimization
+  - Contributor: XXXXRT666
+- 2025.06.27 [PR#2489](https://github.com/RVC-Boss/GPT-SoVITS/pull/2489)
+  - Content: Optimize onnxruntime loading logic (GPU/CPU detection)
+  - Type: Optimization
+  - Contributor: KamioRinn
+- 2025.06.27 [PR#2488](https://github.com/RVC-Boss/GPT-SoVITS/pull/2488)
+  - Content: Optimize language segmentation and formatting
+  - Type: Optimization
+  - Contributor: KamioRinn
+
+## 202507
+
+- 2025.07.10 [Commit#426e1a2bb](https://github.com/RVC-Boss/GPT-SoVITS/commit/426e1a2bb43614af2479b877c37acfb0591e952f)
+  - Content: Increase inference process priority (fix possible GPU utilization limitation on Win11)
+  - Type: Optimization
+  - Contributor: XianYue0125
+- 2025.07.16 [PR#2490](https://github.com/RVC-Boss/GPT-SoVITS/pull/2490)
+  - Content: Fix TTS.py not recognizing actually supported versions v2Pro and v2ProPlus, and update default configuration
+  - Type: Fix
+  - Contributor: jiangsier-xyz
+- 2025.07.16 [Commit#4d8ebf85](https://github.com/RVC-Boss/GPT-SoVITS/commit/4d8ebf85233d4f1166d7cc02fdc595602975ca8f)
+  - Content: Fix v2pro model recognition issue in parallel inference mode
+  - Type: Fix
+  - Contributor: RVC-Boss
+- 2025.07.17 [PR#2531](https://github.com/RVC-Boss/GPT-SoVITS/pull/2531)
+  - Content: Whisper ASR supports more cost-effective distill models
+  - Type: Optimization
+  - Contributor: XXXXRT666
+- 2025.07.18 [PR#2536](https://github.com/RVC-Boss/GPT-SoVITS/pull/2536)
+  - Content: Optimize `TTS_Config` code logic
+  - Type: Optimization
+  - Contributor: ChasonJiang
+- 2025.07.18 [PR#2537](https://github.com/RVC-Boss/GPT-SoVITS/pull/2537)
+  - Content: Fix GPT loss calculation issue
+  - Type: Fix
+  - Contributor: ChasonJiang
+
+## 202508
+
+- 2025.08.02 [PR#2561](https://github.com/RVC-Boss/GPT-SoVITS/pull/2561)
+  - Content: WSL Rocm
+  - Type: Fix
+  - Contributor: XXXXRT666
+
+## 202509
+
+- 2025.09.10 [Commit#11aa78bd](https://github.com/RVC-Boss/GPT-SoVITS/commit/11aa78bd9bda8b53047cfcae03abf7ca94d27391)
+  - Content: Fix issue where environment variable may not be a string
+  - Type: Fix
+  - Contributor: RVC-Boss
+
+## 202511
+
+- 2025.11.28 [PR#2671](https://github.com/RVC-Boss/GPT-SoVITS/pull/2671) [PR#2678](https://github.com/RVC-Boss/GPT-SoVITS/pull/2678)
+  - Content: Streaming inference
+  - Type: New Feature
+  - Contributor: ChasonJiang
+- 2025.11.28 [PR#2636](https://github.com/RVC-Boss/GPT-SoVITS/pull/2636)
+  - Content: Optimize text frontend logic for mathematical expression text
+  - Type: Optimization
+  - Contributor: KamioRinn
+- 2025.11.28 [PR#2469](https://github.com/RVC-Boss/GPT-SoVITS/pull/2469)
+  - Content: Streaming inference
+  - Type: New Feature
+  - Contributor: L-jasmine
+- 2025.11.28 [PR#2577](https://github.com/RVC-Boss/GPT-SoVITS/pull/2577)
+  - Content: Support VQ distributed training
+  - Type: Optimization
+  - Contributor: wzy3650
+- 2025.11.28 [PR#2627](https://github.com/RVC-Boss/GPT-SoVITS/pull/2627) [PR#2679](https://github.com/RVC-Boss/GPT-SoVITS/pull/2679)
+  - Content: Optimize ASR model download logic
+  - Type: Optimization
+  - Contributor: XXXXRT666
+- 2025.11.28 [PR#2662](https://github.com/RVC-Boss/GPT-SoVITS/pull/2662)
+  - Content: Fix default batch size bug
+  - Type: Fix
+  - Contributor: Spr-Aachen
+
+## 202512
+
+- 2025.12.30 [PR#2703](https://github.com/RVC-Boss/GPT-SoVITS/pull/2703) [PR#2704](https://github.com/RVC-Boss/GPT-SoVITS/pull/2704)
+  - Content: Fix sampling error
+  - Type: Fix
+  - Contributor: ChasonJiang
+
+## 202602
+
+- 2026.02.08 [PR#2727](https://github.com/RVC-Boss/GPT-SoVITS/pull/2727)
+  - Content: Fix build failure caused by unaccepted Conda terms
+  - Type: Fix
+  - Contributor: Oarora
+- 2026.02.09 [PR#2732](https://github.com/RVC-Boss/GPT-SoVITS/pull/2732)
+  - Content: Optimize automatic environment setup
+  - Type: Optimization
+  - Contributor: XXXXRT666
+
+## 202604
+
+- 2026.04.18 [PR#2763](https://github.com/RVC-Boss/GPT-SoVITS/pull/2763)
+  - Content: Optimize G2PW inference input construction and polyphone handling to reduce redundant computation and inference overhead for long sentences
+  - Type: Optimization
+  - Contributor: baicai-1145
+- 2026.04.18 [PR#2767](https://github.com/RVC-Boss/GPT-SoVITS/pull/2767)
+  - Content: Improve the LoRA training flow for GPT-SoVITS v3 on a single card under Windows
+  - Type: Optimization
+  - Contributor: 2409324124
+- 2026.04.18 [PR#2755](https://github.com/RVC-Boss/GPT-SoVITS/pull/2755)
+  - Content: Fix miscellaneous bugs in multiple modules
+  - Type: Fix
+  - Contributor: wishhyt
+- 2026.04.18 [PR#2758](https://github.com/RVC-Boss/GPT-SoVITS/pull/2758)
+  - Content: Add error handling hints for dataset processing
+  - Type: Optimization
+  - Contributor: mushroomcowisheggs
+- 2026.04.18 [PR#2753](https://github.com/RVC-Boss/GPT-SoVITS/pull/2753)
+  - Content: Fix some bugs in parallel inference
+  - Type: Fix
+  - Contributor: wishhyt
+- 2026.04.18 [PR#2733](https://github.com/RVC-Boss/GPT-SoVITS/pull/2733)
+  - Content: Fix bug where DPO training does not support missing word simulation
+  - Type: Fix
+  - Contributor: Mr-Neutr0n
+- 2026.04.18 [Commit#02425ea](https://github.com/RVC-Boss/GPT-SoVITS/commit/02425ea25680c26c700be0bc158756c69103d827)
+  - Content: Fix missing imports (e.g., Optional) in ONNX script
+  - Type: Fix
+  - Contributor: RVC-Boss

docs/ja/Changelog_JA.md

@@ -578,3 +578,160 @@
   - 内容: 自動精度検出ロジックを最適化し、WebUI フロントエンドモジュールに折り畳み（Collapsible）機能を追加
   - タイプ: 新機能
   - 貢献者: XXXXRT666, RVC-Boss
+- 2025.06.06 [PR#2427](https://github.com/RVC-Boss/GPT-SoVITS/pull/2427)
+  - 内容: 「X一X」パターンの多音字検出を修正
+  - タイプ: 修正
+  - 貢献者: wzy3650
+- 2025.06.05 [PR#2439](https://github.com/RVC-Boss/GPT-SoVITS/pull/2439)
+  - 内容: 設定の修正；SoVITSモデル読み込みの修正
+  - タイプ: 修正
+  - 貢献者: wzy3650
+- 2025.06.09 [Commit#8056efe4](https://github.com/RVC-Boss/GPT-SoVITS/commit/8056efe4ab7bbc3610c72ae356a6f37518441f7d)
+  - 内容: `ge.sum`の数値爆発による推論の無音化を修正
+  - タイプ: 修正
+  - 貢献者: RVC-Boss
+- 2025.06.10 [Commit#2c0436b9](https://github.com/RVC-Boss/GPT-SoVITS/commit/2c0436b9ce397424ae03476c836fb64c6e5ebcc6)
+  - 内容: 実験名がスペースで終わる場合のWindowsパスの誤りを修正
+  - タイプ: 修正
+  - 貢献者: RVC-Boss
+- 2025.06.10 [PR#2449](https://github.com/RVC-Boss/GPT-SoVITS/pull/2449)
+  - 内容: 言語分割の最適化
+  - タイプ: 最適化
+  - 貢献者: KamioRinn
+- 2025.06.11 [PR#2450](https://github.com/RVC-Boss/GPT-SoVITS/pull/2450)
+  - 内容: v2proの並列推論対応におけるバグを修正
+  - タイプ: 修正
+  - 貢献者: YYuX-1145
+- 2025.06.11 [Commit#ed89a023](https://github.com/RVC-Boss/GPT-SoVITS/commit/ed89a023378dabba9d4b6580235bb9742245816d)
+  - 内容: v2proの`ge`抽出時の数値オーバーフロー問題を修正
+  - タイプ: 修正
+  - 貢献者: RVC-Boss
+- 2025.06.17 [PR#2464](https://github.com/RVC-Boss/GPT-SoVITS/pull/2464) [PR#2482](https://github.com/RVC-Boss/GPT-SoVITS/pull/2482)
+  - 内容: `install.sh`のロジックを最適化
+  - タイプ: 最適化
+  - 貢献者: XXXXRT666
+- 2025.06.27 [PR#2489](https://github.com/RVC-Boss/GPT-SoVITS/pull/2489)
+  - 内容: onnxruntime読み込みロジックを最適化（GPU/CPU検出）
+  - タイプ: 最適化
+  - 貢献者: KamioRinn
+- 2025.06.27 [PR#2488](https://github.com/RVC-Boss/GPT-SoVITS/pull/2488)
+  - 内容: 言語分割と書式を最適化
+  - タイプ: 最適化
+  - 貢献者: KamioRinn
+
+## 202507
+
+- 2025.07.10 [Commit#426e1a2bb](https://github.com/RVC-Boss/GPT-SoVITS/commit/426e1a2bb43614af2479b877c37acfb0591e952f)
+  - 内容: 推論プロセスの優先度を上げる（Win11でのGPU利用制限の可能性を修正）
+  - タイプ: 最適化
+  - 貢献者: XianYue0125
+- 2025.07.16 [PR#2490](https://github.com/RVC-Boss/GPT-SoVITS/pull/2490)
+  - 内容: TTS.pyが実際にサポートされているバージョンv2Proおよびv2ProPlusを認識しない問題を修正し、デフォルト設定を更新
+  - タイプ: 修正
+  - 貢献者: jiangsier-xyz
+- 2025.07.16 [Commit#4d8ebf85](https://github.com/RVC-Boss/GPT-SoVITS/commit/4d8ebf85233d4f1166d7cc02fdc595602975ca8f)
+  - 内容: 並列推論モードでのv2proモデル認識問題を修正
+  - タイプ: 修正
+  - 貢献者: RVC-Boss
+- 2025.07.17 [PR#2531](https://github.com/RVC-Boss/GPT-SoVITS/pull/2531)
+  - 内容: Whisper ASRがよりコスト効率の高い蒸留モデルをサポート
+  - タイプ: 最適化
+  - 貢献者: XXXXRT666
+- 2025.07.18 [PR#2536](https://github.com/RVC-Boss/GPT-SoVITS/pull/2536)
+  - 内容: `TTS_Config`のコードロジックを最適化
+  - タイプ: 最適化
+  - 貢献者: ChasonJiang
+- 2025.07.18 [PR#2537](https://github.com/RVC-Boss/GPT-SoVITS/pull/2537)
+  - 内容: GPT損失計算の問題を修正
+  - タイプ: 修正
+  - 貢献者: ChasonJiang
+
+## 202508
+
+- 2025.08.02 [PR#2561](https://github.com/RVC-Boss/GPT-SoVITS/pull/2561)
+  - 内容: WSL Rocm対応
+  - タイプ: 修正
+  - 貢献者: XXXXRT666
+
+## 202509
+
+- 2025.09.10 [Commit#11aa78bd](https://github.com/RVC-Boss/GPT-SoVITS/commit/11aa78bd9bda8b53047cfcae03abf7ca94d27391)
+  - 内容: 環境変数が文字列でない可能性がある問題を修正
+  - タイプ: 修正
+  - 貢献者: RVC-Boss
+
+## 202511
+
+- 2025.11.28 [PR#2671](https://github.com/RVC-Boss/GPT-SoVITS/pull/2671) [PR#2678](https://github.com/RVC-Boss/GPT-SoVITS/pull/2678)
+  - 内容: ストリーミング推論
+  - タイプ: 新機能
+  - 貢献者: ChasonJiang
+- 2025.11.28 [PR#2636](https://github.com/RVC-Boss/GPT-SoVITS/pull/2636)
+  - 内容: 数式テキストに対するテキスト前処理ロジックを最適化
+  - タイプ: 最適化
+  - 貢献者: KamioRinn
+- 2025.11.28 [PR#2469](https://github.com/RVC-Boss/GPT-SoVITS/pull/2469)
+  - 内容: ストリーミング推論
+  - タイプ: 新機能
+  - 貢献者: L-jasmine
+- 2025.11.28 [PR#2577](https://github.com/RVC-Boss/GPT-SoVITS/pull/2577)
+  - 内容: VQ分散学習をサポート
+  - タイプ: 最適化
+  - 貢献者: wzy3650
+- 2025.11.28 [PR#2627](https://github.com/RVC-Boss/GPT-SoVITS/pull/2627) [PR#2679](https://github.com/RVC-Boss/GPT-SoVITS/pull/2679)
+  - 内容: ASRモデルダウンロードロジックを最適化
+  - タイプ: 最適化
+  - 貢献者: XXXXRT666
+- 2025.11.28 [PR#2662](https://github.com/RVC-Boss/GPT-SoVITS/pull/2662)
+  - 内容: デフォルトのバッチサイズのバグを修正
+  - タイプ: 修正
+  - 貢献者: Spr-Aachen
+
+## 202512
+
+- 2025.12.30 [PR#2703](https://github.com/RVC-Boss/GPT-SoVITS/pull/2703) [PR#2704](https://github.com/RVC-Boss/GPT-SoVITS/pull/2704)
+  - 内容: サンプリングエラーを修正
+  - タイプ: 修正
+  - 貢献者: ChasonJiang
+
+## 202602
+
+- 2026.02.08 [PR#2727](https://github.com/RVC-Boss/GPT-SoVITS/pull/2727)
+  - 内容: 受け入れられなかったConda利用規約によるビルド失敗を修正
+  - タイプ: 修正
+  - 貢献者: Oarora
+- 2026.02.09 [PR#2732](https://github.com/RVC-Boss/GPT-SoVITS/pull/2732)
+  - 内容: 自動環境セットアップを最適化
+  - タイプ: 最適化
+  - 貢献者: XXXXRT666
+
+## 202604
+
+- 2026.04.18 [PR#2763](https://github.com/RVC-Boss/GPT-SoVITS/pull/2763)
+  - 内容: G2PW推論入力の構築と多音字処理を最適化し、長文における冗長な計算と推論オーバーヘッドを削減
+  - タイプ: 最適化
+  - 貢献者: baicai-1145
+- 2026.04.18 [PR#2767](https://github.com/RVC-Boss/GPT-SoVITS/pull/2767)
+  - 内容: WindowsでのシングルカードにおけるGPT-SoVITS v3のLoRAトレーニングフローを改善
+  - タイプ: 最適化
+  - 貢献者: 2409324124
+- 2026.04.18 [PR#2755](https://github.com/RVC-Boss/GPT-SoVITS/pull/2755)
+  - 内容: 複数モジュールの雑多なバグを修正
+  - タイプ: 修正
+  - 貢献者: wishhyt
+- 2026.04.18 [PR#2758](https://github.com/RVC-Boss/GPT-SoVITS/pull/2758)
+  - 内容: データセット処理時のエラーハンドリングヒントを追加
+  - タイプ: 最適化
+  - 貢献者: mushroomcowisheggs
+- 2026.04.18 [PR#2753](https://github.com/RVC-Boss/GPT-SoVITS/pull/2753)
+  - 内容: 並列推論の一部バグを修正
+  - タイプ: 修正
+  - 貢献者: wishhyt
+- 2026.04.18 [PR#2733](https://github.com/RVC-Boss/GPT-SoVITS/pull/2733)
+  - 内容: DPOトレーニングが欠落単語シミュレーションをサポートしないバグを修正
+  - タイプ: 修正
+  - 貢献者: Mr-Neutr0n
+- 2026.04.18 [Commit#02425ea](https://github.com/RVC-Boss/GPT-SoVITS/commit/02425ea25680c26c700be0bc158756c69103d827)
+  - 内容: ONNXスクリプトでの（Optionalなどの）不足インポートを修正
+  - タイプ: 修正
+  - 貢献者: RVC-Boss

docs/ko/Changelog_KO.md

@@ -578,3 +578,160 @@
   - 내용: 자동 정밀도 감지 로직 최적화; WebUI 프론트엔드 모듈에 접기 기능 추가
   - 유형: 신규 기능
   - 기여자: XXXXRT666, RVC-Boss
+- 2025.06.06 [PR#2427](https://github.com/RVC-Boss/GPT-SoVITS/pull/2427)
+  - 내용: "X一X" 패턴의 다중 발음 감지 오류 수정
+  - 유형: 수정
+  - 기여자: wzy3650
+- 2025.06.05 [PR#2439](https://github.com/RVC-Boss/GPT-SoVITS/pull/2439)
+  - 내용: 설정 오류 수정; SoVITS 모델 로딩 오류 수정
+  - 유형: 수정
+  - 기여자: wzy3650
+- 2025.06.09 [Commit#8056efe4](https://github.com/RVC-Boss/GPT-SoVITS/commit/8056efe4ab7bbc3610c72ae356a6f37518441f7d)
+  - 내용: `ge.sum`의 수치 폭발 가능성으로 인한 추론 무음 현상 수정
+  - 유형: 수정
+  - 기여자: RVC-Boss
+- 2025.06.10 [Commit#2c0436b9](https://github.com/RVC-Boss/GPT-SoVITS/commit/2c0436b9ce397424ae03476c836fb64c6e5ebcc6)
+  - 내용: 실험 이름이 공백으로 끝날 때 발생하는 잘못된 Windows 경로 문제 수정
+  - 유형: 수정
+  - 기여자: RVC-Boss
+- 2025.06.10 [PR#2449](https://github.com/RVC-Boss/GPT-SoVITS/pull/2449)
+  - 내용: 언어 분할 최적화
+  - 유형: 최적화
+  - 기여자: KamioRinn
+- 2025.06.11 [PR#2450](https://github.com/RVC-Boss/GPT-SoVITS/pull/2450)
+  - 내용: v2pro 병렬 추론 지원 버그 수정
+  - 유형: 수정
+  - 기여자: YYuX-1145
+- 2025.06.11 [Commit#ed89a023](https://github.com/RVC-Boss/GPT-SoVITS/commit/ed89a023378dabba9d4b6580235bb9742245816d)
+  - 내용: v2pro의 `ge` 추출 시 수치 오버플로우 문제 수정
+  - 유형: 수정
+  - 기여자: RVC-Boss
+- 2025.06.17 [PR#2464](https://github.com/RVC-Boss/GPT-SoVITS/pull/2464) [PR#2482](https://github.com/RVC-Boss/GPT-SoVITS/pull/2482)
+  - 내용: `install.sh` 로직 최적화
+  - 유형: 최적화
+  - 기여자: XXXXRT666
+- 2025.06.27 [PR#2489](https://github.com/RVC-Boss/GPT-SoVITS/pull/2489)
+  - 내용: onnxruntime 로딩 로직 최적화 (GPU/CPU 감지)
+  - 유형: 최적화
+  - 기여자: KamioRinn
+- 2025.06.27 [PR#2488](https://github.com/RVC-Boss/GPT-SoVITS/pull/2488)
+  - 내용: 언어 분할 및 형식 최적화
+  - 유형: 최적화
+  - 기여자: KamioRinn
+
+## 202507
+
+- 2025.07.10 [Commit#426e1a2bb](https://github.com/RVC-Boss/GPT-SoVITS/commit/426e1a2bb43614af2479b877c37acfb0591e952f)
+  - 내용: 추론 프로세스 우선순위 증가 (Win11에서 GPU 활용 제한 가능성 수정)
+  - 유형: 최적화
+  - 기여자: XianYue0125
+- 2025.07.16 [PR#2490](https://github.com/RVC-Boss/GPT-SoVITS/pull/2490)
+  - 내용: TTS.py가 실제 지원되는 버전 v2Pro 및 v2ProPlus를 인식하지 못하는 문제 수정 및 기본 설정 업데이트
+  - 유형: 수정
+  - 기여자: jiangsier-xyz
+- 2025.07.16 [Commit#4d8ebf85](https://github.com/RVC-Boss/GPT-SoVITS/commit/4d8ebf85233d4f1166d7cc02fdc595602975ca8f)
+  - 내용: 병렬 추론 모드에서 v2pro 모델 인식 문제 수정
+  - 유형: 수정
+  - 기여자: RVC-Boss
+- 2025.07.17 [PR#2531](https://github.com/RVC-Boss/GPT-SoVITS/pull/2531)
+  - 내용: Whisper ASR이 더 비용 효율적인 distill 모델 지원
+  - 유형: 최적화
+  - 기여자: XXXXRT666
+- 2025.07.18 [PR#2536](https://github.com/RVC-Boss/GPT-SoVITS/pull/2536)
+  - 내용: `TTS_Config` 코드 로직 최적화
+  - 유형: 최적화
+  - 기여자: ChasonJiang
+- 2025.07.18 [PR#2537](https://github.com/RVC-Boss/GPT-SoVITS/pull/2537)
+  - 내용: GPT 손실(loss) 계산 문제 수정
+  - 유형: 수정
+  - 기여자: ChasonJiang
+
+## 202508
+
+- 2025.08.02 [PR#2561](https://github.com/RVC-Boss/GPT-SoVITS/pull/2561)
+  - 내용: WSL Rocm
+  - 유형: 수정
+  - 기여자: XXXXRT666
+
+## 202509
+
+- 2025.09.10 [Commit#11aa78bd](https://github.com/RVC-Boss/GPT-SoVITS/commit/11aa78bd9bda8b53047cfcae03abf7ca94d27391)
+  - 내용: 환경 변수가 문자열이 아닐 수 있는 문제 수정
+  - 유형: 수정
+  - 기여자: RVC-Boss
+
+## 202511
+
+- 2025.11.28 [PR#2671](https://github.com/RVC-Boss/GPT-SoVITS/pull/2671) [PR#2678](https://github.com/RVC-Boss/GPT-SoVITS/pull/2678)
+  - 내용: 스트리밍 추론
+  - 유형: 새 기능
+  - 기여자: ChasonJiang
+- 2025.11.28 [PR#2636](https://github.com/RVC-Boss/GPT-SoVITS/pull/2636)
+  - 내용: 수학 표현식 텍스트에 대한 텍스트 전처리 로직 최적화
+  - 유형: 최적화
+  - 기여자: KamioRinn
+- 2025.11.28 [PR#2469](https://github.com/RVC-Boss/GPT-SoVITS/pull/2469)
+  - 내용: 스트리밍 추론
+  - 유형: 새 기능
+  - 기여자: L-jasmine
+- 2025.11.28 [PR#2577](https://github.com/RVC-Boss/GPT-SoVITS/pull/2577)
+  - 내용: VQ 분산 학습 지원
+  - 유형: 최적화
+  - 기여자: wzy3650
+- 2025.11.28 [PR#2627](https://github.com/RVC-Boss/GPT-SoVITS/pull/2627) [PR#2679](https://github.com/RVC-Boss/GPT-SoVITS/pull/2679)
+  - 내용: ASR 모델 다운로드 로직 최적화
+  - 유형: 최적화
+  - 기여자: XXXXRT666
+- 2025.11.28 [PR#2662](https://github.com/RVC-Boss/GPT-SoVITS/pull/2662)
+  - 내용: 기본 배치 크기 버그 수정
+  - 유형: 수정
+  - 기여자: Spr-Aachen
+
+## 202512
+
+- 2025.12.30 [PR#2703](https://github.com/RVC-Boss/GPT-SoVITS/pull/2703) [PR#2704](https://github.com/RVC-Boss/GPT-SoVITS/pull/2704)
+  - 내용: 샘플링 오류 수정
+  - 유형: 수정
+  - 기여자: ChasonJiang
+
+## 202602
+
+- 2026.02.08 [PR#2727](https://github.com/RVC-Boss/GPT-SoVITS/pull/2727)
+  - 내용: Conda 약관 미동의로 인한 빌드 실패 수정
+  - 유형: 수정
+  - 기여자: Oarora
+- 2026.02.09 [PR#2732](https://github.com/RVC-Boss/GPT-SoVITS/pull/2732)
+  - 내용: 자동 환경 설정 최적화
+  - 유형: 최적화
+  - 기여자: XXXXRT666
+
+## 202604
+
+- 2026.04.18 [PR#2763](https://github.com/RVC-Boss/GPT-SoVITS/pull/2763)
+  - 내용: G2PW 추론 입력 구성 및 다중 발음 처리를 최적화하여 긴 문장에 대한 중복 계산 및 추론 오버헤드 감소
+  - 유형: 최적화
+  - 기여자: baicai-1145
+- 2026.04.18 [PR#2767](https://github.com/RVC-Boss/GPT-SoVITS/pull/2767)
+  - 내용: Windows 환경 단일 GPU에서 GPT-SoVITS v3의 LoRA 학습 흐름 개선
+  - 유형: 최적화
+  - 기여자: 2409324124
+- 2026.04.18 [PR#2755](https://github.com/RVC-Boss/GPT-SoVITS/pull/2755)
+  - 내용: 여러 모듈의 잡다한 버그 수정
+  - 유형: 수정
+  - 기여자: wishhyt
+- 2026.04.18 [PR#2758](https://github.com/RVC-Boss/GPT-SoVITS/pull/2758)
+  - 내용: 데이터셋 처리를 위한 오류 처리 힌트 추가
+  - 유형: 최적화
+  - 기여자: mushroomcowisheggs
+- 2026.04.18 [PR#2753](https://github.com/RVC-Boss/GPT-SoVITS/pull/2753)
+  - 내용: 병렬 추론의 일부 버그 수정
+  - 유형: 수정
+  - 기여자: wishhyt
+- 2026.04.18 [PR#2733](https://github.com/RVC-Boss/GPT-SoVITS/pull/2733)
+  - 내용: DPO 학습이 누락 단어 시뮬레이션을 지원하지 않는 버그 수정
+  - 유형: 수정
+  - 기여자: Mr-Neutr0n
+- 2026.04.18 [Commit#02425ea](https://github.com/RVC-Boss/GPT-SoVITS/commit/02425ea25680c26c700be0bc158756c69103d827)
+  - 내용: ONNX 스크립트에서 Optional 등 누락된 임포트 문제 수정
+  - 유형: 수정
+  - 기여자: RVC-Boss

docs/tr/Changelog_TR.md

@@ -2,8 +2,6 @@
 
 ## 202401
 
-## 202401
-
 - 2024.01.21 [PR#108](https://github.com/RVC-Boss/GPT-SoVITS/pull/108)
   - İçerik: WebUI'ya İngilizce sistem çeviri desteği eklendi.
   - Tür: Dokümantasyon
@@ -332,6 +330,8 @@
   - Tür: Optimizasyon
   - Katkıda Bulunan: RVC-Boss, GoHomeToMacDonal
   - İlgili: [PR#672](https://github.com/RVC-Boss/GPT-SoVITS/pull/672)
+- Gelecek güncellemeler, `fast_inference` dalındaki değişikliklerin tutarlılığını doğrulamaya devam edecek.
+
 - 2024.07.13 [PR#1294](https://github.com/RVC-Boss/GPT-SoVITS/pull/1294), [PR#1298](https://github.com/RVC-Boss/GPT-SoVITS/pull/1298)
   - İçerik: i18n taraması yeniden düzenlendi ve çok dilli yapılandırma dosyaları güncellendi
   - Tür: Dokümantasyon
@@ -578,3 +578,160 @@
   - İçerik: Otomatik hassasiyet algılama mantığı optimize edildi; WebUI önyüz modüllerine katlanabilir özellik eklendi
   - Tür: Yeni Özellik
   - Katkıda Bulunanlar: XXXXRT666, RVC-Boss
+- 2025.06.06 [PR#2427](https://github.com/RVC-Boss/GPT-SoVITS/pull/2427)
+  - İçerik: "X一X" kalıbı için çok sesli harf tespitini düzelt
+  - Tür: Düzeltme
+  - Katkıda Bulunan: wzy3650
+- 2025.06.05 [PR#2439](https://github.com/RVC-Boss/GPT-SoVITS/pull/2439)
+  - İçerik: Yapılandırma düzeltmesi; SoVITS model yüklemesini düzelt
+  - Tür: Düzeltme
+  - Katkıda Bulunan: wzy3650
+- 2025.06.09 [Commit#8056efe4](https://github.com/RVC-Boss/GPT-SoVITS/commit/8056efe4ab7bbc3610c72ae356a6f37518441f7d)
+  - İçerik: `ge.sum` kaynaklı olası sayısal patlamayı (sessiz çıkarıma yol açan) düzelt
+  - Tür: Düzeltme
+  - Katkıda Bulunan: RVC-Boss
+- 2025.06.10 [Commit#2c0436b9](https://github.com/RVC-Boss/GPT-SoVITS/commit/2c0436b9ce397424ae03476c836fb64c6e5ebcc6)
+  - İçerik: Deney adı boşlukla bittiğinde oluşan hatalı Windows yolunu düzelt
+  - Tür: Düzeltme
+  - Katkıda Bulunan: RVC-Boss
+- 2025.06.10 [PR#2449](https://github.com/RVC-Boss/GPT-SoVITS/pull/2449)
+  - İçerik: Dil bölütlemeyi optimize et
+  - Tür: Optimizasyon
+  - Katkıda Bulunan: KamioRinn
+- 2025.06.11 [PR#2450](https://github.com/RVC-Boss/GPT-SoVITS/pull/2450)
+  - İçerik: v2pro için paralel çıkarım desteğindeki hatayı düzelt
+  - Tür: Düzeltme
+  - Katkıda Bulunan: YYuX-1145
+- 2025.06.11 [Commit#ed89a023](https://github.com/RVC-Boss/GPT-SoVITS/commit/ed89a023378dabba9d4b6580235bb9742245816d)
+  - İçerik: v2pro için `ge` çıkarımındaki sayısal taşma sorununu düzelt
+  - Tür: Düzeltme
+  - Katkıda Bulunan: RVC-Boss
+- 2025.06.17 [PR#2464](https://github.com/RVC-Boss/GPT-SoVITS/pull/2464) [PR#2482](https://github.com/RVC-Boss/GPT-SoVITS/pull/2482)
+  - İçerik: `install.sh` mantığını optimize et
+  - Tür: Optimizasyon
+  - Katkıda Bulunan: XXXXRT666
+- 2025.06.27 [PR#2489](https://github.com/RVC-Boss/GPT-SoVITS/pull/2489)
+  - İçerik: onnxruntime yükleme mantığını optimize et (GPU/CPU algılama)
+  - Tür: Optimizasyon
+  - Katkıda Bulunan: KamioRinn
+- 2025.06.27 [PR#2488](https://github.com/RVC-Boss/GPT-SoVITS/pull/2488)
+  - İçerik: Dil bölütleme ve biçimlendirmeyi optimize et
+  - Tür: Optimizasyon
+  - Katkıda Bulunan: KamioRinn
+
+## 202507
+
+- 2025.07.10 [Commit#426e1a2bb](https://github.com/RVC-Boss/GPT-SoVITS/commit/426e1a2bb43614af2479b877c37acfb0591e952f)
+  - İçerik: Çıkarım işlem önceliğini artır (Win11'de olası GPU kullanım sınırlamasını düzelt)
+  - Tür: Optimizasyon
+  - Katkıda Bulunan: XianYue0125
+- 2025.07.16 [PR#2490](https://github.com/RVC-Boss/GPT-SoVITS/pull/2490)
+  - İçerik: TTS.py'nin gerçekte desteklenen sürümler olan v2Pro ve v2ProPlus'ı tanımaması sorununu düzelt ve varsayılan yapılandırmayı güncelle
+  - Tür: Düzeltme
+  - Katkıda Bulunan: jiangsier-xyz
+- 2025.07.16 [Commit#4d8ebf85](https://github.com/RVC-Boss/GPT-SoVITS/commit/4d8ebf85233d4f1166d7cc02fdc595602975ca8f)
+  - İçerik: Paralel çıkarım modunda v2pro model tanıma sorununu düzelt
+  - Tür: Düzeltme
+  - Katkıda Bulunan: RVC-Boss
+- 2025.07.17 [PR#2531](https://github.com/RVC-Boss/GPT-SoVITS/pull/2531)
+  - İçerik: Whisper ASR daha uygun maliyetli distill modellerini destekler
+  - Tür: Optimizasyon
+  - Katkıda Bulunan: XXXXRT666
+- 2025.07.18 [PR#2536](https://github.com/RVC-Boss/GPT-SoVITS/pull/2536)
+  - İçerik: `TTS_Config` kod mantığını optimize et
+  - Tür: Optimizasyon
+  - Katkıda Bulunan: ChasonJiang
+- 2025.07.18 [PR#2537](https://github.com/RVC-Boss/GPT-SoVITS/pull/2537)
+  - İçerik: GPT kayıp (loss) hesaplama sorununu düzelt
+  - Tür: Düzeltme
+  - Katkıda Bulunan: ChasonJiang
+
+## 202508
+
+- 2025.08.02 [PR#2561](https://github.com/RVC-Boss/GPT-SoVITS/pull/2561)
+  - İçerik: WSL Rocm
+  - Tür: Düzeltme
+  - Katkıda Bulunan: XXXXRT666
+
+## 202509
+
+- 2025.09.10 [Commit#11aa78bd](https://github.com/RVC-Boss/GPT-SoVITS/commit/11aa78bd9bda8b53047cfcae03abf7ca94d27391)
+  - İçerik: Ortam değişkeninin dize (string) olmaması sorununu düzelt
+  - Tür: Düzeltme
+  - Katkıda Bulunan: RVC-Boss
+
+## 202511
+
+- 2025.11.28 [PR#2671](https://github.com/RVC-Boss/GPT-SoVITS/pull/2671) [PR#2678](https://github.com/RVC-Boss/GPT-SoVITS/pull/2678)
+  - İçerik: Akışlı çıkarım (streaming inference)
+  - Tür: Yeni Özellik
+  - Katkıda Bulunan: ChasonJiang
+- 2025.11.28 [PR#2636](https://github.com/RVC-Boss/GPT-SoVITS/pull/2636)
+  - İçerik: Matematiksel ifade metinleri için metin ön uç (frontend) mantığını optimize et
+  - Tür: Optimizasyon
+  - Katkıda Bulunan: KamioRinn
+- 2025.11.28 [PR#2469](https://github.com/RVC-Boss/GPT-SoVITS/pull/2469)
+  - İçerik: Akışlı çıkarım (streaming inference)
+  - Tür: Yeni Özellik
+  - Katkıda Bulunan: L-jasmine
+- 2025.11.28 [PR#2577](https://github.com/RVC-Boss/GPT-SoVITS/pull/2577)
+  - İçerik: VQ dağıtılmış eğitimi destekle
+  - Tür: Optimizasyon
+  - Katkıda Bulunan: wzy3650
+- 2025.11.28 [PR#2627](https://github.com/RVC-Boss/GPT-SoVITS/pull/2627) [PR#2679](https://github.com/RVC-Boss/GPT-SoVITS/pull/2679)
+  - İçerik: ASR model indirme mantığını optimize et
+  - Tür: Optimizasyon
+  - Katkıda Bulunan: XXXXRT666
+- 2025.11.28 [PR#2662](https://github.com/RVC-Boss/GPT-SoVITS/pull/2662)
+  - İçerik: Varsayılan parti boyutu (batch size) hatasını düzelt
+  - Tür: Düzeltme
+  - Katkıda Bulunan: Spr-Aachen
+
+## 202512
+
+- 2025.12.30 [PR#2703](https://github.com/RVC-Boss/GPT-SoVITS/pull/2703) [PR#2704](https://github.com/RVC-Boss/GPT-SoVITS/pull/2704)
+  - İçerik: Örnekleme (sampling) hatasını düzelt
+  - Tür: Düzeltme
+  - Katkıda Bulunan: ChasonJiang
+
+## 202602
+
+- 2026.02.08 [PR#2727](https://github.com/RVC-Boss/GPT-SoVITS/pull/2727)
+  - İçerik: Kabul edilmeyen Conda koşullarının neden olduğu derleme hatasını düzelt
+  - Tür: Düzeltme
+  - Katkıda Bulunan: Oarora
+- 2026.02.09 [PR#2732](https://github.com/RVC-Boss/GPT-SoVITS/pull/2732)
+  - İçerik: Otomatik ortam kurulumunu optimize et
+  - Tür: Optimizasyon
+  - Katkıda Bulunan: XXXXRT666
+
+# 202604
+
+- 2026.04.18 [PR#2763](https://github.com/RVC-Boss/GPT-SoVITS/pull/2763)
+  - İçerik: Uzun cümlelerde gereksiz hesaplama ve çıkarım yükünü azaltmak için G2PW çıkarım girdi oluşturmayı ve çok sesli harf işlemeyi optimize et
+  - Tür: Optimizasyon
+  - Katkıda Bulunan: baicai-1145
+- 2026.04.18 [PR#2767](https://github.com/RVC-Boss/GPT-SoVITS/pull/2767)
+  - İçerik: Windows altında tek kartta GPT-SoVITS v3 için LoRA eğitim akışını iyileştir
+  - Tür: Optimizasyon
+  - Katkıda Bulunan: 2409324124
+- 2026.04.18 [PR#2755](https://github.com/RVC-Boss/GPT-SoVITS/pull/2755)
+  - İçerik: Birden çok modüldeki çeşitli hataları düzelt
+  - Tür: Düzeltme
+  - Katkıda Bulunan: wishhyt
+- 2026.04.18 [PR#2758](https://github.com/RVC-Boss/GPT-SoVITS/pull/2758)
+  - İçerik: Veri kümesi işleme için hata işleme ipuçları ekle
+  - Tür: Optimizasyon
+  - Katkıda Bulunan: mushroomcowisheggs
+- 2026.04.18 [PR#2753](https://github.com/RVC-Boss/GPT-SoVITS/pull/2753)
+  - İçerik: Paralel çıkarımdaki bazı hataları düzelt
+  - Tür: Düzeltme
+  - Katkıda Bulunan: wishhyt
+- 2026.04.18 [PR#2733](https://github.com/RVC-Boss/GPT-SoVITS/pull/2733)
+  - İçerik: DPO eğitiminin eksik kelime simülasyonunu desteklememe hatasını düzelt
+  - Tür: Düzeltme
+  - Katkıda Bulunan: Mr-Neutr0n
+- 2026.04.18 [Commit#02425ea](https://github.com/RVC-Boss/GPT-SoVITS/commit/02425ea25680c26c700be0bc158756c69103d827)
+  - İçerik: ONNX betiğinde (Optional vb.) eksik içe aktarmaları düzelt
+  - Tür: Düzeltme
+  - Katkıda Bulunan: RVC-Boss

tools/asr/funasr_asr.py

@@ -39,6 +39,7 @@ def create_model(language="zh"):
             local_dir="tools/asr/models/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch",
         )
         model_revision = "v2.0.4"
+        vad_model_revision = punc_model_revision = "v2.0.4"
     elif language == "yue":
         path_asr = "tools/asr/models/speech_UniASR_asr_2pass-cantonese-CHS-16k-common-vocab1468-tensorflow1-online"
         snapshot_download(
@@ -51,8 +52,6 @@ def create_model(language="zh"):
     else:
         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]
     else:

tools/uvr5/lib/lib_v5/spec_utils.py

@@ -485,6 +485,8 @@ def istft(spec, hl):
     wave_right = librosa.istft(spec_right, hop_length=hl)
     wave = np.asfortranarray([wave_left, wave_right])
 
+    return wave
+
 
 if __name__ == "__main__":
     import argparse