Update Changelog_CN.md

Fixed issues such as missing imports for types like `Optional`.

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)
-        assert os.path.exists(self.path6)
+        if not os.path.exists(self.path2):
+            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,
-    num_heads: int,
+    embed_dim_to_check,
+    num_heads,
     in_proj_weight,
-    in_proj_bias: Optional[Tensor],
-    bias_k: Optional[Tensor],
-    bias_v: Optional[Tensor],
-    add_zero_attn: bool,
-    dropout_p: float,
-    out_proj_weight: Tensor,
-    out_proj_bias: Optional[Tensor],
-    training: bool = True,
-    key_padding_mask: Optional[Tensor] = None,
-    need_weights: bool = True,
-    attn_mask: Optional[Tensor] = None,
-    use_separate_proj_weight: bool = False,
-    q_proj_weight: Optional[Tensor] = None,
-    k_proj_weight: Optional[Tensor] = None,
-    v_proj_weight: Optional[Tensor] = None,
-    static_k: Optional[Tensor] = None,
-    static_v: Optional[Tensor] = None,
-    average_attn_weights: bool = True,
-    is_causal: bool = False,
+    in_proj_bias,
+    bias_k,
+    bias_v,
+    add_zero_attn,
+    dropout_p,
+    out_proj_weight,
+    out_proj_bias,
+    training=True,
+    key_padding_mask=None,
+    need_weights=True,
+    attn_mask=None,
+    use_separate_proj_weight=False,
+    q_proj_weight=None,
+    k_proj_weight=None,
+    v_proj_weight=None,
+    static_k=None,
+    static_v=None,
+    average_attn_weights=True,
+    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(
-        backend="gloo" if os.name == "nt" or not torch.cuda.is_available() else "nccl",
-        init_method="env://?use_libuv=False",
-        world_size=n_gpus,
-        rank=rank,
-    )
+    use_ddp = n_gpus > 1
+    if use_ddp:
+        dist.init_process_group(
+            backend="gloo" if os.name == "nt" or not torch.cuda.is_available() else "nccl",
+            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(
-        train_dataset,
-        num_workers=5,
+    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)
+    loader_kwargs = dict(
+        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
-            pinyins = g2pw.lazy_pinyin(seg, neutral_tone_with_five=True, style=Style.TONE3)
+            # g2pw采用整句推理（批量推理，逐句取结果）
+            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:
-            tokens, text2token, token2text = tokenize_and_map(tokenizer=tokenizer, text=text)
-        except Exception:
-            print(f'warning: text "{text}" is invalid')
-            return {}
+        cached = tokenized_cache.get(text)
+        if cached is None:
+            try:
+                tokens, text2token, token2text = tokenize_and_map(tokenizer=tokenizer, text=text)
+            except Exception:
+                print(f'warning: text "{text}" is invalid')
+                return {}
 
-        text, query_id, tokens, text2token, token2text = _truncate(
-            max_len=max_len, text=text, query_id=query_id, tokens=tokens, text2token=text2token, token2text=token2text
-        )
+            if len(tokens) <= max_len - 2:
+                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)))
-        token_type_id = list(np.zeros((len(processed_tokens),), dtype=int))
-        attention_mask = list(np.ones((len(processed_tokens),), dtype=int))
-
-        query_char = text[query_id]
-        phoneme_mask = (
-            [1 if i in char2phonemes[query_char] else 0 for i in range(len(labels))] if use_mask else [1] * len(labels)
-        )
-        char_id = chars.index(query_char)
-        position_id = text2token[query_id] + 1  # [CLS] token locate at first place
+        query_char = text_for_query[query_id_for_query]
+        if use_mask:
+            phoneme_mask = char_phoneme_masks[query_char]
+        else:
+            phoneme_mask = full_phoneme_mask
+        char_id = char2id[query_char]
+        position_id = text2token_for_query[query_id_for_query] + 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),
-        "token_type_ids": np.array(token_type_ids).astype(np.int64),
-        "attention_masks": np.array(attention_masks).astype(np.int64),
+        "input_ids": np.array(_pad_sequences(input_ids, pad_value=0)).astype(np.int64),
+        "token_type_ids": np.array(_pad_sequences(token_type_ids, pad_value=0)).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)
-
-        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_dir = download_and_decompress(model_dir)
+        self.config = load_config(config_path=os.path.join(self.model_dir, "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")
-        monophonic_chars_path = os.path.join(uncompress_path, "MONOPHONIC_CHARS.txt")
+        polyphonic_chars_path = os.path.join(self.model_dir, "POLYPHONIC_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.remove(char)
+            self.polyphonic_chars_new.discard(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)
+            self.monophonic_chars_dict.pop(char, None)
 
-        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')
-            return None
+        print(f'Warning: "{bopomofo}" cannot convert to pinyin')
+        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]]]:
-        texts, query_ids, sent_ids, partial_results = [], [], [], []
+    def _prepare_data(
+        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)
-                    query_ids.append(i)
-                    sent_ids.append(sent_id)
+                    polyphonic_indices.append(i)
                 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)

docs/cn/Changelog_CN.md

@@ -606,7 +606,7 @@
   - 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)
+- 2025.06.11 [Commit#37f5abfc](https://github.com/RVC-Boss/GPT-SoVITS/commit/6fdc67ca83418306f11e90b9139278313ac5c3e9) [Commit#6fdc67ca](https://github.com/RVC-Boss/GPT-SoVITS/commit/37f5abfcb4a6553652235909db2e124b6f8ff3a5)
   - 内容: install.sh逻辑优化
   - 类型: 优化
   - 提交: XXXXRT666
@@ -618,9 +618,102 @@
   - 内容: 语言分割及格式化优化
   - 类型: 优化
   - 提交: KamioRinn
+
+## after 202506
 - 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 [Commit#e476b01f3](https://github.com/RVC-Boss/GPT-SoVITS/commit/e476b01f30312139555d45a78cbd830f557d892c)
+  - 内容: 解决 TTS.py 无法识别真正支持版本 v2Pro、v2ProPlus 的问题 (#2490)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: 
+- 2025.07.17 [Commit#cefafee3](https://github.com/RVC-Boss/GPT-SoVITS/commit/cefafee32cfc08f0f622ef460578b09485cc189e)
+  - 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 [Commit#b9211657](https://github.com/RVC-Boss/GPT-SoVITS/commit/b9211657d8dfe8cd46f6b6eb9cfc55d5989e6548)
+  - 内容: 优化TTS_Config的代码逻辑 (#2536)
+  - 类型: 优化
+  - 提交: ChasonJiang
+- 2025.07.18 [Commit#b5a67e62](https://github.com/RVC-Boss/GPT-SoVITS/commit/b5a67e62470fb87e7cea28ecad7c7c3bf7a58afd)
+  - 内容: 修复gpt的loss计算问题 (#2537)
+  - 类型: 修复
+  - 提交: ChasonJiang
+- 2025.08.02 [Commit#fdf794e3](https://github.com/RVC-Boss/GPT-SoVITS/commit/fdf794e31d1fd6f91c5cb4fbb0396094491a31ac)
+  - 内容: WSL Rocm (#2561)
+  - 类型: 修复
+  - 提交: XXXXRT666
+- 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
+- 2025.11.28 [Commit#92ab59c5](https://github.com/RVC-Boss/GPT-SoVITS/commit/92ab59c5533a5dea368ddb8dad89e14474307145) [Commit#6fb441f](https://github.com/RVC-Boss/GPT-SoVITS/commit/6fb441f65e4b0573d7f7b16d96dc1917d38eda64)
+  - 内容: 流式推理
+  - 类型: 新功能
+  - 提交: ChasonJiang
+- 2025.11.28 [Commit#e00ca921](https://github.com/RVC-Boss/GPT-SoVITS/commit/e00ca92140542e6d947b9f660e24ed757aabc793)
+  - 内容: 数学计算文本前端逻辑优化
+  - 类型: 优化
+  - 提交: KamioRinn
+- 2025.11.28 [Commit#11aa78bd](https://github.com/RVC-Boss/GPT-SoVITS/commit/11aa78bd9bda8b53047cfcae03abf7ca94d27391)
+  - 内容: 流式推理 (#2469)
+  - 类型: 新功能
+  - 提交: L-jasmine
+- 2025.11.28 [Commit#60a4a214](https://github.com/RVC-Boss/GPT-SoVITS/commit/60a4a214aff18057bb4ce76643d3b85de4bb67a4)
+  - 内容: 支持vq分布式训练 (#2577)
+  - 类型: 优化
+  - 提交: wzy3650
+- 2025.11.28 [Commit#c85c54ec](https://github.com/RVC-Boss/GPT-SoVITS/commit/c85c54eca99a2fd01d6b574584217d0ecfbd90c1) [Commit#8577992](https://github.com/RVC-Boss/GPT-SoVITS/commit/857799276c3e8adcda7d662a55b07bf00bc1f01b)
+  - 内容: ASR模型下载逻辑优化
+  - 类型: 优化
+  - 提交: XXXXRT666
+- 2025.11.28 [Commit#92d2d33](https://github.com/RVC-Boss/GPT-SoVITS/commit/92d2d337fd98673c126fd40727e067204e4523ae)
+  - 内容: default batch size bug修复
+  - 类型: 修复
+  - 提交: Spr-Aachen
+- 2025.12.30 [Commit#9080a96](https://github.com/RVC-Boss/GPT-SoVITS/commit/9080a967d5e64f4bfb5a9ea33afc7252136b0256) [Commit#c767f0b](https://github.com/RVC-Boss/GPT-SoVITS/commit/c767f0b83b998e996a4d230d86da575a03f54a3f)
+  - 内容: 修复采样错误
+  - 类型: 修复
+  - 提交: ChasonJiang
+- 2026.02.08 [Commit#9986880](https://github.com/RVC-Boss/GPT-SoVITS/commit/9986880b3f13b3076989db17cc1a7227aa0186c9)
+  - 内容: 修复Conda 条款未同意导致的构建失败 (#2727)
+  - 类型: 修复
+  - 提交: Oarora
+- 2026.02.09 [Commit#2d9193b](https://github.com/RVC-Boss/GPT-SoVITS/commit/2d9193b0d3c0eae0c3a14d8c68a839f1bae157dc)
+  - 内容: 环境自动构建优化 (#2732)
+  - 类型: 优化
+  - 提交: XXXXRT666
+- 2026.04.18 [Commit#ba8de9b](https://github.com/RVC-Boss/GPT-SoVITS/commit/ba8de9b760a4bd5b3eb827a594416e71b44f2510)
+  - 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 (#2763)
+  - 类型: 优化
+  - 提交: baicai-1145
+- 2026.04.18 [Commit#780383d](https://github.com/RVC-Boss/GPT-SoVITS/commit/780383d5bd0d09a4f132b5ab1e80c04c9606b48a)
+  - 内容: 改进 Windows 单卡 v3 LoRA 训练流程 (#2767)
+  - 类型: 优化
+  - 提交: 2409324124
+- 2026.04.18 [Commit#1419190](https://github.com/RVC-Boss/GPT-SoVITS/commit/14191901cdb7e791d8fee1ff31dffe107f9e28fb)
+  - 内容: 修复多个模块中的独立 bug (#2755)
+  - 类型: 修复
+  - 提交: wishhyt
+- 2026.04.18 [Commit#00ce973](https://github.com/RVC-Boss/GPT-SoVITS/commit/00ce973412384e92a44836f168de2a9a8827259c)
+  - 内容: 添加数据集的错误处理提示 (#2758)
+  - 类型: 优化
+  - 提交: mushroomcowisheggs
+- 2026.04.18 [Commit#445d18c](https://github.com/RVC-Boss/GPT-SoVITS/commit/445d18ccce0b4ea7cb6f8c93ff688b662bc61338)
+  - 内容: 并行推理部分bug修复 (#2753)
+  - 类型: 修复
+  - 提交: wishhyt
+- 2026.04.18 [Commit#938f05f](https://github.com/RVC-Boss/GPT-SoVITS/commit/938f05fce8bcfb2407b8311fbbc10ac4d9ffe1c0)
+  - 内容: bug修复：dpo训练不支持漏字模拟 (#2733)
+  - 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
 

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