Merge 7604f36bb270d0a897df0b8d4dd9d35f860d06cb 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)
-        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/inference_webui.py

@@ -27,11 +27,14 @@ import re
 import sys
 import traceback
 import warnings
+import soundfile  # 新增导入
 
 import torch
 import torchaudio
 from text.LangSegmenter import LangSegmenter
 
+
+
 logging.getLogger("markdown_it").setLevel(logging.ERROR)
 logging.getLogger("urllib3").setLevel(logging.ERROR)
 logging.getLogger("httpcore").setLevel(logging.ERROR)
@@ -1001,6 +1004,255 @@ def get_tts_wav(
     yield opt_sr, (audio_opt * 32767).astype(np.int16)
 
 
+import uuid
+import shutil
+from pydub import AudioSegment
+
+TEMP_FOLDER = "TEMP"  # 临时文件夹路径
+os.makedirs(TEMP_FOLDER, exist_ok=True)
+
+def clean_temp_folder_on_startup():
+    """启动时清理临时文件夹"""
+    try:
+        if os.path.exists(TEMP_FOLDER):
+            shutil.rmtree(TEMP_FOLDER)
+            os.makedirs(TEMP_FOLDER, exist_ok=True)
+            print("启动时已清理临时文件夹")
+        else:
+            os.makedirs(TEMP_FOLDER, exist_ok=True)
+            print("临时文件夹已创建")
+    except Exception as e:
+        print(f"启动时清理临时文件夹失败: {str(e)}")
+
+def split_text_by_punctuation(text, period_pause=0.3, comma_pause=0.15):
+    """改进的文本分割函数"""
+    if not text or not isinstance(text, str):
+        print("收到空或非字符串文本输入")
+        return []
+    
+    segments = []
+    current_segment = ""
+    punctuation_marks = ['，', ',', '。', '.']
+    
+    for char in text:
+        current_segment += char
+        if char in punctuation_marks:
+            # 根据标点类型设置停顿时间
+            pause = period_pause if char in ['。', '.'] else comma_pause
+            segments.append({
+                "text": current_segment.strip(),
+                "pause": pause,
+                "punctuation": char
+            })
+            current_segment = ""
+    
+    if current_segment:
+        segments.append({
+            "text": current_segment.strip(),
+            "pause": comma_pause,  # 默认使用非句号停顿时间
+            "punctuation": ""
+        })
+    
+    print(f"分割结果: {[seg['text'] for seg in segments]}")
+    return segments
+
+def generate_segment_audio(segment_data, ref_wav_path, prompt_text, prompt_language, text_language, top_k, top_p, temperature):
+    """增强的音频生成函数"""
+    try:
+        if not os.path.exists(ref_wav_path):
+            raise FileNotFoundError(f"参考音频不存在: {ref_wav_path}")
+        
+        # 生成音频
+        sr, audio_data = next(get_tts_wav(
+            ref_wav_path=ref_wav_path,
+            prompt_text=prompt_text,
+            prompt_language=prompt_language,
+            text=segment_data["text"],
+            text_language=text_language,
+            how_to_cut=i18n("不切"),
+            top_k=top_k,
+            top_p=top_p,
+            temperature=temperature,
+            pause_second=0  # 不在内部添加停顿
+        ))
+        
+        # 这里不再生成ID，由调用者提供
+        temp_path = os.path.join(TEMP_FOLDER, "temp_generate.wav")
+        soundfile.write(temp_path, audio_data, sr)
+        
+        # 添加停顿
+        audio = AudioSegment.from_wav(temp_path)
+        pause = AudioSegment.silent(duration=int(segment_data["pause"]*1000))
+        final_audio = audio + pause
+        final_audio.export(temp_path, format="wav")
+        
+        return {
+            "success": True,
+            "audio_path": temp_path,
+            "text": segment_data["text"],
+            "pause": segment_data["pause"],
+            "message": "生成成功"
+        }
+    except Exception as e:
+        print(f"生成片段失败: {str(e)}", exc_info=True)
+        return {
+            "success": False,
+            "audio_path": None,
+            "text": segment_data["text"],
+            "pause": segment_data["pause"],
+            "message": f"生成失败: {str(e)}"
+        }
+
+def process_all_segments(text, ref_wav_path, prompt_text, prompt_language, text_language, 
+                        top_k, top_p, temperature, period_pause, comma_pause):
+    """完整处理流程"""
+    # 输入验证
+    if not text or not isinstance(text, str):
+        error_msg = "输入文本无效"
+        print(error_msg)
+        return [[1, error_msg, "错误"]], None
+    
+    if not os.path.exists(ref_wav_path):
+        error_msg = f"参考音频不存在: {ref_wav_path}"
+        print(error_msg)
+        return [[1, error_msg, "错误"]], None
+    
+    # 处理分段
+    segments = split_text_by_punctuation(text, period_pause, comma_pause)
+    if not segments:
+        error_msg = "无法分割文本"
+        print(error_msg)
+        return [[1, error_msg, "错误"]], None
+    
+    results = []
+    audio_files = []
+    
+    # 修改这里：使用enumerate从1开始编号，而不是基于文件夹内容
+    for i, segment in enumerate(segments, 1):
+        result = generate_segment_audio(
+            segment, ref_wav_path, prompt_text,
+            prompt_language, text_language, top_k, top_p, temperature
+        )
+        
+        # 更新结果中的segment_id
+        result["segment_id"] = f"{i}temp"
+        if result["success"] and result["audio_path"]:
+            # 重命名文件以匹配新的编号
+            new_path = os.path.join(TEMP_FOLDER, f"{result['segment_id']}.wav")
+            os.rename(result["audio_path"], new_path)
+            result["audio_path"] = new_path
+            audio_files.append(new_path)
+        
+        results.append(result)
+        print(f"处理进度: {i}/{len(segments)} - {result['message']}")
+
+    # 准备显示数据
+    df_data = []
+    for i, result in enumerate(results, 1):
+        df_data.append([
+            f"{i}temp",
+            result["text"],
+            result["message"]
+        ])
+    
+    first_audio = audio_files[0] if audio_files else None
+    return df_data, first_audio
+
+def regenerate_segment(segment_id, new_text, ref_wav_path, prompt_text, 
+                      prompt_language, text_language, top_k, top_p, temperature,
+                      period_pause, comma_pause):
+    try:
+        if not segment_id or not new_text:
+            raise ValueError("缺少片段ID或新文本内容")
+        
+        # 从文件名解析原始停顿时间
+        try:
+            pause = 0.25 if segment_id.endswith(("。", ".")) else 0.1
+        except:
+            pause = 0.1  # 默认值
+        
+        is_period = segment_id.endswith(("。", "."))
+        pause = period_pause if is_period else comma_pause
+        
+        segment_data = {
+            "text": new_text,
+            "pause": pause,
+            "punctuation": "。" if is_period else "，"
+        }
+        result = generate_segment_audio(
+            segment_data, ref_wav_path, prompt_text,
+            prompt_language, text_language, top_k, top_p, temperature
+        )
+        
+        # 更新文件
+        if result["success"]:
+            old_path = os.path.join(TEMP_FOLDER, f"{segment_id}.wav")
+            if os.path.exists(old_path):
+                os.remove(old_path)
+            os.rename(result["audio_path"], old_path)
+            result["audio_path"] = old_path
+        
+        return (
+            result["audio_path"],
+            segment_id,
+            result["message"]
+        )
+    except Exception as e:
+        print(f"重新生成片段失败: {str(e)}", exc_info=True)
+        return None, segment_id, f"重新生成失败: {str(e)}"
+
+def merge_all_segments():
+    try:
+        # 获取并按编号排序片段
+        segments = sorted(
+            [f for f in os.listdir(TEMP_FOLDER) if f.endswith(".wav") and f != "final_output.wav"],
+            key=lambda x: int(x.split("temp")[0])
+        )
+        if not segments:
+            raise ValueError("没有找到可合并的音频片段")
+        
+        combined = AudioSegment.empty()
+        for seg in segments:
+            seg_path = os.path.join(TEMP_FOLDER, seg)
+            audio = AudioSegment.from_wav(seg_path)
+            combined += audio
+        
+        # 保存最终结果
+        output_path = os.path.join(TEMP_FOLDER, "final_output.wav")
+        combined.export(output_path, format="wav")
+        
+        print(f"成功合并 {len(segments)} 个片段")
+        return output_path, "合并成功"
+    except Exception as e:
+        print(f"合并片段失败: {str(e)}", exc_info=True)
+        return None, f"合并失败: {str(e)}"
+
+def clean_temp_files():
+    """清理临时文件函数"""
+    try:
+        for filename in os.listdir(TEMP_FOLDER):
+            file_path = os.path.join(TEMP_FOLDER, filename)
+            try:
+                if os.path.isfile(file_path):
+                    os.unlink(file_path)
+            except Exception as e:
+                print(f"删除文件 {file_path} 失败: {e}")
+        return "临时文件已清理"
+    except Exception as e:
+        return f"清理失败: {str(e)}"
+
+def on_segment_select(df, evt: gr.SelectData):
+    """当选择分段列表中的项目时更新显示"""
+    if evt.index:
+        selected_row = df.iloc[evt.index[0]]
+        audio_path = os.path.join(TEMP_FOLDER, f"{selected_row['编号']}.wav")
+        return (
+            selected_row["编号"],
+            selected_row["文本内容"],
+            audio_path if os.path.exists(audio_path) else None
+        )
+    return "1temp", "", None
+
 def split(todo_text):
     todo_text = todo_text.replace("……", "。").replace("——", "，")
     if todo_text[-1] not in splits:
@@ -1018,6 +1270,236 @@ def split(todo_text):
         else:
             i_split_head += 1
     return todo_texts
+# ======================== 合并功能实现 ========================
+def merge_selected_segments(merge_range, segment_list_data, ref_wav_path, prompt_text, 
+                           prompt_language, text_language, top_k, top_p, temperature,
+                           pause_period, pause_comma):
+    """合并选中的句子并立即生成新音频"""
+    try:
+        if not merge_range:
+            return segment_list_data, "请输入合并范围（例如：1-3）", None
+        
+        # 解析合并范围
+        start, end = map(int, merge_range.split('-'))
+        if start <= 0 or end <= 0 or start > end:
+            return segment_list_data, "无效的合并范围", None
+        
+        # 检查范围是否有效
+        if end > len(segment_list_data):
+            return segment_list_data, f"结束编号 {end} 超过总段数 {len(segment_list_data)}", None
+        
+        # === 第一步：收集要删除的文件并立即删除 ===
+        files_to_delete = []
+        for i in range(start-1, end):
+            file_id = segment_list_data.iloc[i, 0]
+            file_path = os.path.join(TEMP_FOLDER, f"{file_id}.wav")
+            if os.path.exists(file_path):
+                os.remove(file_path)
+                files_to_delete.append(file_id)
+        
+        # === 第二步：合并文本 ===
+        merged_text = ""
+        for i in range(start-1, end):
+            merged_text += segment_list_data.iloc[i, 1]  # 文本内容在第二列
+        
+        # 确定合并后的停顿类型（取最后一个片段的标点）
+        last_punctuation = segment_list_data.iloc[end-1, 1][-1] if segment_list_data.iloc[end-1, 1] else ""
+        is_period = last_punctuation in ["。", "."]
+        pause = pause_period if is_period else pause_comma
+        
+        # === 第三步：立即生成合并后的音频 ===
+        segment_data = {
+            "text": merged_text,
+            "pause": pause,
+            "punctuation": last_punctuation
+        }
+        # 生成音频
+        result = generate_segment_audio(
+            segment_data, ref_wav_path, prompt_text,
+            prompt_language, text_language, top_k, top_p, temperature
+        )
+        if not result["success"]:
+            return segment_list_data, result["message"], None
+        
+        # === 第四步：构建新的分段列表 ===
+        # 创建新的合并条目 - 使用起始编号
+        new_id = f"{start}temp"
+        merged_entry = [new_id, merged_text, "已生成"]
+        
+        # 移动生成的音频到正确位置
+        new_path = os.path.join(TEMP_FOLDER, f"{new_id}.wav")
+        shutil.move(result["audio_path"], new_path)
+        
+        # 构建新的分段列表
+        new_segment_list = []
+        
+        # 添加合并前的部分
+        if start > 1:
+            new_segment_list.extend(segment_list_data.iloc[:start-1].values.tolist())
+        
+        # 添加合并条目
+        new_segment_list.append(merged_entry)
+        
+        # 添加合并后的部分
+        if end < len(segment_list_data):
+            new_segment_list.extend(segment_list_data.iloc[end:].values.tolist())
+        
+        # === 第五步：重新编号 ===
+        reindexed_list = []
+        new_id_counter = 1
+        for segment in new_segment_list:
+            old_id = segment[0]
+            # 为新列表生成连续编号
+            new_id = f"{new_id_counter}temp"
+            
+            # 重命名文件（如果存在）
+            old_path = os.path.join(TEMP_FOLDER, f"{old_id}.wav")
+            new_path = os.path.join(TEMP_FOLDER, f"{new_id}.wav")
+            
+            if os.path.exists(old_path) and old_id != new_id:
+                os.rename(old_path, new_path)
+            
+            # 更新ID
+            segment[0] = new_id
+            reindexed_list.append(segment)
+            new_id_counter += 1
+        
+        return reindexed_list, "合并成功", new_id
+    
+    except Exception as e:
+        traceback.print_exc()
+        return segment_list_data, f"合并失败: {str(e)}", None
+
+# ======================== 实现拆分逻辑函数 ========================
+def split_selected_segment(split_id, segment_list_data, ref_wav_path, prompt_text, 
+                         prompt_language, text_language, top_k, top_p, temperature,
+                         pause_period, pause_comma):
+    """拆分选中的句子并重新生成 - 使用倒序重命名避免冲突"""
+    try:
+        if not split_id:
+            return segment_list_data, "请输入要拆分的句子编号", None
+        
+        # 从文件名解析原始停顿时间
+        try:
+            pause = 0.25 if split_id.endswith(("。", ".")) else 0.1
+        except:
+            pause = 0.1  # 默认值
+        
+        # 查找要拆分的句子
+        df = segment_list_data
+        target_row = None
+        target_idx = None
+        
+        # 查找匹配的句子
+        for idx, row in enumerate(df.itertuples()):
+            if str(row[1]) == split_id:  # 第一列是ID
+                target_row = row
+                target_idx = idx
+                break
+        
+        if target_row is None:
+            return segment_list_data, f"未找到编号为 {split_id} 的句子", None
+        
+        # 获取原始文本
+        original_text = target_row[2]  # 第二列是文本
+        
+        # 根据标点符号拆分文本
+        segments = []
+        current_segment = ""
+        punctuation_marks = ['，', ',', '。', '.', '？', '?', '！', '!', '；', ';']
+        
+        for char in original_text:
+            current_segment += char
+            if char in punctuation_marks:
+                # 根据标点类型设置停顿时间
+                pause = pause_period if char in ['。', '.', '？', '?', '！', '!'] else pause_comma
+                segments.append({
+                    "text": current_segment.strip(),
+                    "pause": pause,
+                    "punctuation": char
+                })
+                current_segment = ""
+        
+        if current_segment:
+            segments.append({
+                "text": current_segment.strip(),
+                "pause": pause_comma,  # 默认使用非句号停顿时间
+                "punctuation": ""
+            })
+        
+        if len(segments) <= 1:
+            return segment_list_data, "句子无法拆分（没有标点符号）", None
+        
+        # 计算拆分后新增的句子数量
+        num_new_segments = len(segments)
+        offset = num_new_segments - 1  # 拆分后增加的句子数
+        
+        # === 第一步：删除原始音频文件 ===
+        original_path = os.path.join(TEMP_FOLDER, f"{split_id}.wav")
+        if os.path.exists(original_path):
+            os.remove(original_path)
+        
+        # === 第二步：倒序重命名后续句子避免冲突 ===
+        new_segment_list = []
+        
+        # 添加拆分前的句子
+        for i in range(target_idx):
+            new_segment_list.append(df.iloc[i].tolist())
+        
+        # 倒序重命名从最后一个句子开始，避免冲突
+        total_segments = len(df)
+        
+        # 从最后一个句子开始倒序遍历
+        for i in range(total_segments - 1, target_idx, -1):
+            old_id = df.iloc[i, 0]
+            old_num = int(old_id.replace("temp", ""))
+            new_id_num = old_num + offset
+            new_id = f"{new_id_num}temp"
+            
+            # 重命名音频文件
+            old_path = os.path.join(TEMP_FOLDER, f"{old_id}.wav")
+            new_path = os.path.join(TEMP_FOLDER, f"{new_id}.wav")
+            if os.path.exists(old_path):
+                os.rename(old_path, new_path)
+            
+            # 更新列表条目
+            new_segment_list.append([new_id, df.iloc[i, 1], df.iloc[i, 2]])
+        
+        # === 第三步：添加拆分后的新句子 ===
+        for i, segment in enumerate(segments):
+            new_id = f"{target_idx+1+i}temp"  # 新ID从原始位置开始
+            new_segment_list.append([new_id, segment["text"], "待生成"])
+        
+        # === 第四步：重新排序列表 ===
+        # 按照ID中的数字排序
+        new_segment_list.sort(key=lambda x: int(x[0].replace("temp", "")))
+        
+        # === 第五步：生成拆分后的新句子 ===
+        for i, segment in enumerate(segments):
+            segment_id = f"{target_idx+1+i}temp"
+            result = generate_segment_audio(
+                segment, ref_wav_path, prompt_text,
+                prompt_language, text_language, top_k, top_p, temperature
+            )
+            
+            if result["success"]:
+                # 更新文件
+                new_path = os.path.join(TEMP_FOLDER, f"{segment_id}.wav")
+                if os.path.exists(result["audio_path"]):
+                    os.rename(result["audio_path"], new_path)
+                
+                # 更新状态
+                for j, item in enumerate(new_segment_list):
+                    if item[0] == segment_id:
+                        new_segment_list[j][2] = "已生成"
+                        break
+        
+        return new_segment_list, f"成功拆分为 {num_new_segments} 个句子", f"{target_idx+1}temp"
+    
+    except Exception as e:
+        traceback.print_exc()
+        return segment_list_data, f"拆分失败: {str(e)}", None
+
 
 
 def cut1(inp):
@@ -1327,6 +1809,160 @@ with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False, js=js, css=css
         )
         GPT_dropdown.change(change_gpt_weights, [GPT_dropdown], [])
 
+
+
+        # ======================== 插入分段合成UI开始 ========================
+    with gr.Tab(i18n("分段合成模式")):
+        with gr.Row():
+            with gr.Column():
+                segmented_text = gr.Textbox(label=i18n("需要分段的文本"), lines=10, value="")
+                segment_button = gr.Button(i18n("分割文本并生成所有片段"), variant="primary")
+                segmented_output = gr.Audio(label=i18n("当前选中片段"), interactive=False)
+                
+                with gr.Row():
+                    segment_index = gr.Textbox(label=i18n("片段编号"), interactive=False, visible=False)
+                    new_segment_text = gr.Textbox(label=i18n("新文本内容"), lines=2, max_lines=4)
+                    regenerate_button = gr.Button(i18n("重新生成当前片段"), variant="primary")
+                with gr.Row():
+                    pause_period = gr.Slider(
+                        minimum=0.1, 
+                        maximum=1.0, 
+                        step=0.05, 
+                        label=i18n("句号停顿时间(秒)"), 
+                        value=0.3,
+                        interactive=True
+                    )
+                    pause_comma = gr.Slider(
+                        minimum=0.1, 
+                        maximum=0.5, 
+                        step=0.05, 
+                        label=i18n("非句号停顿时间(秒)"), 
+                        value=0.15,
+                        interactive=True
+                    )
+                with gr.Row():
+                    clean_button = gr.Button(i18n("清理临时文件"), variant="secondary", scale=1)
+                    confirm_button = gr.Button(i18n("确认并合并所有片段"), variant="primary", scale=2)
+                
+                final_output = gr.Audio(label=i18n("最终合成结果"), interactive=False)
+                segment_status = gr.Textbox(label=i18n("状态"), interactive=False)
+                # 添加合并功能
+                with gr.Row():
+                    merge_range = gr.Textbox(label=i18n("合并范围（例如：1-3）"), scale=3)
+                    merge_button = gr.Button(i18n("合并句子"), variant="primary", scale=1)
+                # 添加拆分功能控件
+                with gr.Row():
+                    split_index = gr.Textbox(label=i18n("要拆分的句子编号"), scale=1)
+                    split_button = gr.Button(i18n("拆分句子"), variant="primary", scale=1)
+            with gr.Column():
+                segment_list = gr.Dataframe(
+                    headers=["编号", "文本内容", "状态"],
+                    datatype=["str", "str", "str"],
+                    interactive=False,
+                    label=i18n("分段列表"),
+                    value=[]
+                )
+                # 在分段合成UI部分添加以下代码（大约在line 3200附近）
+
+    # ======================== 插入分段合成UI结束 ========================
+
+    # ======================== 插入分段合成事件绑定开始 ========================
+    # 分割文本并生成所有片段
+    segment_button.click(
+        process_all_segments,
+        inputs=[
+            segmented_text, 
+            inp_ref, 
+            prompt_text, 
+            prompt_language, 
+            text_language, 
+            top_k, 
+            top_p, 
+            temperature,
+            pause_period,  # 新增句号停顿参数
+            pause_comma    # 新增非句号停顿参数
+        ],
+        outputs=[segment_list, segmented_output]
+    )
+    # 重新生成当前片段
+    regenerate_button.click(
+        regenerate_segment,
+        inputs=[
+            segment_index,  # 第一个参数应该是segment_id
+            new_segment_text, 
+            inp_ref, 
+            prompt_text, 
+            prompt_language, 
+            text_language, 
+            top_k, 
+            top_p, 
+            temperature,
+            pause_period,
+            pause_comma
+        ],
+        outputs=[segmented_output, segment_index, segment_status]
+    )
+
+    # 合并所有片段
+    confirm_button.click(
+        merge_all_segments,
+        inputs=[],
+        outputs=[final_output, segment_status]
+    )
+
+    # 清理临时文件
+    clean_button.click(
+        clean_temp_files,
+        inputs=[],
+        outputs=[segment_status]
+    )
+
+    # 当选择分段列表中的项目时
+    segment_list.select(
+        fn=on_segment_select,
+        inputs=[segment_list],
+        outputs=[segment_index, new_segment_text, segmented_output],
+    )
+    merge_button.click(
+        fn=merge_selected_segments,
+        inputs=[
+            merge_range, 
+            segment_list,
+            inp_ref,
+            prompt_text,
+            prompt_language,
+            text_language,
+            top_k,
+            top_p,
+            temperature,
+            pause_period,
+            pause_comma
+        ],
+        outputs=[segment_list, segment_status, segment_index]
+    )
+    # ======================== 添加拆分按钮的事件绑定 ========================
+    split_button.click(
+        fn=split_selected_segment,
+        inputs=[
+            split_index, 
+            segment_list,
+            inp_ref,
+            prompt_text,
+            prompt_language,
+            text_language,
+            top_k,
+            top_p,
+            temperature,
+            pause_period,
+            pause_comma
+        ],
+        outputs=[segment_list, segment_status, segment_index]
+    )
+    # ======================== 插入分段合成事件绑定结束 ========================
+
+
+
+
         # gr.Markdown(value=i18n("文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。"))
         # with gr.Row():
         #     text_inp = gr.Textbox(label=i18n("需要合成的切分前文本"), value="")
@@ -1351,3 +1987,4 @@ if __name__ == "__main__":
         server_port=infer_ttswebui,
         # quiet=True,
     )
+

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)

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

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

webui.py

@@ -1,3 +1,6 @@
+import os
+os.environ["CUDA_VISIBLE_DEVICES"] = "0"  # 限制为单卡
+
 import os
 import sys
 
@@ -123,13 +126,13 @@ def set_default():
     if version not in v3v4set:
         default_sovits_epoch = 8
         default_sovits_save_every_epoch = 4
-        max_sovits_epoch = 25  # 40
-        max_sovits_save_every_epoch = 25  # 10
+        max_sovits_epoch = 255  # 40
+        max_sovits_save_every_epoch = 255  # 10
     else:
         default_sovits_epoch = 2
         default_sovits_save_every_epoch = 1
-        max_sovits_epoch = 16  # 40 # 3 #训太多=作死
-        max_sovits_save_every_epoch = 10  # 10 # 3
+        max_sovits_epoch = 255  # 40 # 3 #训太多=作死
+        max_sovits_save_every_epoch = 255  # 10 # 3
 
     default_batch_size = max(1, default_batch_size)
     default_batch_size_s1 = max(1, default_batch_size_s1)
@@ -503,7 +506,7 @@ def open1Ba(
 ):
     global p_train_SoVITS
     if p_train_SoVITS == None:
-        exp_name = exp_name.rstrip(" ")
+        exp_name=exp_name.rstrip(" ")
         config_file = (
             "GPT_SoVITS/configs/s2.json"
             if version not in {"v2Pro", "v2ProPlus"}
@@ -600,7 +603,7 @@ def open1Bb(
 ):
     global p_train_GPT
     if p_train_GPT == None:
-        exp_name = exp_name.rstrip(" ")
+        exp_name=exp_name.rstrip(" ")
         with open(
             "GPT_SoVITS/configs/s1longer.yaml" if version == "v1" else "GPT_SoVITS/configs/s1longer-v2.yaml"
         ) as f:
@@ -1724,8 +1727,8 @@ with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False, js=js, css=css
                                 )
                             with gr.Row():
                                 text_low_lr_rate = gr.Slider(
-                                    minimum=0.2,
-                                    maximum=0.6,
+                                    minimum=0,
+                                    maximum=1,
                                     step=0.05,
                                     label=i18n("文本模块学习率权重"),
                                     value=0.4,
@@ -1734,7 +1737,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False, js=js, css=css
                                 lora_rank = gr.Radio(
                                     label=i18n("LoRA秩"),
                                     value="32",
-                                    choices=["16", "32", "64", "128"],
+                                    choices=["16", "32", "64", "128", "256", "512", "1024","2048", "4096"],
                                     visible=True if version in v3v4set else False,
                                 )  # v1v2 not need
                                 save_every_epoch = gr.Slider(
@@ -1796,7 +1799,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False, js=js, css=css
                                 )
                                 total_epoch1Bb = gr.Slider(
                                     minimum=2,
-                                    maximum=50,
+                                    maximum=max_sovits_epoch,
                                     step=1,
                                     label=i18n("总训练轮数total_epoch"),
                                     value=15,