diff --git a/app.py b/app.py
index 89dd5ec71dd48b5a523878dbe6e2bc930faf5cfc..3df291c341af0f20a19ec4da48f6ea37b2f4545a 100644
--- a/app.py
+++ b/app.py
@@ -66,6 +66,30 @@ async def load_model(character_name: str = Form(...), model_path: str = Form(...
     try:
         print(f"📦 Loading character: {character_name} from {full_path}")
         genie_tts.load_character(character_name, full_path, language)
+        
+        # 自动探测参考音频配置
+        prompt_json_path = os.path.join(full_path, "prompt_wav.json")
+        ref_wav_path = os.path.join(full_path, "ref.wav")
+        
+        if os.path.exists(prompt_json_path):
+            import json
+            with open(prompt_json_path, "r", encoding="utf-8") as f:
+                data = json.load(f)
+                config = data.get("default", {})
+                REF_CACHE[character_name] = {
+                    "path": os.path.join(full_path, config.get("wav_path", "ref.wav")),
+                    "text": config.get("prompt_text", ""),
+                    "lang": config.get("prompt_lang", language)
+                }
+                print(f"📖 Loaded ref info from JSON for {character_name}")
+        elif os.path.exists(ref_wav_path):
+            REF_CACHE[character_name] = {
+                "path": ref_wav_path,
+                "text": "", 
+                "lang": language
+            }
+            print(f"🎵 Found ref.wav for {character_name}")
+
         return {"status": "success", "message": f"Character '{character_name}' loaded."}
     except Exception as e:
         raise HTTPException(status_code=500, detail=str(e))
@@ -76,12 +100,21 @@ async def upload_and_tts(
     prompt_text: str = Form(...),
     text: str = Form(...),
     language: str = Form("zh"),
+    text_lang: str = Form(None),
     file: UploadFile = File(...)
 ):
     """
     上传临时参考音频并生成语音
     """
     try:
+        # 🟢 确保模型已加载
+        if not genie_tts.model_manager.get(character_name):
+             print(f"⚠️ Character {character_name} not loaded, trying to load...")
+             char_path = os.path.join(MODELS_ROOT, character_name.lower())
+             if not os.path.exists(char_path):
+                 char_path = os.path.join(MODELS_ROOT, "mzm") # 兜底逻辑
+             genie_tts.load_character(character_name, char_path, language)
+
         ts = int(time.time() * 1000)
         save_path = f"/tmp/ref_{ts}.wav"
         os.makedirs("/tmp", exist_ok=True)
@@ -89,23 +122,37 @@ async def upload_and_tts(
         with open(save_path, "wb") as buffer:
             shutil.copyfileobj(file.file, buffer)
         
-        print(f"🔥 [Custom] Using temp audio for {character_name}: {save_path}")
+        print(f"🔥 [Custom] Using temp audio: {save_path}")
         genie_tts.set_reference_audio(character_name, save_path, prompt_text, language)
         
         out_path = f"/tmp/out_{ts}.wav"
-        genie_tts.tts(character_name, text, save_path=out_path, play=False)
+        # 🟢 执行 TTS
+        genie_tts.tts(character_name, text, save_path=out_path, play=False, text_language=text_lang)
         
+        # 🟢 关键：强制等待文件出现（最多等5秒）
+        wait_time = 0
+        while not os.path.exists(out_path) and wait_time < 50:
+            time.sleep(0.1)
+            wait_time += 1
+
+        if not os.path.exists(out_path):
+            raise HTTPException(status_code=500, detail="Audio file generation timed out or failed.")
+
         def iterfile():
-            with open(out_path, "rb") as f:
-                yield from f
-            try: 
-                os.remove(save_path)
-                os.remove(out_path)
-            except: pass
+            try:
+                with open(out_path, "rb") as f:
+                    yield from f
+            finally:
+                # 给一点延迟确保读取完毕后再删除
+                time.sleep(1)
+                try:
+                    if os.path.exists(save_path): os.remove(save_path)
+                    if os.path.exists(out_path): os.remove(out_path)
+                except: pass
 
         return StreamingResponse(iterfile(), media_type="audio/wav")
     except Exception as e:
-        print(f"❌ Error in upload/tts: {e}")
+        print(f"❌ Error in upload/tts: {str(e)}")
         raise HTTPException(status_code=500, detail=str(e))
 
 @app.post("/tts")
@@ -138,6 +185,12 @@ async def dynamic_tts(
         out_path = f"/tmp/out_dyn_{int(time.time())}.wav"
         genie_tts.tts(character_name, text, save_path=out_path, play=False, text_language=text_lang)
         
+        # 🟢 同样增加文件等待
+        wait_time = 0
+        while not os.path.exists(out_path) and wait_time < 50:
+            time.sleep(0.1)
+            wait_time += 1
+
         return StreamingResponse(open(out_path, "rb"), media_type="audio/wav")
     except Exception as e:
         print(f"❌ Error: {e}")
diff --git a/genie_tts/Audio/Audio.py b/genie_tts/Audio/Audio.py
index 2c519b4da2c38bb9d88f2dfca96c6000ddb95a8a..3819c96ebc94e5d29bee980ff3e9b3c8bd93dc33 100644
--- a/genie_tts/Audio/Audio.py
+++ b/genie_tts/Audio/Audio.py
@@ -1,51 +1,51 @@
-import os
-import soundfile as sf
-import soxr
-import numpy as np
-import logging
-from typing import Optional
-
-logger = logging.getLogger(__name__)
-
-# 音频时长建议范围 (秒)
-MIN_DURATION_S = 3
-MAX_DURATION_S = 10
-# 在音频末尾追加的静音时长 (秒)
-SILENCE_TO_APPEND_S = 0.3
-# 模型期望的目标采样率
-TARGET_SAMPLING_RATE = 16000
-
-
-def load_audio(
-        audio_path: str,
-        target_sampling_rate: int = TARGET_SAMPLING_RATE
-) -> Optional[np.ndarray]:
-    try:
-        wav, original_sr = sf.read(audio_path, dtype='float32')
-        if wav.ndim > 1:
-            wav = np.mean(wav, axis=1)  # 多声道转单声道。
-        if original_sr != target_sampling_rate:
-            wav = soxr.resample(wav, original_sr, target_sampling_rate, quality='hq')  # 重采样。
-
-    except Exception as e:
-        logger.error(f"Failed to load reference audio: {audio_path}. Error: {e}")
-        return None
-
-    # 检查音频长度是否在建议范围之外
-    min_samples = int(MIN_DURATION_S * target_sampling_rate)
-    max_samples = int(MAX_DURATION_S * target_sampling_rate)
-    if not (min_samples <= wav.shape[0] <= max_samples):
-        duration = len(wav) / target_sampling_rate
-        logger.warning(
-            f"The reference audio '{os.path.basename(audio_path)}' has a duration of {duration:.2f} seconds, "
-            f"which is outside the recommended range of {MIN_DURATION_S} to {MAX_DURATION_S} seconds!"
-        )
-
-    # 创建并拼接静音
-    silence_samples = int(SILENCE_TO_APPEND_S * target_sampling_rate)
-    silence_array = np.zeros(silence_samples, dtype=np.float32)
-    wav_processed = np.concatenate([wav, silence_array])
-
-    # 为模型输入增加批次维度
-    # wav_processed = np.expand_dims(wav_processed, axis=0)
-    return wav_processed
+import os
+import soundfile as sf
+import soxr
+import numpy as np
+import logging
+from typing import Optional
+
+logger = logging.getLogger(__name__)
+
+# 音频时长建议范围 (秒)
+MIN_DURATION_S = 3
+MAX_DURATION_S = 10
+# 在音频末尾追加的静音时长 (秒)
+SILENCE_TO_APPEND_S = 0.3
+# 模型期望的目标采样率
+TARGET_SAMPLING_RATE = 16000
+
+
+def load_audio(
+        audio_path: str,
+        target_sampling_rate: int = TARGET_SAMPLING_RATE
+) -> Optional[np.ndarray]:
+    try:
+        wav, original_sr = sf.read(audio_path, dtype='float32')
+        if wav.ndim > 1:
+            wav = np.mean(wav, axis=1)  # 多声道转单声道。
+        if original_sr != target_sampling_rate:
+            wav = soxr.resample(wav, original_sr, target_sampling_rate, quality='hq')  # 重采样。
+
+    except Exception as e:
+        logger.error(f"Failed to load reference audio: {audio_path}. Error: {e}")
+        return None
+
+    # 检查音频长度是否在建议范围之外
+    min_samples = int(MIN_DURATION_S * target_sampling_rate)
+    max_samples = int(MAX_DURATION_S * target_sampling_rate)
+    if not (min_samples <= wav.shape[0] <= max_samples):
+        duration = len(wav) / target_sampling_rate
+        logger.warning(
+            f"The reference audio '{os.path.basename(audio_path)}' has a duration of {duration:.2f} seconds, "
+            f"which is outside the recommended range of {MIN_DURATION_S} to {MAX_DURATION_S} seconds!"
+        )
+
+    # 创建并拼接静音
+    silence_samples = int(SILENCE_TO_APPEND_S * target_sampling_rate)
+    silence_array = np.zeros(silence_samples, dtype=np.float32)
+    wav_processed = np.concatenate([wav, silence_array])
+
+    # 为模型输入增加批次维度
+    # wav_processed = np.expand_dims(wav_processed, axis=0)
+    return wav_processed
diff --git a/genie_tts/Audio/__pycache__/Audio.cpython-311.pyc b/genie_tts/Audio/__pycache__/Audio.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..8feec30071c864483b7e21a331801203ceae1aed
Binary files /dev/null and b/genie_tts/Audio/__pycache__/Audio.cpython-311.pyc differ
diff --git a/genie_tts/Audio/__pycache__/ReferenceAudio.cpython-311.pyc b/genie_tts/Audio/__pycache__/ReferenceAudio.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2ce41a436ae82831a1927ac8425f66fb9de37b88
Binary files /dev/null and b/genie_tts/Audio/__pycache__/ReferenceAudio.cpython-311.pyc differ
diff --git a/genie_tts/Audio/__pycache__/__init__.cpython-311.pyc b/genie_tts/Audio/__pycache__/__init__.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..140f93d8fda57a05d42a44ff5abf388d72191846
Binary files /dev/null and b/genie_tts/Audio/__pycache__/__init__.cpython-311.pyc differ
diff --git a/genie_tts/Converter/Converter.py b/genie_tts/Converter/Converter.py
index 6951e71aedaf4d08338f12c277c0da3e7cf917bf..d0fc7e794de0affd450df31fb64547c2d6b9a743 100644
--- a/genie_tts/Converter/Converter.py
+++ b/genie_tts/Converter/Converter.py
@@ -1,11 +1,11 @@
-from .v2.Converter import convert as convert_v2
-from .v2ProPlus.Converter import convert as convert_v2pp
-
-import os
-
-
-def convert(torch_ckpt_path: str, torch_pth_path: str, output_dir: str) -> None:
-    if os.path.getsize(torch_pth_path) > 150 * 1024 * 1024:  # 大于 150 MB
-        convert_v2pp(torch_ckpt_path, torch_pth_path, output_dir)
-    else:
-        convert_v2(torch_ckpt_path, torch_pth_path, output_dir)
+from .v2.Converter import convert as convert_v2
+from .v2ProPlus.Converter import convert as convert_v2pp
+
+import os
+
+
+def convert(torch_ckpt_path: str, torch_pth_path: str, output_dir: str) -> None:
+    if os.path.getsize(torch_pth_path) > 150 * 1024 * 1024:  # 大于 150 MB
+        convert_v2pp(torch_ckpt_path, torch_pth_path, output_dir)
+    else:
+        convert_v2(torch_ckpt_path, torch_pth_path, output_dir)
diff --git a/genie_tts/Converter/__pycache__/Converter.cpython-311.pyc b/genie_tts/Converter/__pycache__/Converter.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..98bd57b82f7dfdade427c904a50f6924c97ffd23
Binary files /dev/null and b/genie_tts/Converter/__pycache__/Converter.cpython-311.pyc differ
diff --git a/genie_tts/Converter/__pycache__/__init__.cpython-311.pyc b/genie_tts/Converter/__pycache__/__init__.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..63a605374d3f1d160e4c3fe59fe575ec3f9d9c08
Binary files /dev/null and b/genie_tts/Converter/__pycache__/__init__.cpython-311.pyc differ
diff --git a/genie_tts/Converter/__pycache__/load_state_dict.cpython-311.pyc b/genie_tts/Converter/__pycache__/load_state_dict.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..621e06b8a12e5bebd19c3f53e0a8a9ff8551657c
Binary files /dev/null and b/genie_tts/Converter/__pycache__/load_state_dict.cpython-311.pyc differ
diff --git a/genie_tts/Converter/__pycache__/utils.cpython-311.pyc b/genie_tts/Converter/__pycache__/utils.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..e89f6248c90c29b0b65e54251be25e630ba7057c
Binary files /dev/null and b/genie_tts/Converter/__pycache__/utils.cpython-311.pyc differ
diff --git a/genie_tts/Converter/load_state_dict.py b/genie_tts/Converter/load_state_dict.py
index 952d3580c094ca9b4c74b7ac94633b7817a3cc81..ceff2c0183fdcfaa1685b3202c41f2d57f4e3262 100644
--- a/genie_tts/Converter/load_state_dict.py
+++ b/genie_tts/Converter/load_state_dict.py
@@ -1,26 +1,26 @@
-import sys
-import os
-
-sys.path.append(os.path.dirname(__file__))
-
-import torch
-from io import BytesIO
-import utils
-
-
-def load_sovits_model(pth_path: str, device: str = 'cpu'):
-    f = open(pth_path, "rb")
-    meta = f.read(2)
-    if meta != b"PK":
-        # noinspection PyTypeChecker
-        data = b"PK" + f.read()
-        bio = BytesIO()
-        # noinspection PyTypeChecker
-        bio.write(data)
-        bio.seek(0)
-        return torch.load(bio, map_location=device, weights_only=False)
-    return torch.load(pth_path, map_location=device, weights_only=False)
-
-
-def load_gpt_model(ckpt_path: str, device: str = 'cpu'):
-    return torch.load(ckpt_path, map_location=device, weights_only=True)
+import sys
+import os
+
+sys.path.append(os.path.dirname(__file__))
+
+import torch
+from io import BytesIO
+import utils
+
+
+def load_sovits_model(pth_path: str, device: str = 'cpu'):
+    f = open(pth_path, "rb")
+    meta = f.read(2)
+    if meta != b"PK":
+        # noinspection PyTypeChecker
+        data = b"PK" + f.read()
+        bio = BytesIO()
+        # noinspection PyTypeChecker
+        bio.write(data)
+        bio.seek(0)
+        return torch.load(bio, map_location=device, weights_only=False)
+    return torch.load(pth_path, map_location=device, weights_only=False)
+
+
+def load_gpt_model(ckpt_path: str, device: str = 'cpu'):
+    return torch.load(ckpt_path, map_location=device, weights_only=True)
diff --git a/genie_tts/Converter/v2/Converter.py b/genie_tts/Converter/v2/Converter.py
index 39910b97c9e6cb8aeda2a4d89594dbe1106af0af..301edd060a51c3052a2b05121d9d96da729e5a79 100644
--- a/genie_tts/Converter/v2/Converter.py
+++ b/genie_tts/Converter/v2/Converter.py
@@ -1,146 +1,146 @@
-from .VITSConverter import VITSConverter
-from .T2SConverter import T2SModelConverter
-from .EncoderConverter import EncoderConverter
-from ...Utils.Constants import PACKAGE_NAME
-
-import logging
-from typing import Optional, Tuple
-import re
-import os
-import shutil
-import traceback
-import importlib.resources
-import contextlib
-
-logger = logging.getLogger()
-
-CACHE_DIR = os.path.join(os.getcwd(), "Cache")
-ENCODER_RESOURCE_PATH = "Data/v2/Models/t2s_encoder_fp32.onnx"
-STAGE_DECODER_RESOURCE_PATH = "Data/v2/Models/t2s_stage_decoder_fp32.onnx"
-FIRST_STAGE_DECODER_RESOURCE_PATH = "Data/v2/Models/t2s_first_stage_decoder_fp32.onnx"
-VITS_RESOURCE_PATH = "Data/v2/Models/vits_fp32.onnx"
-T2S_KEYS_RESOURCE_PATH = "Data/v2/Keys/t2s_onnx_keys.txt"
-VITS_KEYS_RESOURCE_PATH = "Data/v2/Keys/vits_onnx_keys.txt"
-
-
-def find_ckpt_and_pth(directory: str) -> Tuple[Optional[str], Optional[str]]:
-    """
-    在 directory（不递归子目录）里查找：
-    - .ckpt：从所有 .ckpt 文件名中搜索 'e{正整数}' 作为 epoch（找不到则视为 e0），
-             选择 epoch 最大的那个文件（若无则为 None）
-    - .pth ：从所有 .pth 文件名中搜索 'e{正整数}' 作为 epoch（找不到则视为 e0），
-             选择 epoch 最大的那个文件（若无则为 None）
-    若出现相同 epoch，选修改时间较新的文件以打破平手。
-    """
-    best_ckpt_path: Optional[str] = None
-    best_ckpt_epoch: int = -1
-
-    best_pth_path: Optional[str] = None
-    best_pth_epoch: int = -1
-
-    for filename in os.listdir(directory):
-        full_path = os.path.join(directory, filename)
-
-        if not os.path.isfile(full_path):
-            continue
-
-        # 提取 epoch
-        m = re.search(r"e(\d+)", filename, flags=re.IGNORECASE)
-        epoch = int(m.group(1)) if m else 0
-
-        # .ckpt 文件处理
-        if filename.lower().endswith(".ckpt"):
-            if (
-                    epoch > best_ckpt_epoch
-                    or (
-                    epoch == best_ckpt_epoch
-                    and best_ckpt_path is not None
-                    and os.path.getmtime(full_path) > os.path.getmtime(best_ckpt_path)
-            )
-            ):
-                best_ckpt_epoch = epoch
-                best_ckpt_path = full_path
-
-        # .pth 文件处理
-        elif filename.lower().endswith(".pth"):
-            if (
-                    epoch > best_pth_epoch
-                    or (
-                    epoch == best_pth_epoch
-                    and best_pth_path is not None
-                    and os.path.getmtime(full_path) > os.path.getmtime(best_pth_path)
-            )
-            ):
-                best_pth_epoch = epoch
-                best_pth_path = full_path
-
-    return best_ckpt_path, best_pth_path
-
-
-def remove_folder(folder: str) -> None:
-    try:
-        if os.path.exists(folder):
-            shutil.rmtree(folder)
-            logger.info(f"🧹 Folder cleaned: {folder}")
-    except Exception as e:
-        logger.error(f"❌ Failed to clean folder {folder}: {e}")
-
-
-def convert(torch_ckpt_path: str,
-            torch_pth_path: str,
-            output_dir: str):
-    # 确保缓存和输出目录存在
-    os.makedirs(CACHE_DIR, exist_ok=True)
-    os.makedirs(output_dir, exist_ok=True)
-
-    if len(os.listdir(output_dir)) > 0:
-        logger.warning(f"The output directory {output_dir} is not empty!")
-
-    with contextlib.ExitStack() as stack:
-        files = importlib.resources.files(PACKAGE_NAME)
-
-        def enter(p):
-            return stack.enter_context(importlib.resources.as_file(files.joinpath(p)))
-
-        encoder_onnx_path = enter(ENCODER_RESOURCE_PATH)
-        stage_decoder_path = enter(STAGE_DECODER_RESOURCE_PATH)
-        first_stage_decoder_path = enter(FIRST_STAGE_DECODER_RESOURCE_PATH)
-        vits_onnx_path = enter(VITS_RESOURCE_PATH)
-        t2s_keys_path = enter(T2S_KEYS_RESOURCE_PATH)
-        vits_keys_path = enter(VITS_KEYS_RESOURCE_PATH)
-
-        converter_1 = T2SModelConverter(
-            torch_ckpt_path=torch_ckpt_path,
-            stage_decoder_onnx_path=str(stage_decoder_path),
-            first_stage_decoder_onnx_path=str(first_stage_decoder_path),
-            key_list_file=str(t2s_keys_path),
-            output_dir=output_dir,
-            cache_dir=CACHE_DIR,
-        )
-        converter_2 = VITSConverter(
-            torch_pth_path=torch_pth_path,
-            vits_onnx_path=str(vits_onnx_path),
-            key_list_file=str(vits_keys_path),
-            output_dir=output_dir,
-            cache_dir=CACHE_DIR,
-        )
-        converter_3 = EncoderConverter(
-            ckpt_path=torch_ckpt_path,
-            pth_path=torch_pth_path,
-            onnx_input_path=str(encoder_onnx_path),
-            output_dir=output_dir,
-        )
-
-        try:
-            converter_1.run_full_process()
-            converter_2.run_full_process()
-            converter_3.run_full_process()
-            logger.info(f"🎉 Conversion successful! Saved to: {os.path.abspath(output_dir)}\n"
-                        f"- Model Type: V2")
-        except Exception:
-            logger.error(f"❌ A critical error occurred during the conversion process")
-            logger.error(traceback.format_exc())
-            remove_folder(output_dir)  # 只在失败时清理输出目录
-        finally:
-            # 无论成功还是失败，都尝试清理缓存目录
-            remove_folder(CACHE_DIR)
+from .VITSConverter import VITSConverter
+from .T2SConverter import T2SModelConverter
+from .EncoderConverter import EncoderConverter
+from ...Utils.Constants import PACKAGE_NAME
+
+import logging
+from typing import Optional, Tuple
+import re
+import os
+import shutil
+import traceback
+import importlib.resources
+import contextlib
+
+logger = logging.getLogger()
+
+CACHE_DIR = os.path.join(os.getcwd(), "Cache")
+ENCODER_RESOURCE_PATH = "Data/v2/Models/t2s_encoder_fp32.onnx"
+STAGE_DECODER_RESOURCE_PATH = "Data/v2/Models/t2s_stage_decoder_fp32.onnx"
+FIRST_STAGE_DECODER_RESOURCE_PATH = "Data/v2/Models/t2s_first_stage_decoder_fp32.onnx"
+VITS_RESOURCE_PATH = "Data/v2/Models/vits_fp32.onnx"
+T2S_KEYS_RESOURCE_PATH = "Data/v2/Keys/t2s_onnx_keys.txt"
+VITS_KEYS_RESOURCE_PATH = "Data/v2/Keys/vits_onnx_keys.txt"
+
+
+def find_ckpt_and_pth(directory: str) -> Tuple[Optional[str], Optional[str]]:
+    """
+    在 directory（不递归子目录）里查找：
+    - .ckpt：从所有 .ckpt 文件名中搜索 'e{正整数}' 作为 epoch（找不到则视为 e0），
+             选择 epoch 最大的那个文件（若无则为 None）
+    - .pth ：从所有 .pth 文件名中搜索 'e{正整数}' 作为 epoch（找不到则视为 e0），
+             选择 epoch 最大的那个文件（若无则为 None）
+    若出现相同 epoch，选修改时间较新的文件以打破平手。
+    """
+    best_ckpt_path: Optional[str] = None
+    best_ckpt_epoch: int = -1
+
+    best_pth_path: Optional[str] = None
+    best_pth_epoch: int = -1
+
+    for filename in os.listdir(directory):
+        full_path = os.path.join(directory, filename)
+
+        if not os.path.isfile(full_path):
+            continue
+
+        # 提取 epoch
+        m = re.search(r"e(\d+)", filename, flags=re.IGNORECASE)
+        epoch = int(m.group(1)) if m else 0
+
+        # .ckpt 文件处理
+        if filename.lower().endswith(".ckpt"):
+            if (
+                    epoch > best_ckpt_epoch
+                    or (
+                    epoch == best_ckpt_epoch
+                    and best_ckpt_path is not None
+                    and os.path.getmtime(full_path) > os.path.getmtime(best_ckpt_path)
+            )
+            ):
+                best_ckpt_epoch = epoch
+                best_ckpt_path = full_path
+
+        # .pth 文件处理
+        elif filename.lower().endswith(".pth"):
+            if (
+                    epoch > best_pth_epoch
+                    or (
+                    epoch == best_pth_epoch
+                    and best_pth_path is not None
+                    and os.path.getmtime(full_path) > os.path.getmtime(best_pth_path)
+            )
+            ):
+                best_pth_epoch = epoch
+                best_pth_path = full_path
+
+    return best_ckpt_path, best_pth_path
+
+
+def remove_folder(folder: str) -> None:
+    try:
+        if os.path.exists(folder):
+            shutil.rmtree(folder)
+            logger.info(f"🧹 Folder cleaned: {folder}")
+    except Exception as e:
+        logger.error(f"❌ Failed to clean folder {folder}: {e}")
+
+
+def convert(torch_ckpt_path: str,
+            torch_pth_path: str,
+            output_dir: str):
+    # 确保缓存和输出目录存在
+    os.makedirs(CACHE_DIR, exist_ok=True)
+    os.makedirs(output_dir, exist_ok=True)
+
+    if len(os.listdir(output_dir)) > 0:
+        logger.warning(f"The output directory {output_dir} is not empty!")
+
+    with contextlib.ExitStack() as stack:
+        files = importlib.resources.files(PACKAGE_NAME)
+
+        def enter(p):
+            return stack.enter_context(importlib.resources.as_file(files.joinpath(p)))
+
+        encoder_onnx_path = enter(ENCODER_RESOURCE_PATH)
+        stage_decoder_path = enter(STAGE_DECODER_RESOURCE_PATH)
+        first_stage_decoder_path = enter(FIRST_STAGE_DECODER_RESOURCE_PATH)
+        vits_onnx_path = enter(VITS_RESOURCE_PATH)
+        t2s_keys_path = enter(T2S_KEYS_RESOURCE_PATH)
+        vits_keys_path = enter(VITS_KEYS_RESOURCE_PATH)
+
+        converter_1 = T2SModelConverter(
+            torch_ckpt_path=torch_ckpt_path,
+            stage_decoder_onnx_path=str(stage_decoder_path),
+            first_stage_decoder_onnx_path=str(first_stage_decoder_path),
+            key_list_file=str(t2s_keys_path),
+            output_dir=output_dir,
+            cache_dir=CACHE_DIR,
+        )
+        converter_2 = VITSConverter(
+            torch_pth_path=torch_pth_path,
+            vits_onnx_path=str(vits_onnx_path),
+            key_list_file=str(vits_keys_path),
+            output_dir=output_dir,
+            cache_dir=CACHE_DIR,
+        )
+        converter_3 = EncoderConverter(
+            ckpt_path=torch_ckpt_path,
+            pth_path=torch_pth_path,
+            onnx_input_path=str(encoder_onnx_path),
+            output_dir=output_dir,
+        )
+
+        try:
+            converter_1.run_full_process()
+            converter_2.run_full_process()
+            converter_3.run_full_process()
+            logger.info(f"🎉 Conversion successful! Saved to: {os.path.abspath(output_dir)}\n"
+                        f"- Model Type: V2")
+        except Exception:
+            logger.error(f"❌ A critical error occurred during the conversion process")
+            logger.error(traceback.format_exc())
+            remove_folder(output_dir)  # 只在失败时清理输出目录
+        finally:
+            # 无论成功还是失败，都尝试清理缓存目录
+            remove_folder(CACHE_DIR)
diff --git a/genie_tts/Converter/v2/EncoderConverter.py b/genie_tts/Converter/v2/EncoderConverter.py
index 33ed75ba3277ec535f9c2f88304c7c07cc2752b4..84eaf47856716ecf5646d81f439ae335390ba6f8 100644
--- a/genie_tts/Converter/v2/EncoderConverter.py
+++ b/genie_tts/Converter/v2/EncoderConverter.py
@@ -1,106 +1,106 @@
-import torch
-import onnx
-import os
-
-from ..load_state_dict import load_gpt_model, load_sovits_model
-
-
-class EncoderConverter:
-    """
-    一个转换器，用于为 t2s_encoder 模型创建：
-    1. 一个从 .ckpt 和 .pth 文件中合并而来的全精度 (fp32) .bin 权重文件。
-    2. 一个链接到该 .bin 文件的 ONNX 模型。
-    """
-
-    def __init__(self,
-                 ckpt_path: str,
-                 pth_path: str,
-                 onnx_input_path: str,
-                 output_dir: str,
-                 ):
-        self.ckpt_path: str = ckpt_path
-        self.pth_path: str = pth_path
-        self.onnx_input_path: str = onnx_input_path
-        self.output_dir: str = output_dir
-
-        # 定义最终输出文件的路径
-        self.output_bin_path: str = os.path.join(self.output_dir, "t2s_encoder_fp32.bin")
-        self.output_onnx_path: str = os.path.join(self.output_dir, "t2s_encoder_fp32.onnx")
-
-        # 确保输出目录存在
-        os.makedirs(self.output_dir, exist_ok=True)
-
-        # 检查所有输入文件是否存在
-        for path in [self.ckpt_path, self.pth_path, self.onnx_input_path]:
-            if not os.path.exists(path):
-                raise FileNotFoundError(f"Error: Input file not found! Path: {path}")
-
-    def run_full_process(self):
-        # 1. 定义固定的 ONNX 权重键列表 (此顺序决定了 .bin 文件的布局)
-        onnx_keys = [
-            "encoder.ar_text_embedding.word_embeddings.weight",
-            "encoder.bert_proj.weight",
-            "encoder.bert_proj.bias",
-            "encoder.ar_text_position.alpha",
-            "vits.ssl_proj.weight",
-            "vits.ssl_proj.bias",
-            "vits.quantizer.vq.layers.0._codebook.embed"
-        ]
-
-        # 2. 加载所有必要的模型和权重
-        ckpt_state_dict = load_gpt_model(self.ckpt_path)['weight']
-        pth_state_dict = load_sovits_model(self.pth_path)['weight']
-        model = onnx.load(self.onnx_input_path, load_external_data=False)
-        initializer_map = {init.name: init for init in model.graph.initializer}
-        current_offset = 0
-        bin_filename = os.path.basename(self.output_bin_path)
-
-        # 3. 生成 .bin 文件并同步修改 ONNX 模型
-        with open(self.output_bin_path, 'wb') as f_bin:
-            for onnx_key in onnx_keys:
-                source_key = ""
-                source_dict = None
-
-                if onnx_key.startswith("encoder."):
-                    source_key = "model." + onnx_key[len("encoder."):]
-                    source_dict = ckpt_state_dict
-                elif onnx_key.startswith("vits."):
-                    source_key = onnx_key[len("vits."):]
-                    source_dict = pth_state_dict
-
-                if source_dict is None:
-                    raise ValueError(
-                        f"❌ Critical error: Unable to determine the weight source for ONNX key '{onnx_key}'.")
-                # 从源文件中提取张量
-                tensor = source_dict.get(source_key)
-                if tensor is None:
-                    raise ValueError(
-                        f"❌ Critical error: Key '{source_key}' (corresponding to ONNX key '{onnx_key}') not found in the source file.")
-
-                # 转换为 fp32 numpy 数组并获取字节
-                numpy_array_fp32 = tensor.to(torch.float32).cpu().numpy()
-                tensor_bytes = numpy_array_fp32.tobytes()
-                tensor_length = len(tensor_bytes)
-                f_bin.write(tensor_bytes)
-
-                # 在 ONNX 模型中找到对应的 initializer 并修改它
-                if onnx_key in initializer_map:
-                    tensor_proto = initializer_map[onnx_key]
-
-                    tensor_proto.ClearField('raw_data')
-                    tensor_proto.data_location = onnx.TensorProto.EXTERNAL
-                    del tensor_proto.external_data[:]
-
-                    keys_to_set = ["location", "offset", "length"]
-                    values_to_set = [bin_filename, str(current_offset), str(tensor_length)]
-
-                    for k, v in zip(keys_to_set, values_to_set):
-                        entry = tensor_proto.external_data.add()
-                        entry.key = k
-                        entry.value = v
-
-                # 更新下一个权重的偏移量
-                current_offset += tensor_length
-
-        # 4. 保存修改后的 ONNX 模型
-        onnx.save(model, self.output_onnx_path)
+import torch
+import onnx
+import os
+
+from ..load_state_dict import load_gpt_model, load_sovits_model
+
+
+class EncoderConverter:
+    """
+    一个转换器，用于为 t2s_encoder 模型创建：
+    1. 一个从 .ckpt 和 .pth 文件中合并而来的全精度 (fp32) .bin 权重文件。
+    2. 一个链接到该 .bin 文件的 ONNX 模型。
+    """
+
+    def __init__(self,
+                 ckpt_path: str,
+                 pth_path: str,
+                 onnx_input_path: str,
+                 output_dir: str,
+                 ):
+        self.ckpt_path: str = ckpt_path
+        self.pth_path: str = pth_path
+        self.onnx_input_path: str = onnx_input_path
+        self.output_dir: str = output_dir
+
+        # 定义最终输出文件的路径
+        self.output_bin_path: str = os.path.join(self.output_dir, "t2s_encoder_fp32.bin")
+        self.output_onnx_path: str = os.path.join(self.output_dir, "t2s_encoder_fp32.onnx")
+
+        # 确保输出目录存在
+        os.makedirs(self.output_dir, exist_ok=True)
+
+        # 检查所有输入文件是否存在
+        for path in [self.ckpt_path, self.pth_path, self.onnx_input_path]:
+            if not os.path.exists(path):
+                raise FileNotFoundError(f"Error: Input file not found! Path: {path}")
+
+    def run_full_process(self):
+        # 1. 定义固定的 ONNX 权重键列表 (此顺序决定了 .bin 文件的布局)
+        onnx_keys = [
+            "encoder.ar_text_embedding.word_embeddings.weight",
+            "encoder.bert_proj.weight",
+            "encoder.bert_proj.bias",
+            "encoder.ar_text_position.alpha",
+            "vits.ssl_proj.weight",
+            "vits.ssl_proj.bias",
+            "vits.quantizer.vq.layers.0._codebook.embed"
+        ]
+
+        # 2. 加载所有必要的模型和权重
+        ckpt_state_dict = load_gpt_model(self.ckpt_path)['weight']
+        pth_state_dict = load_sovits_model(self.pth_path)['weight']
+        model = onnx.load(self.onnx_input_path, load_external_data=False)
+        initializer_map = {init.name: init for init in model.graph.initializer}
+        current_offset = 0
+        bin_filename = os.path.basename(self.output_bin_path)
+
+        # 3. 生成 .bin 文件并同步修改 ONNX 模型
+        with open(self.output_bin_path, 'wb') as f_bin:
+            for onnx_key in onnx_keys:
+                source_key = ""
+                source_dict = None
+
+                if onnx_key.startswith("encoder."):
+                    source_key = "model." + onnx_key[len("encoder."):]
+                    source_dict = ckpt_state_dict
+                elif onnx_key.startswith("vits."):
+                    source_key = onnx_key[len("vits."):]
+                    source_dict = pth_state_dict
+
+                if source_dict is None:
+                    raise ValueError(
+                        f"❌ Critical error: Unable to determine the weight source for ONNX key '{onnx_key}'.")
+                # 从源文件中提取张量
+                tensor = source_dict.get(source_key)
+                if tensor is None:
+                    raise ValueError(
+                        f"❌ Critical error: Key '{source_key}' (corresponding to ONNX key '{onnx_key}') not found in the source file.")
+
+                # 转换为 fp32 numpy 数组并获取字节
+                numpy_array_fp32 = tensor.to(torch.float32).cpu().numpy()
+                tensor_bytes = numpy_array_fp32.tobytes()
+                tensor_length = len(tensor_bytes)
+                f_bin.write(tensor_bytes)
+
+                # 在 ONNX 模型中找到对应的 initializer 并修改它
+                if onnx_key in initializer_map:
+                    tensor_proto = initializer_map[onnx_key]
+
+                    tensor_proto.ClearField('raw_data')
+                    tensor_proto.data_location = onnx.TensorProto.EXTERNAL
+                    del tensor_proto.external_data[:]
+
+                    keys_to_set = ["location", "offset", "length"]
+                    values_to_set = [bin_filename, str(current_offset), str(tensor_length)]
+
+                    for k, v in zip(keys_to_set, values_to_set):
+                        entry = tensor_proto.external_data.add()
+                        entry.key = k
+                        entry.value = v
+
+                # 更新下一个权重的偏移量
+                current_offset += tensor_length
+
+        # 4. 保存修改后的 ONNX 模型
+        onnx.save(model, self.output_onnx_path)
diff --git a/genie_tts/Converter/v2/T2SConverter.py b/genie_tts/Converter/v2/T2SConverter.py
index 6a8de6925697ec911f6a22cfa4cdb31f4f73cbac..3cc386379cc56f008592bfeecc25852feae72d61 100644
--- a/genie_tts/Converter/v2/T2SConverter.py
+++ b/genie_tts/Converter/v2/T2SConverter.py
@@ -1,125 +1,125 @@
-import torch
-import onnx
-import numpy as np
-import json
-import os
-from collections import OrderedDict
-
-from ..load_state_dict import load_gpt_model
-
-
-class T2SModelConverter:
-    """
-    一个专门的转换器，用于处理 t2s (Text-to-Speech) 模型。
-    - PyTorch 模型: .ckpt 文件
-    - ONNX 模型: t2s_stage_decoder_fp32.onnx
-    - 遵循特定的键名映射规则。
-    """
-
-    def __init__(self,
-                 torch_ckpt_path: str,
-                 stage_decoder_onnx_path: str,
-                 first_stage_decoder_onnx_path: str,
-                 key_list_file: str,
-                 output_dir: str,
-                 cache_dir: str,
-                 ):
-        self.torch_ckpt_path: str = torch_ckpt_path
-        self.stage_decoder_onnx_path: str = stage_decoder_onnx_path
-        self.first_stage_decoder_onnx_path: str = first_stage_decoder_onnx_path
-        self.key_list_file: str = key_list_file
-        self.output_dir: str = output_dir
-        self.cache_dir: str = cache_dir
-
-        os.makedirs(self.output_dir, exist_ok=True)
-        os.makedirs(self.output_dir, exist_ok=True)
-
-        # 定义输出文件路径
-        self.fp16_bin_path: str = os.path.join(self.output_dir, "t2s_shared_fp16.bin")
-        self.index_table_path: str = os.path.join(self.cache_dir, "t2s_weights_index_fp32.json")
-        self.relinked_encoder_path: str = os.path.join(self.output_dir, "t2s_encoder_fp32.onnx")
-        self.relinked_stage_decoder_path: str = os.path.join(self.output_dir, "t2s_stage_decoder_fp32.onnx")
-        self.relinked_first_stage_decoder_path: str = os.path.join(self.output_dir, "t2s_first_stage_decoder_fp32.onnx")
-        self.reconstructed_fp32_bin_path = os.path.join(self.output_dir, "t2s_shared_fp32.bin")
-
-    def step1_create_fp16_bin_with_key_mapping(self):
-        """
-        (1) 根据特定的键映射规则，从 .ckpt 创建 fp16 .bin 和 fp32 索引。
-            (已根据用户验证脚本的正确逻辑进行最终修正)
-        """
-        if not os.path.exists(self.key_list_file):
-            raise FileNotFoundError(
-                f"Error: Stage 1 requires the key list file, but it was not found: {self.key_list_file}")
-
-        with open(self.key_list_file, 'r') as f:
-            onnx_keys = [line.strip() for line in f.readlines()]
-
-        ckpt_data = load_gpt_model(self.torch_ckpt_path)
-        if 'weight' not in ckpt_data:
-            raise KeyError(
-                f"❌ Error: 'weight' key not found in the .ckpt file. Top-level keys in the file are: {list(ckpt_data.keys())}")
-
-        torch_state_dict = ckpt_data['weight']
-
-        index_table = OrderedDict()
-        current_fp32_offset = 0
-
-        with open(self.fp16_bin_path, 'wb') as f_bin:
-            for onnx_key in onnx_keys:
-                transformed_onnx_key = onnx_key.replace('transformer_encoder', 'h')
-                torch_lookup_key = f"model.{transformed_onnx_key}"
-                torch_tensor = torch_state_dict.get(torch_lookup_key)
-                numpy_array_fp16 = torch_tensor.to(torch.float16).cpu().numpy()
-                f_bin.write(numpy_array_fp16.tobytes())
-                tensor_length_fp32 = numpy_array_fp16.nbytes * 2
-                index_table[onnx_key] = {'offset': current_fp32_offset, 'length': tensor_length_fp32}
-                current_fp32_offset += tensor_length_fp32
-
-        with open(self.index_table_path, 'w') as f_json:
-            json.dump(index_table, f_json, indent=4)  # type: ignore
-
-    def step2_relink_onnx_for_fp32(self, old_model: str, new_model: str):
-        """
-        (2) 根据 fp32 索引表，修改 ONNX 模型，使其链接到未来的全精度 .bin。
-            (使用与第一个脚本相同的、更稳定的底层方法)
-        """
-        if not os.path.exists(self.index_table_path):
-            raise FileNotFoundError(
-                f"Error: Stage 2 requires the index file, but it was not found: {self.index_table_path}")
-
-        # 加载描述 fp32 布局的索引表
-        with open(self.index_table_path, 'r') as f:
-            index_table = json.load(f)
-
-        model = onnx.load_model(old_model, load_external_data=False)
-        reconstructed_bin_filename = os.path.basename(self.reconstructed_fp32_bin_path)
-
-        for tensor in model.graph.initializer:
-            if tensor.name in index_table:
-                tensor.ClearField('raw_data')
-                tensor.data_location = onnx.TensorProto.EXTERNAL
-                info = index_table[tensor.name]
-                del tensor.external_data[:]
-                keys = ["location", "offset", "length"]
-                values = [reconstructed_bin_filename, str(info['offset']), str(info['length'])]
-
-                for k, v in zip(keys, values):
-                    entry = tensor.external_data.add()
-                    entry.key = k
-                    entry.value = v
-
-        onnx.save(model, new_model)
-
-    @staticmethod
-    def step3_reconstruct_fp32_bin_from_fp16(fp16_bin_path: str, output_fp32_bin_path: str):
-        """
-        (3) 静态工具函数：从半精度 .bin 文件还原出全精度 .bin 文件。
-        """
-        fp16_array = np.fromfile(fp16_bin_path, dtype=np.float16)
-        fp32_array = fp16_array.astype(np.float32)
-        fp32_array.tofile(output_fp32_bin_path)
-
-    def run_full_process(self):
-        self.step1_create_fp16_bin_with_key_mapping()
-        self.step2_relink_onnx_for_fp32(self.stage_decoder_onnx_path, self.relinked_stage_decoder_path)
-        self.step2_relink_onnx_for_fp32(self.first_stage_decoder_onnx_path, self.relinked_first_stage_decoder_path)
+import torch
+import onnx
+import numpy as np
+import json
+import os
+from collections import OrderedDict
+
+from ..load_state_dict import load_gpt_model
+
+
+class T2SModelConverter:
+    """
+    一个专门的转换器，用于处理 t2s (Text-to-Speech) 模型。
+    - PyTorch 模型: .ckpt 文件
+    - ONNX 模型: t2s_stage_decoder_fp32.onnx
+    - 遵循特定的键名映射规则。
+    """
+
+    def __init__(self,
+                 torch_ckpt_path: str,
+                 stage_decoder_onnx_path: str,
+                 first_stage_decoder_onnx_path: str,
+                 key_list_file: str,
+                 output_dir: str,
+                 cache_dir: str,
+                 ):
+        self.torch_ckpt_path: str = torch_ckpt_path
+        self.stage_decoder_onnx_path: str = stage_decoder_onnx_path
+        self.first_stage_decoder_onnx_path: str = first_stage_decoder_onnx_path
+        self.key_list_file: str = key_list_file
+        self.output_dir: str = output_dir
+        self.cache_dir: str = cache_dir
+
+        os.makedirs(self.output_dir, exist_ok=True)
+        os.makedirs(self.output_dir, exist_ok=True)
+
+        # 定义输出文件路径
+        self.fp16_bin_path: str = os.path.join(self.output_dir, "t2s_shared_fp16.bin")
+        self.index_table_path: str = os.path.join(self.cache_dir, "t2s_weights_index_fp32.json")
+        self.relinked_encoder_path: str = os.path.join(self.output_dir, "t2s_encoder_fp32.onnx")
+        self.relinked_stage_decoder_path: str = os.path.join(self.output_dir, "t2s_stage_decoder_fp32.onnx")
+        self.relinked_first_stage_decoder_path: str = os.path.join(self.output_dir, "t2s_first_stage_decoder_fp32.onnx")
+        self.reconstructed_fp32_bin_path = os.path.join(self.output_dir, "t2s_shared_fp32.bin")
+
+    def step1_create_fp16_bin_with_key_mapping(self):
+        """
+        (1) 根据特定的键映射规则，从 .ckpt 创建 fp16 .bin 和 fp32 索引。
+            (已根据用户验证脚本的正确逻辑进行最终修正)
+        """
+        if not os.path.exists(self.key_list_file):
+            raise FileNotFoundError(
+                f"Error: Stage 1 requires the key list file, but it was not found: {self.key_list_file}")
+
+        with open(self.key_list_file, 'r') as f:
+            onnx_keys = [line.strip() for line in f.readlines()]
+
+        ckpt_data = load_gpt_model(self.torch_ckpt_path)
+        if 'weight' not in ckpt_data:
+            raise KeyError(
+                f"❌ Error: 'weight' key not found in the .ckpt file. Top-level keys in the file are: {list(ckpt_data.keys())}")
+
+        torch_state_dict = ckpt_data['weight']
+
+        index_table = OrderedDict()
+        current_fp32_offset = 0
+
+        with open(self.fp16_bin_path, 'wb') as f_bin:
+            for onnx_key in onnx_keys:
+                transformed_onnx_key = onnx_key.replace('transformer_encoder', 'h')
+                torch_lookup_key = f"model.{transformed_onnx_key}"
+                torch_tensor = torch_state_dict.get(torch_lookup_key)
+                numpy_array_fp16 = torch_tensor.to(torch.float16).cpu().numpy()
+                f_bin.write(numpy_array_fp16.tobytes())
+                tensor_length_fp32 = numpy_array_fp16.nbytes * 2
+                index_table[onnx_key] = {'offset': current_fp32_offset, 'length': tensor_length_fp32}
+                current_fp32_offset += tensor_length_fp32
+
+        with open(self.index_table_path, 'w') as f_json:
+            json.dump(index_table, f_json, indent=4)  # type: ignore
+
+    def step2_relink_onnx_for_fp32(self, old_model: str, new_model: str):
+        """
+        (2) 根据 fp32 索引表，修改 ONNX 模型，使其链接到未来的全精度 .bin。
+            (使用与第一个脚本相同的、更稳定的底层方法)
+        """
+        if not os.path.exists(self.index_table_path):
+            raise FileNotFoundError(
+                f"Error: Stage 2 requires the index file, but it was not found: {self.index_table_path}")
+
+        # 加载描述 fp32 布局的索引表
+        with open(self.index_table_path, 'r') as f:
+            index_table = json.load(f)
+
+        model = onnx.load_model(old_model, load_external_data=False)
+        reconstructed_bin_filename = os.path.basename(self.reconstructed_fp32_bin_path)
+
+        for tensor in model.graph.initializer:
+            if tensor.name in index_table:
+                tensor.ClearField('raw_data')
+                tensor.data_location = onnx.TensorProto.EXTERNAL
+                info = index_table[tensor.name]
+                del tensor.external_data[:]
+                keys = ["location", "offset", "length"]
+                values = [reconstructed_bin_filename, str(info['offset']), str(info['length'])]
+
+                for k, v in zip(keys, values):
+                    entry = tensor.external_data.add()
+                    entry.key = k
+                    entry.value = v
+
+        onnx.save(model, new_model)
+
+    @staticmethod
+    def step3_reconstruct_fp32_bin_from_fp16(fp16_bin_path: str, output_fp32_bin_path: str):
+        """
+        (3) 静态工具函数：从半精度 .bin 文件还原出全精度 .bin 文件。
+        """
+        fp16_array = np.fromfile(fp16_bin_path, dtype=np.float16)
+        fp32_array = fp16_array.astype(np.float32)
+        fp32_array.tofile(output_fp32_bin_path)
+
+    def run_full_process(self):
+        self.step1_create_fp16_bin_with_key_mapping()
+        self.step2_relink_onnx_for_fp32(self.stage_decoder_onnx_path, self.relinked_stage_decoder_path)
+        self.step2_relink_onnx_for_fp32(self.first_stage_decoder_onnx_path, self.relinked_first_stage_decoder_path)
diff --git a/genie_tts/Converter/v2/VITSConverter.py b/genie_tts/Converter/v2/VITSConverter.py
index 8b04086a174d0edd3e0da75b7a76da9fc09491c6..41dedc0096c300b7c9d302586f4daff46909a298 100644
--- a/genie_tts/Converter/v2/VITSConverter.py
+++ b/genie_tts/Converter/v2/VITSConverter.py
@@ -1,129 +1,129 @@
-import torch
-import onnx
-import numpy as np
-import json
-import os
-from collections import OrderedDict
-
-from ..load_state_dict import load_sovits_model
-
-
-class VITSConverter:
-    """
-    一个转换器，用于从 PyTorch 模型创建：
-    1. 一个用于分发的半精度 (fp16) .bin 权重文件。
-    2. 一个与全精度 (fp32) 布局兼容的 ONNX 模型。
-    3. 一个可以将 fp16 .bin 文件还原为 fp32 .bin 的工具函数。
-    """
-
-    def __init__(self,
-                 torch_pth_path: str,
-                 vits_onnx_path: str,
-                 key_list_file: str,
-                 output_dir: str,
-                 cache_dir: str,
-                 ):
-        self.torch_pth_path: str = torch_pth_path
-        self.vits_onnx_path: str = vits_onnx_path
-        self.key_list_file: str = key_list_file
-        self.output_dir: str = output_dir
-        self.cache_dir: str = cache_dir
-        # 定义输出文件路径
-        self.fp16_bin_path: str = os.path.join(self.output_dir, "vits_fp16.bin")
-        self.index_table_path: str = os.path.join(self.cache_dir, "vits_weights_index_fp32.json")
-        self.relinked_fp32_onnx_path: str = os.path.join(self.output_dir, "vits_fp32.onnx")
-        self.reconstructed_fp32_bin_path: str = os.path.join(self.output_dir, "vits_fp32.bin")
-
-        # 确保输出目录存在
-        os.makedirs(self.cache_dir, exist_ok=True)
-        os.makedirs(self.output_dir, exist_ok=True)
-
-        if not os.path.exists(self.key_list_file):
-            raise FileNotFoundError(f"Error: Key list file not found! Path: {self.key_list_file}")
-
-    def step1_create_fp16_bin_and_fp32_index(self):
-        """
-        (1) 创建一个半精度 (fp16) 的 .bin 文件，但生成一个
-            描述全精度 (fp32) 布局的索引表。
-        """
-        # 加载 key 列表
-        with open(self.key_list_file, 'r') as f:
-            onnx_keys = [line.strip() for line in f.readlines()]
-
-        # 加载 PyTorch 模型权重
-        torch_state_dict = load_sovits_model(self.torch_pth_path)['weight']
-
-        index_table = OrderedDict()
-        current_fp32_offset = 0
-
-        with open(self.fp16_bin_path, 'wb') as f_bin:
-            for onnx_key in onnx_keys:
-                torch_key = onnx_key[len("vq_model."):] if onnx_key.startswith("vq_model.") else onnx_key
-
-                torch_tensor = torch_state_dict.get(torch_key)
-                if torch_tensor is None:
-                    raise ValueError(f"❌ Critical error: Key '{torch_key}' not found in the PyTorch weights")
-
-                # 转换为 fp16 并写入文件
-                torch_tensor_fp16 = torch_tensor.to(torch.float16)
-                numpy_array_fp16 = torch_tensor_fp16.cpu().numpy()
-                tensor_bytes_fp16 = numpy_array_fp16.tobytes()
-                f_bin.write(tensor_bytes_fp16)
-                tensor_length_fp32 = len(tensor_bytes_fp16) * 2
-                index_table[onnx_key] = {
-                    'offset': current_fp32_offset,
-                    'length': tensor_length_fp32
-                }
-                current_fp32_offset += tensor_length_fp32
-
-        # 保存描述 fp32 布局的索引表
-        with open(self.index_table_path, 'w') as f_json:
-            json.dump(index_table, f_json, indent=4)  # type: ignore
-
-    def step2_relink_onnx_for_fp32(self):
-        """
-        (2) 根据 fp32 索引表，修改 ONNX 模型，使其链接到一个
-            未来的、全精度的 .bin 文件。
-        """
-        # 加载描述 fp32 布局的索引表
-        with open(self.index_table_path, 'r') as f:
-            index_table = json.load(f)
-
-        model = onnx.load_model(self.vits_onnx_path, load_external_data=False)
-        reconstructed_bin_filename = os.path.basename(self.reconstructed_fp32_bin_path)
-
-        for tensor in model.graph.initializer:
-            if tensor.name in index_table:
-                tensor.ClearField('raw_data')
-                tensor.data_location = onnx.TensorProto.EXTERNAL
-                info = index_table[tensor.name]
-
-                del tensor.external_data[:]
-
-                keys = ["location", "offset", "length"]
-                values = [reconstructed_bin_filename, str(info['offset']), str(info['length'])]
-
-                for k, v in zip(keys, values):
-                    entry = tensor.external_data.add()
-                    entry.key = k
-                    entry.value = v
-
-        # 保存修改后的、链接到 fp32 权重的 ONNX 模型
-        onnx.save(model, self.relinked_fp32_onnx_path)
-
-    @staticmethod
-    def step3_reconstruct_fp32_bin_from_fp16(fp16_bin_path: str, output_fp32_bin_path: str):
-        """
-        (3) 静态工具函数：从半精度 .bin 文件还原出全精度 .bin 文件。
-
-        Args:
-            fp16_bin_path (str): 输入的半精度 .bin 文件路径。
-            output_fp32_bin_path (str): 输出的全精度 .bin 文件路径。
-        """
-        fp16_array = np.fromfile(fp16_bin_path, dtype=np.float16)
-        fp32_array = fp16_array.astype(np.float32)
-        fp32_array.tofile(output_fp32_bin_path)
-
-    def run_full_process(self):
-        self.step1_create_fp16_bin_and_fp32_index()
-        self.step2_relink_onnx_for_fp32()
+import torch
+import onnx
+import numpy as np
+import json
+import os
+from collections import OrderedDict
+
+from ..load_state_dict import load_sovits_model
+
+
+class VITSConverter:
+    """
+    一个转换器，用于从 PyTorch 模型创建：
+    1. 一个用于分发的半精度 (fp16) .bin 权重文件。
+    2. 一个与全精度 (fp32) 布局兼容的 ONNX 模型。
+    3. 一个可以将 fp16 .bin 文件还原为 fp32 .bin 的工具函数。
+    """
+
+    def __init__(self,
+                 torch_pth_path: str,
+                 vits_onnx_path: str,
+                 key_list_file: str,
+                 output_dir: str,
+                 cache_dir: str,
+                 ):
+        self.torch_pth_path: str = torch_pth_path
+        self.vits_onnx_path: str = vits_onnx_path
+        self.key_list_file: str = key_list_file
+        self.output_dir: str = output_dir
+        self.cache_dir: str = cache_dir
+        # 定义输出文件路径
+        self.fp16_bin_path: str = os.path.join(self.output_dir, "vits_fp16.bin")
+        self.index_table_path: str = os.path.join(self.cache_dir, "vits_weights_index_fp32.json")
+        self.relinked_fp32_onnx_path: str = os.path.join(self.output_dir, "vits_fp32.onnx")
+        self.reconstructed_fp32_bin_path: str = os.path.join(self.output_dir, "vits_fp32.bin")
+
+        # 确保输出目录存在
+        os.makedirs(self.cache_dir, exist_ok=True)
+        os.makedirs(self.output_dir, exist_ok=True)
+
+        if not os.path.exists(self.key_list_file):
+            raise FileNotFoundError(f"Error: Key list file not found! Path: {self.key_list_file}")
+
+    def step1_create_fp16_bin_and_fp32_index(self):
+        """
+        (1) 创建一个半精度 (fp16) 的 .bin 文件，但生成一个
+            描述全精度 (fp32) 布局的索引表。
+        """
+        # 加载 key 列表
+        with open(self.key_list_file, 'r') as f:
+            onnx_keys = [line.strip() for line in f.readlines()]
+
+        # 加载 PyTorch 模型权重
+        torch_state_dict = load_sovits_model(self.torch_pth_path)['weight']
+
+        index_table = OrderedDict()
+        current_fp32_offset = 0
+
+        with open(self.fp16_bin_path, 'wb') as f_bin:
+            for onnx_key in onnx_keys:
+                torch_key = onnx_key[len("vq_model."):] if onnx_key.startswith("vq_model.") else onnx_key
+
+                torch_tensor = torch_state_dict.get(torch_key)
+                if torch_tensor is None:
+                    raise ValueError(f"❌ Critical error: Key '{torch_key}' not found in the PyTorch weights")
+
+                # 转换为 fp16 并写入文件
+                torch_tensor_fp16 = torch_tensor.to(torch.float16)
+                numpy_array_fp16 = torch_tensor_fp16.cpu().numpy()
+                tensor_bytes_fp16 = numpy_array_fp16.tobytes()
+                f_bin.write(tensor_bytes_fp16)
+                tensor_length_fp32 = len(tensor_bytes_fp16) * 2
+                index_table[onnx_key] = {
+                    'offset': current_fp32_offset,
+                    'length': tensor_length_fp32
+                }
+                current_fp32_offset += tensor_length_fp32
+
+        # 保存描述 fp32 布局的索引表
+        with open(self.index_table_path, 'w') as f_json:
+            json.dump(index_table, f_json, indent=4)  # type: ignore
+
+    def step2_relink_onnx_for_fp32(self):
+        """
+        (2) 根据 fp32 索引表，修改 ONNX 模型，使其链接到一个
+            未来的、全精度的 .bin 文件。
+        """
+        # 加载描述 fp32 布局的索引表
+        with open(self.index_table_path, 'r') as f:
+            index_table = json.load(f)
+
+        model = onnx.load_model(self.vits_onnx_path, load_external_data=False)
+        reconstructed_bin_filename = os.path.basename(self.reconstructed_fp32_bin_path)
+
+        for tensor in model.graph.initializer:
+            if tensor.name in index_table:
+                tensor.ClearField('raw_data')
+                tensor.data_location = onnx.TensorProto.EXTERNAL
+                info = index_table[tensor.name]
+
+                del tensor.external_data[:]
+
+                keys = ["location", "offset", "length"]
+                values = [reconstructed_bin_filename, str(info['offset']), str(info['length'])]
+
+                for k, v in zip(keys, values):
+                    entry = tensor.external_data.add()
+                    entry.key = k
+                    entry.value = v
+
+        # 保存修改后的、链接到 fp32 权重的 ONNX 模型
+        onnx.save(model, self.relinked_fp32_onnx_path)
+
+    @staticmethod
+    def step3_reconstruct_fp32_bin_from_fp16(fp16_bin_path: str, output_fp32_bin_path: str):
+        """
+        (3) 静态工具函数：从半精度 .bin 文件还原出全精度 .bin 文件。
+
+        Args:
+            fp16_bin_path (str): 输入的半精度 .bin 文件路径。
+            output_fp32_bin_path (str): 输出的全精度 .bin 文件路径。
+        """
+        fp16_array = np.fromfile(fp16_bin_path, dtype=np.float16)
+        fp32_array = fp16_array.astype(np.float32)
+        fp32_array.tofile(output_fp32_bin_path)
+
+    def run_full_process(self):
+        self.step1_create_fp16_bin_and_fp32_index()
+        self.step2_relink_onnx_for_fp32()
diff --git a/genie_tts/Converter/v2/__pycache__/Converter.cpython-311.pyc b/genie_tts/Converter/v2/__pycache__/Converter.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3a4ffe154b36e0fb80500290bfc782cf33672ae5
Binary files /dev/null and b/genie_tts/Converter/v2/__pycache__/Converter.cpython-311.pyc differ
diff --git a/genie_tts/Converter/v2/__pycache__/EncoderConverter.cpython-311.pyc b/genie_tts/Converter/v2/__pycache__/EncoderConverter.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d66294d6f6ebb839a254efccd12d963b34790ff4
Binary files /dev/null and b/genie_tts/Converter/v2/__pycache__/EncoderConverter.cpython-311.pyc differ
diff --git a/genie_tts/Converter/v2/__pycache__/T2SConverter.cpython-311.pyc b/genie_tts/Converter/v2/__pycache__/T2SConverter.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..7414f33a9f18071f8033f23fe76c535a9474b55c
Binary files /dev/null and b/genie_tts/Converter/v2/__pycache__/T2SConverter.cpython-311.pyc differ
diff --git a/genie_tts/Converter/v2/__pycache__/VITSConverter.cpython-311.pyc b/genie_tts/Converter/v2/__pycache__/VITSConverter.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..86a00f5771faa85cce35519fc8f341e6c640655c
Binary files /dev/null and b/genie_tts/Converter/v2/__pycache__/VITSConverter.cpython-311.pyc differ
diff --git a/genie_tts/Converter/v2/__pycache__/__init__.cpython-311.pyc b/genie_tts/Converter/v2/__pycache__/__init__.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..be74ed9afd7309856f38d6340e7b3f79fab95134
Binary files /dev/null and b/genie_tts/Converter/v2/__pycache__/__init__.cpython-311.pyc differ
diff --git a/genie_tts/Converter/v2ProPlus/Converter.py b/genie_tts/Converter/v2ProPlus/Converter.py
index 9fbac43d89704024911e8618a13413bce834ddc2..7f275a2e96ed410c25ea2d9cc320c24d4f9babd9 100644
--- a/genie_tts/Converter/v2ProPlus/Converter.py
+++ b/genie_tts/Converter/v2ProPlus/Converter.py
@@ -1,89 +1,89 @@
-import logging
-import traceback
-import os
-import contextlib
-import importlib.resources
-
-from ...Utils.Constants import PACKAGE_NAME
-from ..v2.VITSConverter import VITSConverter
-from ..v2.T2SConverter import T2SModelConverter
-from ..v2.EncoderConverter import EncoderConverter
-from ..v2.Converter import (ENCODER_RESOURCE_PATH, STAGE_DECODER_RESOURCE_PATH,
-                            FIRST_STAGE_DECODER_RESOURCE_PATH, T2S_KEYS_RESOURCE_PATH, CACHE_DIR, remove_folder)
-from .PromptEncoderConverter import PromptEncoderConverter
-
-logger = logging.getLogger()
-
-# 使用 V2 ProPlus 的文件。
-VITS_RESOURCE_PATH = "Data/v2ProPlus/Models/vits_fp32.onnx"
-PROMPT_ENCODER_RESOURCE_PATH = "Data/v2ProPlus/Models/prompt_encoder_fp32.onnx"
-VITS_KEYS_RESOURCE_PATH = "Data/v2ProPlus/Keys/vits_weights.txt"
-PROMPT_ENCODER_KEYS_RESOURCE_PATH = "Data/v2ProPlus/Keys/prompt_encoder_weights.txt"
-
-
-def convert(torch_ckpt_path: str, torch_pth_path: str, output_dir: str) -> None:
-    # 确保缓存和输出目录存在
-    os.makedirs(CACHE_DIR, exist_ok=True)
-    os.makedirs(output_dir, exist_ok=True)
-
-    if len(os.listdir(output_dir)) > 0:
-        logger.warning(f"The output directory {output_dir} is not empty!")
-
-    with contextlib.ExitStack() as stack:
-        files = importlib.resources.files(PACKAGE_NAME)
-
-        def enter(p: str) -> str:
-            return str(stack.enter_context(importlib.resources.as_file(files.joinpath(p))))
-
-        encoder_onnx_path = enter(ENCODER_RESOURCE_PATH)
-        stage_decoder_path = enter(STAGE_DECODER_RESOURCE_PATH)
-        first_stage_decoder_path = enter(FIRST_STAGE_DECODER_RESOURCE_PATH)
-        vits_onnx_path = enter(VITS_RESOURCE_PATH)
-        t2s_keys_path = enter(T2S_KEYS_RESOURCE_PATH)
-        vits_keys_path = enter(VITS_KEYS_RESOURCE_PATH)
-        prompt_encoder_path = enter(PROMPT_ENCODER_RESOURCE_PATH)
-        prompt_encoder_keys_path = enter(PROMPT_ENCODER_KEYS_RESOURCE_PATH)
-
-        converter_1 = T2SModelConverter(
-            torch_ckpt_path=torch_ckpt_path,
-            stage_decoder_onnx_path=stage_decoder_path,
-            first_stage_decoder_onnx_path=first_stage_decoder_path,
-            key_list_file=t2s_keys_path,
-            output_dir=output_dir,
-            cache_dir=CACHE_DIR,
-        )
-        converter_2 = VITSConverter(
-            torch_pth_path=torch_pth_path,
-            vits_onnx_path=vits_onnx_path,
-            key_list_file=vits_keys_path,
-            output_dir=output_dir,
-            cache_dir=CACHE_DIR,
-        )
-        converter_3 = EncoderConverter(
-            ckpt_path=torch_ckpt_path,
-            pth_path=torch_pth_path,
-            onnx_input_path=encoder_onnx_path,
-            output_dir=output_dir,
-        )
-        converter_4 = PromptEncoderConverter(
-            torch_pth_path=torch_pth_path,
-            prompt_encoder_onnx_path=prompt_encoder_path,
-            key_list_file=prompt_encoder_keys_path,
-            output_dir=output_dir,
-            cache_dir=CACHE_DIR,
-        )
-
-        try:
-            converter_1.run_full_process()
-            converter_2.run_full_process()
-            converter_3.run_full_process()
-            converter_4.run_full_process()
-            logger.info(f"🎉 Conversion successful! Saved to: {os.path.abspath(output_dir)}\n"
-                        f"- Model Type: V2ProPlus")
-        except Exception:
-            logger.error(f"❌ A critical error occurred during the conversion process")
-            logger.error(traceback.format_exc())
-            remove_folder(output_dir)  # 只在失败时清理输出目录
-        finally:
-            # 无论成功还是失败，都尝试清理缓存目录
-            remove_folder(CACHE_DIR)
+import logging
+import traceback
+import os
+import contextlib
+import importlib.resources
+
+from ...Utils.Constants import PACKAGE_NAME
+from ..v2.VITSConverter import VITSConverter
+from ..v2.T2SConverter import T2SModelConverter
+from ..v2.EncoderConverter import EncoderConverter
+from ..v2.Converter import (ENCODER_RESOURCE_PATH, STAGE_DECODER_RESOURCE_PATH,
+                            FIRST_STAGE_DECODER_RESOURCE_PATH, T2S_KEYS_RESOURCE_PATH, CACHE_DIR, remove_folder)
+from .PromptEncoderConverter import PromptEncoderConverter
+
+logger = logging.getLogger()
+
+# 使用 V2 ProPlus 的文件。
+VITS_RESOURCE_PATH = "Data/v2ProPlus/Models/vits_fp32.onnx"
+PROMPT_ENCODER_RESOURCE_PATH = "Data/v2ProPlus/Models/prompt_encoder_fp32.onnx"
+VITS_KEYS_RESOURCE_PATH = "./Data/v2ProPlus/Keys/vits_weights.txt"
+PROMPT_ENCODER_KEYS_RESOURCE_PATH = "./Data/v2ProPlus/Keys/prompt_encoder_weights.txt"
+
+
+def convert(torch_ckpt_path: str, torch_pth_path: str, output_dir: str) -> None:
+    # 确保缓存和输出目录存在
+    os.makedirs(CACHE_DIR, exist_ok=True)
+    os.makedirs(output_dir, exist_ok=True)
+
+    if len(os.listdir(output_dir)) > 0:
+        logger.warning(f"The output directory {output_dir} is not empty!")
+
+    with contextlib.ExitStack() as stack:
+        files = importlib.resources.files(PACKAGE_NAME)
+
+        def enter(p: str) -> str:
+            return str(stack.enter_context(importlib.resources.as_file(files.joinpath(p))))
+
+        encoder_onnx_path = enter(ENCODER_RESOURCE_PATH)
+        stage_decoder_path = enter(STAGE_DECODER_RESOURCE_PATH)
+        first_stage_decoder_path = enter(FIRST_STAGE_DECODER_RESOURCE_PATH)
+        vits_onnx_path = enter(VITS_RESOURCE_PATH)
+        t2s_keys_path = enter(T2S_KEYS_RESOURCE_PATH)
+        vits_keys_path = enter(VITS_KEYS_RESOURCE_PATH)
+        prompt_encoder_path = enter(PROMPT_ENCODER_RESOURCE_PATH)
+        prompt_encoder_keys_path = enter(PROMPT_ENCODER_KEYS_RESOURCE_PATH)
+
+        converter_1 = T2SModelConverter(
+            torch_ckpt_path=torch_ckpt_path,
+            stage_decoder_onnx_path=stage_decoder_path,
+            first_stage_decoder_onnx_path=first_stage_decoder_path,
+            key_list_file=t2s_keys_path,
+            output_dir=output_dir,
+            cache_dir=CACHE_DIR,
+        )
+        converter_2 = VITSConverter(
+            torch_pth_path=torch_pth_path,
+            vits_onnx_path=vits_onnx_path,
+            key_list_file=vits_keys_path,
+            output_dir=output_dir,
+            cache_dir=CACHE_DIR,
+        )
+        converter_3 = EncoderConverter(
+            ckpt_path=torch_ckpt_path,
+            pth_path=torch_pth_path,
+            onnx_input_path=encoder_onnx_path,
+            output_dir=output_dir,
+        )
+        converter_4 = PromptEncoderConverter(
+            torch_pth_path=torch_pth_path,
+            prompt_encoder_onnx_path=prompt_encoder_path,
+            key_list_file=prompt_encoder_keys_path,
+            output_dir=output_dir,
+            cache_dir=CACHE_DIR,
+        )
+
+        try:
+            converter_1.run_full_process()
+            converter_2.run_full_process()
+            converter_3.run_full_process()
+            converter_4.run_full_process()
+            logger.info(f"🎉 Conversion successful! Saved to: {os.path.abspath(output_dir)}\n"
+                        f"- Model Type: V2ProPlus")
+        except Exception:
+            logger.error(f"❌ A critical error occurred during the conversion process")
+            logger.error(traceback.format_exc())
+            remove_folder(output_dir)  # 只在失败时清理输出目录
+        finally:
+            # 无论成功还是失败，都尝试清理缓存目录
+            remove_folder(CACHE_DIR)
diff --git a/genie_tts/Converter/v2ProPlus/PromptEncoderConverter.py b/genie_tts/Converter/v2ProPlus/PromptEncoderConverter.py
index 55bac8664c15b2bbaae6fcfc71f13f4feb8b4b1d..395645d87ac20615b834f00d86e8932cdf6f1d49 100644
--- a/genie_tts/Converter/v2ProPlus/PromptEncoderConverter.py
+++ b/genie_tts/Converter/v2ProPlus/PromptEncoderConverter.py
@@ -1,128 +1,128 @@
-import torch
-import onnx
-import json
-import os
-from collections import OrderedDict
-
-from ..load_state_dict import load_sovits_model
-
-
-class PromptEncoderConverter:
-    """
-    一个转换器，用于从 PyTorch 模型创建：
-    1. 一个用于分发的半精度 (fp16) .bin 权重文件。
-    2. 一个与全精度 (fp32) 布局兼容的 ONNX 模型。
-    3. 一个可以将 fp16 .bin 文件还原为 fp32 .bin 的工具函数。
-    """
-
-    def __init__(self,
-                 torch_pth_path: str,
-                 prompt_encoder_onnx_path: str,
-                 key_list_file: str,
-                 output_dir: str,
-                 cache_dir: str,
-                 ):
-        self.torch_pth_path: str = torch_pth_path
-        self.vits_onnx_path: str = prompt_encoder_onnx_path
-        self.key_list_file: str = key_list_file
-        self.output_dir: str = output_dir
-        self.cache_dir: str = cache_dir
-        # 定义输出文件路径
-        self.fp16_bin_path: str = os.path.join(self.output_dir, "prompt_encoder_fp16.bin")
-        self.index_table_path: str = os.path.join(self.cache_dir, "prompt_encoder_weights_index_fp32.json")
-        self.relinked_fp32_onnx_path: str = os.path.join(self.output_dir, "prompt_encoder_fp32.onnx")
-        self.reconstructed_fp32_bin_path: str = os.path.join(self.output_dir, "prompt_encoder_fp32.bin")
-
-        # 确保输出目录存在
-        os.makedirs(self.cache_dir, exist_ok=True)
-        os.makedirs(self.output_dir, exist_ok=True)
-
-        if not os.path.exists(self.key_list_file):
-            raise FileNotFoundError(f"错误: Key 列表文件未找到! 路径: {self.key_list_file}")
-
-    def step1_create_fp16_bin_and_fp32_index(self):
-        """
-        (1) 创建一个半精度 (fp16) 的 .bin 文件，但生成一个
-            描述全精度 (fp32) 布局的索引表。
-        """
-        # 加载 key 列表
-        with open(self.key_list_file, 'r') as f:
-            onnx_keys = [line.strip() for line in f.readlines()]
-
-        # 加载 PyTorch 模型权重
-        torch_state_dict = load_sovits_model(self.torch_pth_path)['weight']
-
-        index_table = OrderedDict()
-        # 这个偏移量将按照 fp32 的大小进行累加
-        current_fp32_offset = 0
-
-        with open(self.fp16_bin_path, 'wb') as f_bin:
-            for onnx_key in onnx_keys:
-                torch_key = onnx_key[len("vq_model."):] if onnx_key.startswith("vq_model.") else onnx_key
-
-                torch_tensor = torch_state_dict.get(torch_key)
-                if torch_tensor is None:
-                    raise ValueError(f"❌ 严重错误: 在 PyTorch 权重中找不到 Key '{torch_key}'")
-
-                # 转换为 fp16 并写入文件
-                torch_tensor_fp16 = torch_tensor.to(torch.float16)
-                numpy_array_fp16 = torch_tensor_fp16.cpu().numpy()
-                tensor_bytes_fp16 = numpy_array_fp16.tobytes()
-                f_bin.write(tensor_bytes_fp16)
-
-                # 关键步骤：计算并记录 fp32 的长度和偏移量
-                # 一个 fp32 = 4 字节, 一个 fp16 = 2 字节。所以 fp32 长度是 fp16 的两倍。
-                tensor_length_fp32 = len(tensor_bytes_fp16) * 2
-
-                index_table[onnx_key] = {
-                    'offset': current_fp32_offset,
-                    'length': tensor_length_fp32
-                }
-
-                # 偏移量也按照 fp32 的长度进行累加
-                current_fp32_offset += tensor_length_fp32
-
-        # 保存描述 fp32 布局的索引表
-        with open(self.index_table_path, 'w') as f_json:
-            json.dump(index_table, f_json, indent=4)  # type: ignore
-
-    def step2_relink_onnx_for_fp32(self):
-        """
-        (2) 根据 fp32 索引表，修改 ONNX 模型，使其链接到一个
-            未来的、全精度的 .bin 文件。
-        """
-        # 加载描述 fp32 布局的索引表
-        with open(self.index_table_path, 'r') as f:
-            index_table = json.load(f)
-
-        # 加载 ONNX 模型结构
-        model = onnx.load_model(self.vits_onnx_path, load_external_data=False)
-
-        # 这个 ONNX 模型将要链接的 .bin 文件名
-        reconstructed_bin_filename = os.path.basename(self.reconstructed_fp32_bin_path)
-
-        for tensor in model.graph.initializer:
-            if tensor.name in index_table:
-                tensor.ClearField('raw_data')
-                tensor.data_location = onnx.TensorProto.EXTERNAL
-                info = index_table[tensor.name]
-
-                del tensor.external_data[:]
-
-                keys = ["location", "offset", "length"]
-                values = [reconstructed_bin_filename, str(info['offset']), str(info['length'])]
-
-                for k, v in zip(keys, values):
-                    entry = tensor.external_data.add()
-                    entry.key = k
-                    entry.value = v
-
-        # 保存修改后的、链接到 fp32 权重的 ONNX 模型
-        onnx.save(model, self.relinked_fp32_onnx_path)
-
-    def run_full_process(self):
-        """
-        按顺序执行核心的转换步骤 (1 和 2)。
-        """
-        self.step1_create_fp16_bin_and_fp32_index()
-        self.step2_relink_onnx_for_fp32()
+import torch
+import onnx
+import json
+import os
+from collections import OrderedDict
+
+from ..load_state_dict import load_sovits_model
+
+
+class PromptEncoderConverter:
+    """
+    一个转换器，用于从 PyTorch 模型创建：
+    1. 一个用于分发的半精度 (fp16) .bin 权重文件。
+    2. 一个与全精度 (fp32) 布局兼容的 ONNX 模型。
+    3. 一个可以将 fp16 .bin 文件还原为 fp32 .bin 的工具函数。
+    """
+
+    def __init__(self,
+                 torch_pth_path: str,
+                 prompt_encoder_onnx_path: str,
+                 key_list_file: str,
+                 output_dir: str,
+                 cache_dir: str,
+                 ):
+        self.torch_pth_path: str = torch_pth_path
+        self.vits_onnx_path: str = prompt_encoder_onnx_path
+        self.key_list_file: str = key_list_file
+        self.output_dir: str = output_dir
+        self.cache_dir: str = cache_dir
+        # 定义输出文件路径
+        self.fp16_bin_path: str = os.path.join(self.output_dir, "prompt_encoder_fp16.bin")
+        self.index_table_path: str = os.path.join(self.cache_dir, "prompt_encoder_weights_index_fp32.json")
+        self.relinked_fp32_onnx_path: str = os.path.join(self.output_dir, "prompt_encoder_fp32.onnx")
+        self.reconstructed_fp32_bin_path: str = os.path.join(self.output_dir, "prompt_encoder_fp32.bin")
+
+        # 确保输出目录存在
+        os.makedirs(self.cache_dir, exist_ok=True)
+        os.makedirs(self.output_dir, exist_ok=True)
+
+        if not os.path.exists(self.key_list_file):
+            raise FileNotFoundError(f"错误: Key 列表文件未找到! 路径: {self.key_list_file}")
+
+    def step1_create_fp16_bin_and_fp32_index(self):
+        """
+        (1) 创建一个半精度 (fp16) 的 .bin 文件，但生成一个
+            描述全精度 (fp32) 布局的索引表。
+        """
+        # 加载 key 列表
+        with open(self.key_list_file, 'r') as f:
+            onnx_keys = [line.strip() for line in f.readlines()]
+
+        # 加载 PyTorch 模型权重
+        torch_state_dict = load_sovits_model(self.torch_pth_path)['weight']
+
+        index_table = OrderedDict()
+        # 这个偏移量将按照 fp32 的大小进行累加
+        current_fp32_offset = 0
+
+        with open(self.fp16_bin_path, 'wb') as f_bin:
+            for onnx_key in onnx_keys:
+                torch_key = onnx_key[len("vq_model."):] if onnx_key.startswith("vq_model.") else onnx_key
+
+                torch_tensor = torch_state_dict.get(torch_key)
+                if torch_tensor is None:
+                    raise ValueError(f"❌ 严重错误: 在 PyTorch 权重中找不到 Key '{torch_key}'")
+
+                # 转换为 fp16 并写入文件
+                torch_tensor_fp16 = torch_tensor.to(torch.float16)
+                numpy_array_fp16 = torch_tensor_fp16.cpu().numpy()
+                tensor_bytes_fp16 = numpy_array_fp16.tobytes()
+                f_bin.write(tensor_bytes_fp16)
+
+                # 关键步骤：计算并记录 fp32 的长度和偏移量
+                # 一个 fp32 = 4 字节, 一个 fp16 = 2 字节。所以 fp32 长度是 fp16 的两倍。
+                tensor_length_fp32 = len(tensor_bytes_fp16) * 2
+
+                index_table[onnx_key] = {
+                    'offset': current_fp32_offset,
+                    'length': tensor_length_fp32
+                }
+
+                # 偏移量也按照 fp32 的长度进行累加
+                current_fp32_offset += tensor_length_fp32
+
+        # 保存描述 fp32 布局的索引表
+        with open(self.index_table_path, 'w') as f_json:
+            json.dump(index_table, f_json, indent=4)  # type: ignore
+
+    def step2_relink_onnx_for_fp32(self):
+        """
+        (2) 根据 fp32 索引表，修改 ONNX 模型，使其链接到一个
+            未来的、全精度的 .bin 文件。
+        """
+        # 加载描述 fp32 布局的索引表
+        with open(self.index_table_path, 'r') as f:
+            index_table = json.load(f)
+
+        # 加载 ONNX 模型结构
+        model = onnx.load_model(self.vits_onnx_path, load_external_data=False)
+
+        # 这个 ONNX 模型将要链接的 .bin 文件名
+        reconstructed_bin_filename = os.path.basename(self.reconstructed_fp32_bin_path)
+
+        for tensor in model.graph.initializer:
+            if tensor.name in index_table:
+                tensor.ClearField('raw_data')
+                tensor.data_location = onnx.TensorProto.EXTERNAL
+                info = index_table[tensor.name]
+
+                del tensor.external_data[:]
+
+                keys = ["location", "offset", "length"]
+                values = [reconstructed_bin_filename, str(info['offset']), str(info['length'])]
+
+                for k, v in zip(keys, values):
+                    entry = tensor.external_data.add()
+                    entry.key = k
+                    entry.value = v
+
+        # 保存修改后的、链接到 fp32 权重的 ONNX 模型
+        onnx.save(model, self.relinked_fp32_onnx_path)
+
+    def run_full_process(self):
+        """
+        按顺序执行核心的转换步骤 (1 和 2)。
+        """
+        self.step1_create_fp16_bin_and_fp32_index()
+        self.step2_relink_onnx_for_fp32()
diff --git a/genie_tts/Converter/v2ProPlus/__pycache__/Converter.cpython-311.pyc b/genie_tts/Converter/v2ProPlus/__pycache__/Converter.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..53d4b93523656afa0dabd1351e944fd529539fd9
Binary files /dev/null and b/genie_tts/Converter/v2ProPlus/__pycache__/Converter.cpython-311.pyc differ
diff --git a/genie_tts/Converter/v2ProPlus/__pycache__/PromptEncoderConverter.cpython-311.pyc b/genie_tts/Converter/v2ProPlus/__pycache__/PromptEncoderConverter.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..34330bba3078569bbe4533fe7b042bc877739f19
Binary files /dev/null and b/genie_tts/Converter/v2ProPlus/__pycache__/PromptEncoderConverter.cpython-311.pyc differ
diff --git a/genie_tts/Core/Resources.py b/genie_tts/Core/Resources.py
index d049cec7c68450ed5abc59f0188838583fa053ab..71e99ead66abc53ad81ed36535a620f7d6f36e45 100644
--- a/genie_tts/Core/Resources.py
+++ b/genie_tts/Core/Resources.py
@@ -1,76 +1,76 @@
-import os
-from huggingface_hub import snapshot_download
-
-
-def download_genie_data() -> None:
-    print(f"🚀 Starting download Genie-TTS resources… This may take a few moments. ⏳")
-    snapshot_download(
-        repo_id="High-Logic/Genie",
-        repo_type="model",
-        allow_patterns="GenieData/*",
-        local_dir=".",
-        local_dir_use_symlinks=True,  # 软链接
-    )
-    print("✅ Genie-TTS resources downloaded successfully.")
-
-
-def ensure_exists(path: str, name: str):
-    if not os.path.exists(path):
-        raise FileNotFoundError(
-            f"Required directory or file '{name}' was not found at: {path}\n"
-            f"Please download the pretrained models and place them under './GenieData', "
-            f"or set the environment variable GENIE_DATA_DIR to the correct directory."
-        )
-
-
-"""
-文件结构与项目 Midori 同步。
-"""
-
-GENIE_DATA_DIR: str = os.getenv(
-    "GENIE_DATA_DIR",
-    "./GenieData"
-)
-
-"""
-Japanese_G2P_DIR: str = os.getenv(
-    "Japanese_G2P_DIR",
-    f"{GENIE_DATA_DIR}/G2P/JapaneseG2P"
-)
-"""
-
-English_G2P_DIR: str = os.getenv(
-    "English_G2P_DIR",
-    f"{GENIE_DATA_DIR}/G2P/EnglishG2P"
-)
-
-Chinese_G2P_DIR: str = os.getenv(
-    "Chinese_G2P_DIR",
-    f"{GENIE_DATA_DIR}/G2P/ChineseG2P"
-)
-
-HUBERT_MODEL_DIR: str = os.getenv(
-    "HUBERT_MODEL_DIR",
-    f"{GENIE_DATA_DIR}/chinese-hubert-base"
-)
-
-SV_MODEL: str = os.getenv(
-    "SV_MODEL",
-    f"{GENIE_DATA_DIR}/speaker_encoder.onnx"
-)
-
-ROBERTA_MODEL_DIR: str = os.getenv(
-    "ROBERTA_MODEL_DIR",
-    f"{GENIE_DATA_DIR}/RoBERTa"
-)
-
-if not os.path.exists(GENIE_DATA_DIR):
-    print("⚠️ GenieData folder not found.")
-    choice = input("Would you like to download it automatically from HuggingFace? (y/N): ").strip().lower()
-    if choice == "y":
-        download_genie_data()
-
-# ---- Run directory checks ----
-ensure_exists(HUBERT_MODEL_DIR, "HUBERT_MODEL_DIR")
-ensure_exists(SV_MODEL, "SV_MODEL")
-# ensure_exists(ROBERTA_MODEL_DIR, "ROBERTA_MODEL_DIR")
+import os
+from huggingface_hub import snapshot_download
+
+
+def download_genie_data() -> None:
+    print(f"🚀 Starting download Genie-TTS resources… This may take a few moments. ⏳")
+    snapshot_download(
+        repo_id="High-Logic/Genie",
+        repo_type="model",
+        allow_patterns="GenieData/*",
+        local_dir=".",
+        local_dir_use_symlinks=True,  # 软链接
+    )
+    print("✅ Genie-TTS resources downloaded successfully.")
+
+
+def ensure_exists(path: str, name: str):
+    if not os.path.exists(path):
+        raise FileNotFoundError(
+            f"Required directory or file '{name}' was not found at: {path}\n"
+            f"Please download the pretrained models and place them under './GenieData', "
+            f"or set the environment variable GENIE_DATA_DIR to the correct directory."
+        )
+
+
+"""
+文件结构与项目 Midori 同步。
+"""
+
+GENIE_DATA_DIR: str = os.getenv(
+    "GENIE_DATA_DIR",
+    "./GenieData"
+)
+
+"""
+Japanese_G2P_DIR: str = os.getenv(
+    "Japanese_G2P_DIR",
+    f"{GENIE_DATA_DIR}/G2P/JapaneseG2P"
+)
+"""
+
+English_G2P_DIR: str = os.getenv(
+    "English_G2P_DIR",
+    f"{GENIE_DATA_DIR}/G2P/EnglishG2P"
+)
+
+Chinese_G2P_DIR: str = os.getenv(
+    "Chinese_G2P_DIR",
+    f"{GENIE_DATA_DIR}/G2P/ChineseG2P"
+)
+
+HUBERT_MODEL_DIR: str = os.getenv(
+    "HUBERT_MODEL_DIR",
+    f"{GENIE_DATA_DIR}/chinese-hubert-base"
+)
+
+SV_MODEL: str = os.getenv(
+    "SV_MODEL",
+    f"{GENIE_DATA_DIR}/speaker_encoder.onnx"
+)
+
+ROBERTA_MODEL_DIR: str = os.getenv(
+    "ROBERTA_MODEL_DIR",
+    f"{GENIE_DATA_DIR}/RoBERTa"
+)
+
+if not os.path.exists(GENIE_DATA_DIR):
+    print("⚠️ GenieData folder not found.")
+    choice = input("Would you like to download it automatically from HuggingFace? (y/N): ").strip().lower()
+    if choice == "y":
+        download_genie_data()
+
+# ---- Run directory checks ----
+ensure_exists(HUBERT_MODEL_DIR, "HUBERT_MODEL_DIR")
+ensure_exists(SV_MODEL, "SV_MODEL")
+# ensure_exists(ROBERTA_MODEL_DIR, "ROBERTA_MODEL_DIR")
diff --git a/genie_tts/Core/__pycache__/Inference.cpython-311.pyc b/genie_tts/Core/__pycache__/Inference.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..fc408b34c0cb7f1447e50ef34fe72c694ee89476
Binary files /dev/null and b/genie_tts/Core/__pycache__/Inference.cpython-311.pyc differ
diff --git a/genie_tts/Core/__pycache__/Resources.cpython-311.pyc b/genie_tts/Core/__pycache__/Resources.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..40cc6cd804aa4e5079a184dacb63d9ff9468a1db
Binary files /dev/null and b/genie_tts/Core/__pycache__/Resources.cpython-311.pyc differ
diff --git a/genie_tts/Core/__pycache__/TTSPlayer.cpython-311.pyc b/genie_tts/Core/__pycache__/TTSPlayer.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c673c0ffefc1c5f4c838879fed67cbdafae5a33d
Binary files /dev/null and b/genie_tts/Core/__pycache__/TTSPlayer.cpython-311.pyc differ
diff --git a/genie_tts/Core/__pycache__/__init__.cpython-311.pyc b/genie_tts/Core/__pycache__/__init__.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3078343b5ae3e6d139eae4f10e6c7f9e39dbcab4
Binary files /dev/null and b/genie_tts/Core/__pycache__/__init__.cpython-311.pyc differ
diff --git a/genie_tts/Data/v2/Keys/t2s_onnx_keys.txt b/genie_tts/Data/v2/Keys/t2s_onnx_keys.txt
index 0a0b35764b2f696e8b72b6c68809ac463d55987a..0ff9ff6adde7fca9248e2d33de4ddaa32727690b 100644
--- a/genie_tts/Data/v2/Keys/t2s_onnx_keys.txt
+++ b/genie_tts/Data/v2/Keys/t2s_onnx_keys.txt
@@ -1,291 +1,291 @@
-ar_audio_embedding.word_embeddings.weight
-ar_audio_position.alpha
-transformer_encoder.layers.0.self_attn.in_proj_weight
-transformer_encoder.layers.0.self_attn.in_proj_bias
-transformer_encoder.layers.0.self_attn.out_proj.weight
-transformer_encoder.layers.0.self_attn.out_proj.bias
-transformer_encoder.layers.0.linear1.weight
-transformer_encoder.layers.0.linear1.bias
-transformer_encoder.layers.0.linear2.weight
-transformer_encoder.layers.0.linear2.bias
-transformer_encoder.layers.0.norm1.weight
-transformer_encoder.layers.0.norm1.bias
-transformer_encoder.layers.0.norm2.weight
-transformer_encoder.layers.0.norm2.bias
-transformer_encoder.layers.1.self_attn.in_proj_weight
-transformer_encoder.layers.1.self_attn.in_proj_bias
-transformer_encoder.layers.1.self_attn.out_proj.weight
-transformer_encoder.layers.1.self_attn.out_proj.bias
-transformer_encoder.layers.1.linear1.weight
-transformer_encoder.layers.1.linear1.bias
-transformer_encoder.layers.1.linear2.weight
-transformer_encoder.layers.1.linear2.bias
-transformer_encoder.layers.1.norm1.weight
-transformer_encoder.layers.1.norm1.bias
-transformer_encoder.layers.1.norm2.weight
-transformer_encoder.layers.1.norm2.bias
-transformer_encoder.layers.2.self_attn.in_proj_weight
-transformer_encoder.layers.2.self_attn.in_proj_bias
-transformer_encoder.layers.2.self_attn.out_proj.weight
-transformer_encoder.layers.2.self_attn.out_proj.bias
-transformer_encoder.layers.2.linear1.weight
-transformer_encoder.layers.2.linear1.bias
-transformer_encoder.layers.2.linear2.weight
-transformer_encoder.layers.2.linear2.bias
-transformer_encoder.layers.2.norm1.weight
-transformer_encoder.layers.2.norm1.bias
-transformer_encoder.layers.2.norm2.weight
-transformer_encoder.layers.2.norm2.bias
-transformer_encoder.layers.3.self_attn.in_proj_weight
-transformer_encoder.layers.3.self_attn.in_proj_bias
-transformer_encoder.layers.3.self_attn.out_proj.weight
-transformer_encoder.layers.3.self_attn.out_proj.bias
-transformer_encoder.layers.3.linear1.weight
-transformer_encoder.layers.3.linear1.bias
-transformer_encoder.layers.3.linear2.weight
-transformer_encoder.layers.3.linear2.bias
-transformer_encoder.layers.3.norm1.weight
-transformer_encoder.layers.3.norm1.bias
-transformer_encoder.layers.3.norm2.weight
-transformer_encoder.layers.3.norm2.bias
-transformer_encoder.layers.4.self_attn.in_proj_weight
-transformer_encoder.layers.4.self_attn.in_proj_bias
-transformer_encoder.layers.4.self_attn.out_proj.weight
-transformer_encoder.layers.4.self_attn.out_proj.bias
-transformer_encoder.layers.4.linear1.weight
-transformer_encoder.layers.4.linear1.bias
-transformer_encoder.layers.4.linear2.weight
-transformer_encoder.layers.4.linear2.bias
-transformer_encoder.layers.4.norm1.weight
-transformer_encoder.layers.4.norm1.bias
-transformer_encoder.layers.4.norm2.weight
-transformer_encoder.layers.4.norm2.bias
-transformer_encoder.layers.5.self_attn.in_proj_weight
-transformer_encoder.layers.5.self_attn.in_proj_bias
-transformer_encoder.layers.5.self_attn.out_proj.weight
-transformer_encoder.layers.5.self_attn.out_proj.bias
-transformer_encoder.layers.5.linear1.weight
-transformer_encoder.layers.5.linear1.bias
-transformer_encoder.layers.5.linear2.weight
-transformer_encoder.layers.5.linear2.bias
-transformer_encoder.layers.5.norm1.weight
-transformer_encoder.layers.5.norm1.bias
-transformer_encoder.layers.5.norm2.weight
-transformer_encoder.layers.5.norm2.bias
-transformer_encoder.layers.6.self_attn.in_proj_weight
-transformer_encoder.layers.6.self_attn.in_proj_bias
-transformer_encoder.layers.6.self_attn.out_proj.weight
-transformer_encoder.layers.6.self_attn.out_proj.bias
-transformer_encoder.layers.6.linear1.weight
-transformer_encoder.layers.6.linear1.bias
-transformer_encoder.layers.6.linear2.weight
-transformer_encoder.layers.6.linear2.bias
-transformer_encoder.layers.6.norm1.weight
-transformer_encoder.layers.6.norm1.bias
-transformer_encoder.layers.6.norm2.weight
-transformer_encoder.layers.6.norm2.bias
-transformer_encoder.layers.7.self_attn.in_proj_weight
-transformer_encoder.layers.7.self_attn.in_proj_bias
-transformer_encoder.layers.7.self_attn.out_proj.weight
-transformer_encoder.layers.7.self_attn.out_proj.bias
-transformer_encoder.layers.7.linear1.weight
-transformer_encoder.layers.7.linear1.bias
-transformer_encoder.layers.7.linear2.weight
-transformer_encoder.layers.7.linear2.bias
-transformer_encoder.layers.7.norm1.weight
-transformer_encoder.layers.7.norm1.bias
-transformer_encoder.layers.7.norm2.weight
-transformer_encoder.layers.7.norm2.bias
-transformer_encoder.layers.8.self_attn.in_proj_weight
-transformer_encoder.layers.8.self_attn.in_proj_bias
-transformer_encoder.layers.8.self_attn.out_proj.weight
-transformer_encoder.layers.8.self_attn.out_proj.bias
-transformer_encoder.layers.8.linear1.weight
-transformer_encoder.layers.8.linear1.bias
-transformer_encoder.layers.8.linear2.weight
-transformer_encoder.layers.8.linear2.bias
-transformer_encoder.layers.8.norm1.weight
-transformer_encoder.layers.8.norm1.bias
-transformer_encoder.layers.8.norm2.weight
-transformer_encoder.layers.8.norm2.bias
-transformer_encoder.layers.9.self_attn.in_proj_weight
-transformer_encoder.layers.9.self_attn.in_proj_bias
-transformer_encoder.layers.9.self_attn.out_proj.weight
-transformer_encoder.layers.9.self_attn.out_proj.bias
-transformer_encoder.layers.9.linear1.weight
-transformer_encoder.layers.9.linear1.bias
-transformer_encoder.layers.9.linear2.weight
-transformer_encoder.layers.9.linear2.bias
-transformer_encoder.layers.9.norm1.weight
-transformer_encoder.layers.9.norm1.bias
-transformer_encoder.layers.9.norm2.weight
-transformer_encoder.layers.9.norm2.bias
-transformer_encoder.layers.10.self_attn.in_proj_weight
-transformer_encoder.layers.10.self_attn.in_proj_bias
-transformer_encoder.layers.10.self_attn.out_proj.weight
-transformer_encoder.layers.10.self_attn.out_proj.bias
-transformer_encoder.layers.10.linear1.weight
-transformer_encoder.layers.10.linear1.bias
-transformer_encoder.layers.10.linear2.weight
-transformer_encoder.layers.10.linear2.bias
-transformer_encoder.layers.10.norm1.weight
-transformer_encoder.layers.10.norm1.bias
-transformer_encoder.layers.10.norm2.weight
-transformer_encoder.layers.10.norm2.bias
-transformer_encoder.layers.11.self_attn.in_proj_weight
-transformer_encoder.layers.11.self_attn.in_proj_bias
-transformer_encoder.layers.11.self_attn.out_proj.weight
-transformer_encoder.layers.11.self_attn.out_proj.bias
-transformer_encoder.layers.11.linear1.weight
-transformer_encoder.layers.11.linear1.bias
-transformer_encoder.layers.11.linear2.weight
-transformer_encoder.layers.11.linear2.bias
-transformer_encoder.layers.11.norm1.weight
-transformer_encoder.layers.11.norm1.bias
-transformer_encoder.layers.11.norm2.weight
-transformer_encoder.layers.11.norm2.bias
-transformer_encoder.layers.12.self_attn.in_proj_weight
-transformer_encoder.layers.12.self_attn.in_proj_bias
-transformer_encoder.layers.12.self_attn.out_proj.weight
-transformer_encoder.layers.12.self_attn.out_proj.bias
-transformer_encoder.layers.12.linear1.weight
-transformer_encoder.layers.12.linear1.bias
-transformer_encoder.layers.12.linear2.weight
-transformer_encoder.layers.12.linear2.bias
-transformer_encoder.layers.12.norm1.weight
-transformer_encoder.layers.12.norm1.bias
-transformer_encoder.layers.12.norm2.weight
-transformer_encoder.layers.12.norm2.bias
-transformer_encoder.layers.13.self_attn.in_proj_weight
-transformer_encoder.layers.13.self_attn.in_proj_bias
-transformer_encoder.layers.13.self_attn.out_proj.weight
-transformer_encoder.layers.13.self_attn.out_proj.bias
-transformer_encoder.layers.13.linear1.weight
-transformer_encoder.layers.13.linear1.bias
-transformer_encoder.layers.13.linear2.weight
-transformer_encoder.layers.13.linear2.bias
-transformer_encoder.layers.13.norm1.weight
-transformer_encoder.layers.13.norm1.bias
-transformer_encoder.layers.13.norm2.weight
-transformer_encoder.layers.13.norm2.bias
-transformer_encoder.layers.14.self_attn.in_proj_weight
-transformer_encoder.layers.14.self_attn.in_proj_bias
-transformer_encoder.layers.14.self_attn.out_proj.weight
-transformer_encoder.layers.14.self_attn.out_proj.bias
-transformer_encoder.layers.14.linear1.weight
-transformer_encoder.layers.14.linear1.bias
-transformer_encoder.layers.14.linear2.weight
-transformer_encoder.layers.14.linear2.bias
-transformer_encoder.layers.14.norm1.weight
-transformer_encoder.layers.14.norm1.bias
-transformer_encoder.layers.14.norm2.weight
-transformer_encoder.layers.14.norm2.bias
-transformer_encoder.layers.15.self_attn.in_proj_weight
-transformer_encoder.layers.15.self_attn.in_proj_bias
-transformer_encoder.layers.15.self_attn.out_proj.weight
-transformer_encoder.layers.15.self_attn.out_proj.bias
-transformer_encoder.layers.15.linear1.weight
-transformer_encoder.layers.15.linear1.bias
-transformer_encoder.layers.15.linear2.weight
-transformer_encoder.layers.15.linear2.bias
-transformer_encoder.layers.15.norm1.weight
-transformer_encoder.layers.15.norm1.bias
-transformer_encoder.layers.15.norm2.weight
-transformer_encoder.layers.15.norm2.bias
-transformer_encoder.layers.16.self_attn.in_proj_weight
-transformer_encoder.layers.16.self_attn.in_proj_bias
-transformer_encoder.layers.16.self_attn.out_proj.weight
-transformer_encoder.layers.16.self_attn.out_proj.bias
-transformer_encoder.layers.16.linear1.weight
-transformer_encoder.layers.16.linear1.bias
-transformer_encoder.layers.16.linear2.weight
-transformer_encoder.layers.16.linear2.bias
-transformer_encoder.layers.16.norm1.weight
-transformer_encoder.layers.16.norm1.bias
-transformer_encoder.layers.16.norm2.weight
-transformer_encoder.layers.16.norm2.bias
-transformer_encoder.layers.17.self_attn.in_proj_weight
-transformer_encoder.layers.17.self_attn.in_proj_bias
-transformer_encoder.layers.17.self_attn.out_proj.weight
-transformer_encoder.layers.17.self_attn.out_proj.bias
-transformer_encoder.layers.17.linear1.weight
-transformer_encoder.layers.17.linear1.bias
-transformer_encoder.layers.17.linear2.weight
-transformer_encoder.layers.17.linear2.bias
-transformer_encoder.layers.17.norm1.weight
-transformer_encoder.layers.17.norm1.bias
-transformer_encoder.layers.17.norm2.weight
-transformer_encoder.layers.17.norm2.bias
-transformer_encoder.layers.18.self_attn.in_proj_weight
-transformer_encoder.layers.18.self_attn.in_proj_bias
-transformer_encoder.layers.18.self_attn.out_proj.weight
-transformer_encoder.layers.18.self_attn.out_proj.bias
-transformer_encoder.layers.18.linear1.weight
-transformer_encoder.layers.18.linear1.bias
-transformer_encoder.layers.18.linear2.weight
-transformer_encoder.layers.18.linear2.bias
-transformer_encoder.layers.18.norm1.weight
-transformer_encoder.layers.18.norm1.bias
-transformer_encoder.layers.18.norm2.weight
-transformer_encoder.layers.18.norm2.bias
-transformer_encoder.layers.19.self_attn.in_proj_weight
-transformer_encoder.layers.19.self_attn.in_proj_bias
-transformer_encoder.layers.19.self_attn.out_proj.weight
-transformer_encoder.layers.19.self_attn.out_proj.bias
-transformer_encoder.layers.19.linear1.weight
-transformer_encoder.layers.19.linear1.bias
-transformer_encoder.layers.19.linear2.weight
-transformer_encoder.layers.19.linear2.bias
-transformer_encoder.layers.19.norm1.weight
-transformer_encoder.layers.19.norm1.bias
-transformer_encoder.layers.19.norm2.weight
-transformer_encoder.layers.19.norm2.bias
-transformer_encoder.layers.20.self_attn.in_proj_weight
-transformer_encoder.layers.20.self_attn.in_proj_bias
-transformer_encoder.layers.20.self_attn.out_proj.weight
-transformer_encoder.layers.20.self_attn.out_proj.bias
-transformer_encoder.layers.20.linear1.weight
-transformer_encoder.layers.20.linear1.bias
-transformer_encoder.layers.20.linear2.weight
-transformer_encoder.layers.20.linear2.bias
-transformer_encoder.layers.20.norm1.weight
-transformer_encoder.layers.20.norm1.bias
-transformer_encoder.layers.20.norm2.weight
-transformer_encoder.layers.20.norm2.bias
-transformer_encoder.layers.21.self_attn.in_proj_weight
-transformer_encoder.layers.21.self_attn.in_proj_bias
-transformer_encoder.layers.21.self_attn.out_proj.weight
-transformer_encoder.layers.21.self_attn.out_proj.bias
-transformer_encoder.layers.21.linear1.weight
-transformer_encoder.layers.21.linear1.bias
-transformer_encoder.layers.21.linear2.weight
-transformer_encoder.layers.21.linear2.bias
-transformer_encoder.layers.21.norm1.weight
-transformer_encoder.layers.21.norm1.bias
-transformer_encoder.layers.21.norm2.weight
-transformer_encoder.layers.21.norm2.bias
-transformer_encoder.layers.22.self_attn.in_proj_weight
-transformer_encoder.layers.22.self_attn.in_proj_bias
-transformer_encoder.layers.22.self_attn.out_proj.weight
-transformer_encoder.layers.22.self_attn.out_proj.bias
-transformer_encoder.layers.22.linear1.weight
-transformer_encoder.layers.22.linear1.bias
-transformer_encoder.layers.22.linear2.weight
-transformer_encoder.layers.22.linear2.bias
-transformer_encoder.layers.22.norm1.weight
-transformer_encoder.layers.22.norm1.bias
-transformer_encoder.layers.22.norm2.weight
-transformer_encoder.layers.22.norm2.bias
-transformer_encoder.layers.23.self_attn.in_proj_weight
-transformer_encoder.layers.23.self_attn.in_proj_bias
-transformer_encoder.layers.23.self_attn.out_proj.weight
-transformer_encoder.layers.23.self_attn.out_proj.bias
-transformer_encoder.layers.23.linear1.weight
-transformer_encoder.layers.23.linear1.bias
-transformer_encoder.layers.23.linear2.weight
-transformer_encoder.layers.23.linear2.bias
-transformer_encoder.layers.23.norm1.weight
-transformer_encoder.layers.23.norm1.bias
-transformer_encoder.layers.23.norm2.weight
-transformer_encoder.layers.23.norm2.bias
-ar_predict_layer.weight
+ar_audio_embedding.word_embeddings.weight
+ar_audio_position.alpha
+transformer_encoder.layers.0.self_attn.in_proj_weight
+transformer_encoder.layers.0.self_attn.in_proj_bias
+transformer_encoder.layers.0.self_attn.out_proj.weight
+transformer_encoder.layers.0.self_attn.out_proj.bias
+transformer_encoder.layers.0.linear1.weight
+transformer_encoder.layers.0.linear1.bias
+transformer_encoder.layers.0.linear2.weight
+transformer_encoder.layers.0.linear2.bias
+transformer_encoder.layers.0.norm1.weight
+transformer_encoder.layers.0.norm1.bias
+transformer_encoder.layers.0.norm2.weight
+transformer_encoder.layers.0.norm2.bias
+transformer_encoder.layers.1.self_attn.in_proj_weight
+transformer_encoder.layers.1.self_attn.in_proj_bias
+transformer_encoder.layers.1.self_attn.out_proj.weight
+transformer_encoder.layers.1.self_attn.out_proj.bias
+transformer_encoder.layers.1.linear1.weight
+transformer_encoder.layers.1.linear1.bias
+transformer_encoder.layers.1.linear2.weight
+transformer_encoder.layers.1.linear2.bias
+transformer_encoder.layers.1.norm1.weight
+transformer_encoder.layers.1.norm1.bias
+transformer_encoder.layers.1.norm2.weight
+transformer_encoder.layers.1.norm2.bias
+transformer_encoder.layers.2.self_attn.in_proj_weight
+transformer_encoder.layers.2.self_attn.in_proj_bias
+transformer_encoder.layers.2.self_attn.out_proj.weight
+transformer_encoder.layers.2.self_attn.out_proj.bias
+transformer_encoder.layers.2.linear1.weight
+transformer_encoder.layers.2.linear1.bias
+transformer_encoder.layers.2.linear2.weight
+transformer_encoder.layers.2.linear2.bias
+transformer_encoder.layers.2.norm1.weight
+transformer_encoder.layers.2.norm1.bias
+transformer_encoder.layers.2.norm2.weight
+transformer_encoder.layers.2.norm2.bias
+transformer_encoder.layers.3.self_attn.in_proj_weight
+transformer_encoder.layers.3.self_attn.in_proj_bias
+transformer_encoder.layers.3.self_attn.out_proj.weight
+transformer_encoder.layers.3.self_attn.out_proj.bias
+transformer_encoder.layers.3.linear1.weight
+transformer_encoder.layers.3.linear1.bias
+transformer_encoder.layers.3.linear2.weight
+transformer_encoder.layers.3.linear2.bias
+transformer_encoder.layers.3.norm1.weight
+transformer_encoder.layers.3.norm1.bias
+transformer_encoder.layers.3.norm2.weight
+transformer_encoder.layers.3.norm2.bias
+transformer_encoder.layers.4.self_attn.in_proj_weight
+transformer_encoder.layers.4.self_attn.in_proj_bias
+transformer_encoder.layers.4.self_attn.out_proj.weight
+transformer_encoder.layers.4.self_attn.out_proj.bias
+transformer_encoder.layers.4.linear1.weight
+transformer_encoder.layers.4.linear1.bias
+transformer_encoder.layers.4.linear2.weight
+transformer_encoder.layers.4.linear2.bias
+transformer_encoder.layers.4.norm1.weight
+transformer_encoder.layers.4.norm1.bias
+transformer_encoder.layers.4.norm2.weight
+transformer_encoder.layers.4.norm2.bias
+transformer_encoder.layers.5.self_attn.in_proj_weight
+transformer_encoder.layers.5.self_attn.in_proj_bias
+transformer_encoder.layers.5.self_attn.out_proj.weight
+transformer_encoder.layers.5.self_attn.out_proj.bias
+transformer_encoder.layers.5.linear1.weight
+transformer_encoder.layers.5.linear1.bias
+transformer_encoder.layers.5.linear2.weight
+transformer_encoder.layers.5.linear2.bias
+transformer_encoder.layers.5.norm1.weight
+transformer_encoder.layers.5.norm1.bias
+transformer_encoder.layers.5.norm2.weight
+transformer_encoder.layers.5.norm2.bias
+transformer_encoder.layers.6.self_attn.in_proj_weight
+transformer_encoder.layers.6.self_attn.in_proj_bias
+transformer_encoder.layers.6.self_attn.out_proj.weight
+transformer_encoder.layers.6.self_attn.out_proj.bias
+transformer_encoder.layers.6.linear1.weight
+transformer_encoder.layers.6.linear1.bias
+transformer_encoder.layers.6.linear2.weight
+transformer_encoder.layers.6.linear2.bias
+transformer_encoder.layers.6.norm1.weight
+transformer_encoder.layers.6.norm1.bias
+transformer_encoder.layers.6.norm2.weight
+transformer_encoder.layers.6.norm2.bias
+transformer_encoder.layers.7.self_attn.in_proj_weight
+transformer_encoder.layers.7.self_attn.in_proj_bias
+transformer_encoder.layers.7.self_attn.out_proj.weight
+transformer_encoder.layers.7.self_attn.out_proj.bias
+transformer_encoder.layers.7.linear1.weight
+transformer_encoder.layers.7.linear1.bias
+transformer_encoder.layers.7.linear2.weight
+transformer_encoder.layers.7.linear2.bias
+transformer_encoder.layers.7.norm1.weight
+transformer_encoder.layers.7.norm1.bias
+transformer_encoder.layers.7.norm2.weight
+transformer_encoder.layers.7.norm2.bias
+transformer_encoder.layers.8.self_attn.in_proj_weight
+transformer_encoder.layers.8.self_attn.in_proj_bias
+transformer_encoder.layers.8.self_attn.out_proj.weight
+transformer_encoder.layers.8.self_attn.out_proj.bias
+transformer_encoder.layers.8.linear1.weight
+transformer_encoder.layers.8.linear1.bias
+transformer_encoder.layers.8.linear2.weight
+transformer_encoder.layers.8.linear2.bias
+transformer_encoder.layers.8.norm1.weight
+transformer_encoder.layers.8.norm1.bias
+transformer_encoder.layers.8.norm2.weight
+transformer_encoder.layers.8.norm2.bias
+transformer_encoder.layers.9.self_attn.in_proj_weight
+transformer_encoder.layers.9.self_attn.in_proj_bias
+transformer_encoder.layers.9.self_attn.out_proj.weight
+transformer_encoder.layers.9.self_attn.out_proj.bias
+transformer_encoder.layers.9.linear1.weight
+transformer_encoder.layers.9.linear1.bias
+transformer_encoder.layers.9.linear2.weight
+transformer_encoder.layers.9.linear2.bias
+transformer_encoder.layers.9.norm1.weight
+transformer_encoder.layers.9.norm1.bias
+transformer_encoder.layers.9.norm2.weight
+transformer_encoder.layers.9.norm2.bias
+transformer_encoder.layers.10.self_attn.in_proj_weight
+transformer_encoder.layers.10.self_attn.in_proj_bias
+transformer_encoder.layers.10.self_attn.out_proj.weight
+transformer_encoder.layers.10.self_attn.out_proj.bias
+transformer_encoder.layers.10.linear1.weight
+transformer_encoder.layers.10.linear1.bias
+transformer_encoder.layers.10.linear2.weight
+transformer_encoder.layers.10.linear2.bias
+transformer_encoder.layers.10.norm1.weight
+transformer_encoder.layers.10.norm1.bias
+transformer_encoder.layers.10.norm2.weight
+transformer_encoder.layers.10.norm2.bias
+transformer_encoder.layers.11.self_attn.in_proj_weight
+transformer_encoder.layers.11.self_attn.in_proj_bias
+transformer_encoder.layers.11.self_attn.out_proj.weight
+transformer_encoder.layers.11.self_attn.out_proj.bias
+transformer_encoder.layers.11.linear1.weight
+transformer_encoder.layers.11.linear1.bias
+transformer_encoder.layers.11.linear2.weight
+transformer_encoder.layers.11.linear2.bias
+transformer_encoder.layers.11.norm1.weight
+transformer_encoder.layers.11.norm1.bias
+transformer_encoder.layers.11.norm2.weight
+transformer_encoder.layers.11.norm2.bias
+transformer_encoder.layers.12.self_attn.in_proj_weight
+transformer_encoder.layers.12.self_attn.in_proj_bias
+transformer_encoder.layers.12.self_attn.out_proj.weight
+transformer_encoder.layers.12.self_attn.out_proj.bias
+transformer_encoder.layers.12.linear1.weight
+transformer_encoder.layers.12.linear1.bias
+transformer_encoder.layers.12.linear2.weight
+transformer_encoder.layers.12.linear2.bias
+transformer_encoder.layers.12.norm1.weight
+transformer_encoder.layers.12.norm1.bias
+transformer_encoder.layers.12.norm2.weight
+transformer_encoder.layers.12.norm2.bias
+transformer_encoder.layers.13.self_attn.in_proj_weight
+transformer_encoder.layers.13.self_attn.in_proj_bias
+transformer_encoder.layers.13.self_attn.out_proj.weight
+transformer_encoder.layers.13.self_attn.out_proj.bias
+transformer_encoder.layers.13.linear1.weight
+transformer_encoder.layers.13.linear1.bias
+transformer_encoder.layers.13.linear2.weight
+transformer_encoder.layers.13.linear2.bias
+transformer_encoder.layers.13.norm1.weight
+transformer_encoder.layers.13.norm1.bias
+transformer_encoder.layers.13.norm2.weight
+transformer_encoder.layers.13.norm2.bias
+transformer_encoder.layers.14.self_attn.in_proj_weight
+transformer_encoder.layers.14.self_attn.in_proj_bias
+transformer_encoder.layers.14.self_attn.out_proj.weight
+transformer_encoder.layers.14.self_attn.out_proj.bias
+transformer_encoder.layers.14.linear1.weight
+transformer_encoder.layers.14.linear1.bias
+transformer_encoder.layers.14.linear2.weight
+transformer_encoder.layers.14.linear2.bias
+transformer_encoder.layers.14.norm1.weight
+transformer_encoder.layers.14.norm1.bias
+transformer_encoder.layers.14.norm2.weight
+transformer_encoder.layers.14.norm2.bias
+transformer_encoder.layers.15.self_attn.in_proj_weight
+transformer_encoder.layers.15.self_attn.in_proj_bias
+transformer_encoder.layers.15.self_attn.out_proj.weight
+transformer_encoder.layers.15.self_attn.out_proj.bias
+transformer_encoder.layers.15.linear1.weight
+transformer_encoder.layers.15.linear1.bias
+transformer_encoder.layers.15.linear2.weight
+transformer_encoder.layers.15.linear2.bias
+transformer_encoder.layers.15.norm1.weight
+transformer_encoder.layers.15.norm1.bias
+transformer_encoder.layers.15.norm2.weight
+transformer_encoder.layers.15.norm2.bias
+transformer_encoder.layers.16.self_attn.in_proj_weight
+transformer_encoder.layers.16.self_attn.in_proj_bias
+transformer_encoder.layers.16.self_attn.out_proj.weight
+transformer_encoder.layers.16.self_attn.out_proj.bias
+transformer_encoder.layers.16.linear1.weight
+transformer_encoder.layers.16.linear1.bias
+transformer_encoder.layers.16.linear2.weight
+transformer_encoder.layers.16.linear2.bias
+transformer_encoder.layers.16.norm1.weight
+transformer_encoder.layers.16.norm1.bias
+transformer_encoder.layers.16.norm2.weight
+transformer_encoder.layers.16.norm2.bias
+transformer_encoder.layers.17.self_attn.in_proj_weight
+transformer_encoder.layers.17.self_attn.in_proj_bias
+transformer_encoder.layers.17.self_attn.out_proj.weight
+transformer_encoder.layers.17.self_attn.out_proj.bias
+transformer_encoder.layers.17.linear1.weight
+transformer_encoder.layers.17.linear1.bias
+transformer_encoder.layers.17.linear2.weight
+transformer_encoder.layers.17.linear2.bias
+transformer_encoder.layers.17.norm1.weight
+transformer_encoder.layers.17.norm1.bias
+transformer_encoder.layers.17.norm2.weight
+transformer_encoder.layers.17.norm2.bias
+transformer_encoder.layers.18.self_attn.in_proj_weight
+transformer_encoder.layers.18.self_attn.in_proj_bias
+transformer_encoder.layers.18.self_attn.out_proj.weight
+transformer_encoder.layers.18.self_attn.out_proj.bias
+transformer_encoder.layers.18.linear1.weight
+transformer_encoder.layers.18.linear1.bias
+transformer_encoder.layers.18.linear2.weight
+transformer_encoder.layers.18.linear2.bias
+transformer_encoder.layers.18.norm1.weight
+transformer_encoder.layers.18.norm1.bias
+transformer_encoder.layers.18.norm2.weight
+transformer_encoder.layers.18.norm2.bias
+transformer_encoder.layers.19.self_attn.in_proj_weight
+transformer_encoder.layers.19.self_attn.in_proj_bias
+transformer_encoder.layers.19.self_attn.out_proj.weight
+transformer_encoder.layers.19.self_attn.out_proj.bias
+transformer_encoder.layers.19.linear1.weight
+transformer_encoder.layers.19.linear1.bias
+transformer_encoder.layers.19.linear2.weight
+transformer_encoder.layers.19.linear2.bias
+transformer_encoder.layers.19.norm1.weight
+transformer_encoder.layers.19.norm1.bias
+transformer_encoder.layers.19.norm2.weight
+transformer_encoder.layers.19.norm2.bias
+transformer_encoder.layers.20.self_attn.in_proj_weight
+transformer_encoder.layers.20.self_attn.in_proj_bias
+transformer_encoder.layers.20.self_attn.out_proj.weight
+transformer_encoder.layers.20.self_attn.out_proj.bias
+transformer_encoder.layers.20.linear1.weight
+transformer_encoder.layers.20.linear1.bias
+transformer_encoder.layers.20.linear2.weight
+transformer_encoder.layers.20.linear2.bias
+transformer_encoder.layers.20.norm1.weight
+transformer_encoder.layers.20.norm1.bias
+transformer_encoder.layers.20.norm2.weight
+transformer_encoder.layers.20.norm2.bias
+transformer_encoder.layers.21.self_attn.in_proj_weight
+transformer_encoder.layers.21.self_attn.in_proj_bias
+transformer_encoder.layers.21.self_attn.out_proj.weight
+transformer_encoder.layers.21.self_attn.out_proj.bias
+transformer_encoder.layers.21.linear1.weight
+transformer_encoder.layers.21.linear1.bias
+transformer_encoder.layers.21.linear2.weight
+transformer_encoder.layers.21.linear2.bias
+transformer_encoder.layers.21.norm1.weight
+transformer_encoder.layers.21.norm1.bias
+transformer_encoder.layers.21.norm2.weight
+transformer_encoder.layers.21.norm2.bias
+transformer_encoder.layers.22.self_attn.in_proj_weight
+transformer_encoder.layers.22.self_attn.in_proj_bias
+transformer_encoder.layers.22.self_attn.out_proj.weight
+transformer_encoder.layers.22.self_attn.out_proj.bias
+transformer_encoder.layers.22.linear1.weight
+transformer_encoder.layers.22.linear1.bias
+transformer_encoder.layers.22.linear2.weight
+transformer_encoder.layers.22.linear2.bias
+transformer_encoder.layers.22.norm1.weight
+transformer_encoder.layers.22.norm1.bias
+transformer_encoder.layers.22.norm2.weight
+transformer_encoder.layers.22.norm2.bias
+transformer_encoder.layers.23.self_attn.in_proj_weight
+transformer_encoder.layers.23.self_attn.in_proj_bias
+transformer_encoder.layers.23.self_attn.out_proj.weight
+transformer_encoder.layers.23.self_attn.out_proj.bias
+transformer_encoder.layers.23.linear1.weight
+transformer_encoder.layers.23.linear1.bias
+transformer_encoder.layers.23.linear2.weight
+transformer_encoder.layers.23.linear2.bias
+transformer_encoder.layers.23.norm1.weight
+transformer_encoder.layers.23.norm1.bias
+transformer_encoder.layers.23.norm2.weight
+transformer_encoder.layers.23.norm2.bias
+ar_predict_layer.weight
diff --git a/genie_tts/Data/v2/Keys/vits_onnx_keys.txt b/genie_tts/Data/v2/Keys/vits_onnx_keys.txt
index d4f0b4bbe0d8fd259d320e2f2c7967e0a92660a1..cf2afc6423f421fcbacf781f865cd2133b4cd930 100644
--- a/genie_tts/Data/v2/Keys/vits_onnx_keys.txt
+++ b/genie_tts/Data/v2/Keys/vits_onnx_keys.txt
@@ -1,668 +1,668 @@
-vq_model.dec.cond.bias
-vq_model.dec.cond.weight
-vq_model.dec.conv_post.weight
-vq_model.dec.conv_pre.bias
-vq_model.dec.conv_pre.weight
-vq_model.dec.resblocks.0.convs1.0.bias
-vq_model.dec.resblocks.0.convs1.0.weight_g
-vq_model.dec.resblocks.0.convs1.0.weight_v
-vq_model.dec.resblocks.0.convs1.1.bias
-vq_model.dec.resblocks.0.convs1.1.weight_g
-vq_model.dec.resblocks.0.convs1.1.weight_v
-vq_model.dec.resblocks.0.convs1.2.bias
-vq_model.dec.resblocks.0.convs1.2.weight_g
-vq_model.dec.resblocks.0.convs1.2.weight_v
-vq_model.dec.resblocks.0.convs2.0.bias
-vq_model.dec.resblocks.0.convs2.0.weight_g
-vq_model.dec.resblocks.0.convs2.0.weight_v
-vq_model.dec.resblocks.0.convs2.1.bias
-vq_model.dec.resblocks.0.convs2.1.weight_g
-vq_model.dec.resblocks.0.convs2.1.weight_v
-vq_model.dec.resblocks.0.convs2.2.bias
-vq_model.dec.resblocks.0.convs2.2.weight_g
-vq_model.dec.resblocks.0.convs2.2.weight_v
-vq_model.dec.resblocks.1.convs1.0.bias
-vq_model.dec.resblocks.1.convs1.0.weight_g
-vq_model.dec.resblocks.1.convs1.0.weight_v
-vq_model.dec.resblocks.1.convs1.1.bias
-vq_model.dec.resblocks.1.convs1.1.weight_g
-vq_model.dec.resblocks.1.convs1.1.weight_v
-vq_model.dec.resblocks.1.convs1.2.bias
-vq_model.dec.resblocks.1.convs1.2.weight_g
-vq_model.dec.resblocks.1.convs1.2.weight_v
-vq_model.dec.resblocks.1.convs2.0.bias
-vq_model.dec.resblocks.1.convs2.0.weight_g
-vq_model.dec.resblocks.1.convs2.0.weight_v
-vq_model.dec.resblocks.1.convs2.1.bias
-vq_model.dec.resblocks.1.convs2.1.weight_g
-vq_model.dec.resblocks.1.convs2.1.weight_v
-vq_model.dec.resblocks.1.convs2.2.bias
-vq_model.dec.resblocks.1.convs2.2.weight_g
-vq_model.dec.resblocks.1.convs2.2.weight_v
-vq_model.dec.resblocks.10.convs1.0.bias
-vq_model.dec.resblocks.10.convs1.0.weight_g
-vq_model.dec.resblocks.10.convs1.0.weight_v
-vq_model.dec.resblocks.10.convs1.1.bias
-vq_model.dec.resblocks.10.convs1.1.weight_g
-vq_model.dec.resblocks.10.convs1.1.weight_v
-vq_model.dec.resblocks.10.convs1.2.bias
-vq_model.dec.resblocks.10.convs1.2.weight_g
-vq_model.dec.resblocks.10.convs1.2.weight_v
-vq_model.dec.resblocks.10.convs2.0.bias
-vq_model.dec.resblocks.10.convs2.0.weight_g
-vq_model.dec.resblocks.10.convs2.0.weight_v
-vq_model.dec.resblocks.10.convs2.1.bias
-vq_model.dec.resblocks.10.convs2.1.weight_g
-vq_model.dec.resblocks.10.convs2.1.weight_v
-vq_model.dec.resblocks.10.convs2.2.bias
-vq_model.dec.resblocks.10.convs2.2.weight_g
-vq_model.dec.resblocks.10.convs2.2.weight_v
-vq_model.dec.resblocks.11.convs1.0.bias
-vq_model.dec.resblocks.11.convs1.0.weight_g
-vq_model.dec.resblocks.11.convs1.0.weight_v
-vq_model.dec.resblocks.11.convs1.1.bias
-vq_model.dec.resblocks.11.convs1.1.weight_g
-vq_model.dec.resblocks.11.convs1.1.weight_v
-vq_model.dec.resblocks.11.convs1.2.bias
-vq_model.dec.resblocks.11.convs1.2.weight_g
-vq_model.dec.resblocks.11.convs1.2.weight_v
-vq_model.dec.resblocks.11.convs2.0.bias
-vq_model.dec.resblocks.11.convs2.0.weight_g
-vq_model.dec.resblocks.11.convs2.0.weight_v
-vq_model.dec.resblocks.11.convs2.1.bias
-vq_model.dec.resblocks.11.convs2.1.weight_g
-vq_model.dec.resblocks.11.convs2.1.weight_v
-vq_model.dec.resblocks.11.convs2.2.bias
-vq_model.dec.resblocks.11.convs2.2.weight_g
-vq_model.dec.resblocks.11.convs2.2.weight_v
-vq_model.dec.resblocks.12.convs1.0.bias
-vq_model.dec.resblocks.12.convs1.0.weight_g
-vq_model.dec.resblocks.12.convs1.0.weight_v
-vq_model.dec.resblocks.12.convs1.1.bias
-vq_model.dec.resblocks.12.convs1.1.weight_g
-vq_model.dec.resblocks.12.convs1.1.weight_v
-vq_model.dec.resblocks.12.convs1.2.bias
-vq_model.dec.resblocks.12.convs1.2.weight_g
-vq_model.dec.resblocks.12.convs1.2.weight_v
-vq_model.dec.resblocks.12.convs2.0.bias
-vq_model.dec.resblocks.12.convs2.0.weight_g
-vq_model.dec.resblocks.12.convs2.0.weight_v
-vq_model.dec.resblocks.12.convs2.1.bias
-vq_model.dec.resblocks.12.convs2.1.weight_g
-vq_model.dec.resblocks.12.convs2.1.weight_v
-vq_model.dec.resblocks.12.convs2.2.bias
-vq_model.dec.resblocks.12.convs2.2.weight_g
-vq_model.dec.resblocks.12.convs2.2.weight_v
-vq_model.dec.resblocks.13.convs1.0.bias
-vq_model.dec.resblocks.13.convs1.0.weight_g
-vq_model.dec.resblocks.13.convs1.0.weight_v
-vq_model.dec.resblocks.13.convs1.1.bias
-vq_model.dec.resblocks.13.convs1.1.weight_g
-vq_model.dec.resblocks.13.convs1.1.weight_v
-vq_model.dec.resblocks.13.convs1.2.bias
-vq_model.dec.resblocks.13.convs1.2.weight_g
-vq_model.dec.resblocks.13.convs1.2.weight_v
-vq_model.dec.resblocks.13.convs2.0.bias
-vq_model.dec.resblocks.13.convs2.0.weight_g
-vq_model.dec.resblocks.13.convs2.0.weight_v
-vq_model.dec.resblocks.13.convs2.1.bias
-vq_model.dec.resblocks.13.convs2.1.weight_g
-vq_model.dec.resblocks.13.convs2.1.weight_v
-vq_model.dec.resblocks.13.convs2.2.bias
-vq_model.dec.resblocks.13.convs2.2.weight_g
-vq_model.dec.resblocks.13.convs2.2.weight_v
-vq_model.dec.resblocks.14.convs1.0.bias
-vq_model.dec.resblocks.14.convs1.0.weight_g
-vq_model.dec.resblocks.14.convs1.0.weight_v
-vq_model.dec.resblocks.14.convs1.1.bias
-vq_model.dec.resblocks.14.convs1.1.weight_g
-vq_model.dec.resblocks.14.convs1.1.weight_v
-vq_model.dec.resblocks.14.convs1.2.bias
-vq_model.dec.resblocks.14.convs1.2.weight_g
-vq_model.dec.resblocks.14.convs1.2.weight_v
-vq_model.dec.resblocks.14.convs2.0.bias
-vq_model.dec.resblocks.14.convs2.0.weight_g
-vq_model.dec.resblocks.14.convs2.0.weight_v
-vq_model.dec.resblocks.14.convs2.1.bias
-vq_model.dec.resblocks.14.convs2.1.weight_g
-vq_model.dec.resblocks.14.convs2.1.weight_v
-vq_model.dec.resblocks.14.convs2.2.bias
-vq_model.dec.resblocks.14.convs2.2.weight_g
-vq_model.dec.resblocks.14.convs2.2.weight_v
-vq_model.dec.resblocks.2.convs1.0.bias
-vq_model.dec.resblocks.2.convs1.0.weight_g
-vq_model.dec.resblocks.2.convs1.0.weight_v
-vq_model.dec.resblocks.2.convs1.1.bias
-vq_model.dec.resblocks.2.convs1.1.weight_g
-vq_model.dec.resblocks.2.convs1.1.weight_v
-vq_model.dec.resblocks.2.convs1.2.bias
-vq_model.dec.resblocks.2.convs1.2.weight_g
-vq_model.dec.resblocks.2.convs1.2.weight_v
-vq_model.dec.resblocks.2.convs2.0.bias
-vq_model.dec.resblocks.2.convs2.0.weight_g
-vq_model.dec.resblocks.2.convs2.0.weight_v
-vq_model.dec.resblocks.2.convs2.1.bias
-vq_model.dec.resblocks.2.convs2.1.weight_g
-vq_model.dec.resblocks.2.convs2.1.weight_v
-vq_model.dec.resblocks.2.convs2.2.bias
-vq_model.dec.resblocks.2.convs2.2.weight_g
-vq_model.dec.resblocks.2.convs2.2.weight_v
-vq_model.dec.resblocks.3.convs1.0.bias
-vq_model.dec.resblocks.3.convs1.0.weight_g
-vq_model.dec.resblocks.3.convs1.0.weight_v
-vq_model.dec.resblocks.3.convs1.1.bias
-vq_model.dec.resblocks.3.convs1.1.weight_g
-vq_model.dec.resblocks.3.convs1.1.weight_v
-vq_model.dec.resblocks.3.convs1.2.bias
-vq_model.dec.resblocks.3.convs1.2.weight_g
-vq_model.dec.resblocks.3.convs1.2.weight_v
-vq_model.dec.resblocks.3.convs2.0.bias
-vq_model.dec.resblocks.3.convs2.0.weight_g
-vq_model.dec.resblocks.3.convs2.0.weight_v
-vq_model.dec.resblocks.3.convs2.1.bias
-vq_model.dec.resblocks.3.convs2.1.weight_g
-vq_model.dec.resblocks.3.convs2.1.weight_v
-vq_model.dec.resblocks.3.convs2.2.bias
-vq_model.dec.resblocks.3.convs2.2.weight_g
-vq_model.dec.resblocks.3.convs2.2.weight_v
-vq_model.dec.resblocks.4.convs1.0.bias
-vq_model.dec.resblocks.4.convs1.0.weight_g
-vq_model.dec.resblocks.4.convs1.0.weight_v
-vq_model.dec.resblocks.4.convs1.1.bias
-vq_model.dec.resblocks.4.convs1.1.weight_g
-vq_model.dec.resblocks.4.convs1.1.weight_v
-vq_model.dec.resblocks.4.convs1.2.bias
-vq_model.dec.resblocks.4.convs1.2.weight_g
-vq_model.dec.resblocks.4.convs1.2.weight_v
-vq_model.dec.resblocks.4.convs2.0.bias
-vq_model.dec.resblocks.4.convs2.0.weight_g
-vq_model.dec.resblocks.4.convs2.0.weight_v
-vq_model.dec.resblocks.4.convs2.1.bias
-vq_model.dec.resblocks.4.convs2.1.weight_g
-vq_model.dec.resblocks.4.convs2.1.weight_v
-vq_model.dec.resblocks.4.convs2.2.bias
-vq_model.dec.resblocks.4.convs2.2.weight_g
-vq_model.dec.resblocks.4.convs2.2.weight_v
-vq_model.dec.resblocks.5.convs1.0.bias
-vq_model.dec.resblocks.5.convs1.0.weight_g
-vq_model.dec.resblocks.5.convs1.0.weight_v
-vq_model.dec.resblocks.5.convs1.1.bias
-vq_model.dec.resblocks.5.convs1.1.weight_g
-vq_model.dec.resblocks.5.convs1.1.weight_v
-vq_model.dec.resblocks.5.convs1.2.bias
-vq_model.dec.resblocks.5.convs1.2.weight_g
-vq_model.dec.resblocks.5.convs1.2.weight_v
-vq_model.dec.resblocks.5.convs2.0.bias
-vq_model.dec.resblocks.5.convs2.0.weight_g
-vq_model.dec.resblocks.5.convs2.0.weight_v
-vq_model.dec.resblocks.5.convs2.1.bias
-vq_model.dec.resblocks.5.convs2.1.weight_g
-vq_model.dec.resblocks.5.convs2.1.weight_v
-vq_model.dec.resblocks.5.convs2.2.bias
-vq_model.dec.resblocks.5.convs2.2.weight_g
-vq_model.dec.resblocks.5.convs2.2.weight_v
-vq_model.dec.resblocks.6.convs1.0.bias
-vq_model.dec.resblocks.6.convs1.0.weight_g
-vq_model.dec.resblocks.6.convs1.0.weight_v
-vq_model.dec.resblocks.6.convs1.1.bias
-vq_model.dec.resblocks.6.convs1.1.weight_g
-vq_model.dec.resblocks.6.convs1.1.weight_v
-vq_model.dec.resblocks.6.convs1.2.bias
-vq_model.dec.resblocks.6.convs1.2.weight_g
-vq_model.dec.resblocks.6.convs1.2.weight_v
-vq_model.dec.resblocks.6.convs2.0.bias
-vq_model.dec.resblocks.6.convs2.0.weight_g
-vq_model.dec.resblocks.6.convs2.0.weight_v
-vq_model.dec.resblocks.6.convs2.1.bias
-vq_model.dec.resblocks.6.convs2.1.weight_g
-vq_model.dec.resblocks.6.convs2.1.weight_v
-vq_model.dec.resblocks.6.convs2.2.bias
-vq_model.dec.resblocks.6.convs2.2.weight_g
-vq_model.dec.resblocks.6.convs2.2.weight_v
-vq_model.dec.resblocks.7.convs1.0.bias
-vq_model.dec.resblocks.7.convs1.0.weight_g
-vq_model.dec.resblocks.7.convs1.0.weight_v
-vq_model.dec.resblocks.7.convs1.1.bias
-vq_model.dec.resblocks.7.convs1.1.weight_g
-vq_model.dec.resblocks.7.convs1.1.weight_v
-vq_model.dec.resblocks.7.convs1.2.bias
-vq_model.dec.resblocks.7.convs1.2.weight_g
-vq_model.dec.resblocks.7.convs1.2.weight_v
-vq_model.dec.resblocks.7.convs2.0.bias
-vq_model.dec.resblocks.7.convs2.0.weight_g
-vq_model.dec.resblocks.7.convs2.0.weight_v
-vq_model.dec.resblocks.7.convs2.1.bias
-vq_model.dec.resblocks.7.convs2.1.weight_g
-vq_model.dec.resblocks.7.convs2.1.weight_v
-vq_model.dec.resblocks.7.convs2.2.bias
-vq_model.dec.resblocks.7.convs2.2.weight_g
-vq_model.dec.resblocks.7.convs2.2.weight_v
-vq_model.dec.resblocks.8.convs1.0.bias
-vq_model.dec.resblocks.8.convs1.0.weight_g
-vq_model.dec.resblocks.8.convs1.0.weight_v
-vq_model.dec.resblocks.8.convs1.1.bias
-vq_model.dec.resblocks.8.convs1.1.weight_g
-vq_model.dec.resblocks.8.convs1.1.weight_v
-vq_model.dec.resblocks.8.convs1.2.bias
-vq_model.dec.resblocks.8.convs1.2.weight_g
-vq_model.dec.resblocks.8.convs1.2.weight_v
-vq_model.dec.resblocks.8.convs2.0.bias
-vq_model.dec.resblocks.8.convs2.0.weight_g
-vq_model.dec.resblocks.8.convs2.0.weight_v
-vq_model.dec.resblocks.8.convs2.1.bias
-vq_model.dec.resblocks.8.convs2.1.weight_g
-vq_model.dec.resblocks.8.convs2.1.weight_v
-vq_model.dec.resblocks.8.convs2.2.bias
-vq_model.dec.resblocks.8.convs2.2.weight_g
-vq_model.dec.resblocks.8.convs2.2.weight_v
-vq_model.dec.resblocks.9.convs1.0.bias
-vq_model.dec.resblocks.9.convs1.0.weight_g
-vq_model.dec.resblocks.9.convs1.0.weight_v
-vq_model.dec.resblocks.9.convs1.1.bias
-vq_model.dec.resblocks.9.convs1.1.weight_g
-vq_model.dec.resblocks.9.convs1.1.weight_v
-vq_model.dec.resblocks.9.convs1.2.bias
-vq_model.dec.resblocks.9.convs1.2.weight_g
-vq_model.dec.resblocks.9.convs1.2.weight_v
-vq_model.dec.resblocks.9.convs2.0.bias
-vq_model.dec.resblocks.9.convs2.0.weight_g
-vq_model.dec.resblocks.9.convs2.0.weight_v
-vq_model.dec.resblocks.9.convs2.1.bias
-vq_model.dec.resblocks.9.convs2.1.weight_g
-vq_model.dec.resblocks.9.convs2.1.weight_v
-vq_model.dec.resblocks.9.convs2.2.bias
-vq_model.dec.resblocks.9.convs2.2.weight_g
-vq_model.dec.resblocks.9.convs2.2.weight_v
-vq_model.dec.ups.0.bias
-vq_model.dec.ups.0.weight_g
-vq_model.dec.ups.0.weight_v
-vq_model.dec.ups.1.bias
-vq_model.dec.ups.1.weight_g
-vq_model.dec.ups.1.weight_v
-vq_model.dec.ups.2.bias
-vq_model.dec.ups.2.weight_g
-vq_model.dec.ups.2.weight_v
-vq_model.dec.ups.3.bias
-vq_model.dec.ups.3.weight_g
-vq_model.dec.ups.3.weight_v
-vq_model.dec.ups.4.bias
-vq_model.dec.ups.4.weight_g
-vq_model.dec.ups.4.weight_v
-vq_model.enc_p.encoder2.attn_layers.0.conv_k.bias
-vq_model.enc_p.encoder2.attn_layers.0.conv_k.weight
-vq_model.enc_p.encoder2.attn_layers.0.conv_o.bias
-vq_model.enc_p.encoder2.attn_layers.0.conv_o.weight
-vq_model.enc_p.encoder2.attn_layers.0.conv_q.bias
-vq_model.enc_p.encoder2.attn_layers.0.conv_q.weight
-vq_model.enc_p.encoder2.attn_layers.0.conv_v.bias
-vq_model.enc_p.encoder2.attn_layers.0.conv_v.weight
-vq_model.enc_p.encoder2.attn_layers.0.emb_rel_k
-vq_model.enc_p.encoder2.attn_layers.0.emb_rel_v
-vq_model.enc_p.encoder2.attn_layers.1.conv_k.bias
-vq_model.enc_p.encoder2.attn_layers.1.conv_k.weight
-vq_model.enc_p.encoder2.attn_layers.1.conv_o.bias
-vq_model.enc_p.encoder2.attn_layers.1.conv_o.weight
-vq_model.enc_p.encoder2.attn_layers.1.conv_q.bias
-vq_model.enc_p.encoder2.attn_layers.1.conv_q.weight
-vq_model.enc_p.encoder2.attn_layers.1.conv_v.bias
-vq_model.enc_p.encoder2.attn_layers.1.conv_v.weight
-vq_model.enc_p.encoder2.attn_layers.1.emb_rel_k
-vq_model.enc_p.encoder2.attn_layers.1.emb_rel_v
-vq_model.enc_p.encoder2.attn_layers.2.conv_k.bias
-vq_model.enc_p.encoder2.attn_layers.2.conv_k.weight
-vq_model.enc_p.encoder2.attn_layers.2.conv_o.bias
-vq_model.enc_p.encoder2.attn_layers.2.conv_o.weight
-vq_model.enc_p.encoder2.attn_layers.2.conv_q.bias
-vq_model.enc_p.encoder2.attn_layers.2.conv_q.weight
-vq_model.enc_p.encoder2.attn_layers.2.conv_v.bias
-vq_model.enc_p.encoder2.attn_layers.2.conv_v.weight
-vq_model.enc_p.encoder2.attn_layers.2.emb_rel_k
-vq_model.enc_p.encoder2.attn_layers.2.emb_rel_v
-vq_model.enc_p.encoder2.ffn_layers.0.conv_1.bias
-vq_model.enc_p.encoder2.ffn_layers.0.conv_1.weight
-vq_model.enc_p.encoder2.ffn_layers.0.conv_2.bias
-vq_model.enc_p.encoder2.ffn_layers.0.conv_2.weight
-vq_model.enc_p.encoder2.ffn_layers.1.conv_1.bias
-vq_model.enc_p.encoder2.ffn_layers.1.conv_1.weight
-vq_model.enc_p.encoder2.ffn_layers.1.conv_2.bias
-vq_model.enc_p.encoder2.ffn_layers.1.conv_2.weight
-vq_model.enc_p.encoder2.ffn_layers.2.conv_1.bias
-vq_model.enc_p.encoder2.ffn_layers.2.conv_1.weight
-vq_model.enc_p.encoder2.ffn_layers.2.conv_2.bias
-vq_model.enc_p.encoder2.ffn_layers.2.conv_2.weight
-vq_model.enc_p.encoder2.norm_layers_1.0.beta
-vq_model.enc_p.encoder2.norm_layers_1.0.gamma
-vq_model.enc_p.encoder2.norm_layers_1.1.beta
-vq_model.enc_p.encoder2.norm_layers_1.1.gamma
-vq_model.enc_p.encoder2.norm_layers_1.2.beta
-vq_model.enc_p.encoder2.norm_layers_1.2.gamma
-vq_model.enc_p.encoder2.norm_layers_2.0.beta
-vq_model.enc_p.encoder2.norm_layers_2.0.gamma
-vq_model.enc_p.encoder2.norm_layers_2.1.beta
-vq_model.enc_p.encoder2.norm_layers_2.1.gamma
-vq_model.enc_p.encoder2.norm_layers_2.2.beta
-vq_model.enc_p.encoder2.norm_layers_2.2.gamma
-vq_model.enc_p.encoder_ssl.attn_layers.0.conv_k.bias
-vq_model.enc_p.encoder_ssl.attn_layers.0.conv_k.weight
-vq_model.enc_p.encoder_ssl.attn_layers.0.conv_o.bias
-vq_model.enc_p.encoder_ssl.attn_layers.0.conv_o.weight
-vq_model.enc_p.encoder_ssl.attn_layers.0.conv_q.bias
-vq_model.enc_p.encoder_ssl.attn_layers.0.conv_q.weight
-vq_model.enc_p.encoder_ssl.attn_layers.0.conv_v.bias
-vq_model.enc_p.encoder_ssl.attn_layers.0.conv_v.weight
-vq_model.enc_p.encoder_ssl.attn_layers.0.emb_rel_k
-vq_model.enc_p.encoder_ssl.attn_layers.0.emb_rel_v
-vq_model.enc_p.encoder_ssl.attn_layers.1.conv_k.bias
-vq_model.enc_p.encoder_ssl.attn_layers.1.conv_k.weight
-vq_model.enc_p.encoder_ssl.attn_layers.1.conv_o.bias
-vq_model.enc_p.encoder_ssl.attn_layers.1.conv_o.weight
-vq_model.enc_p.encoder_ssl.attn_layers.1.conv_q.bias
-vq_model.enc_p.encoder_ssl.attn_layers.1.conv_q.weight
-vq_model.enc_p.encoder_ssl.attn_layers.1.conv_v.bias
-vq_model.enc_p.encoder_ssl.attn_layers.1.conv_v.weight
-vq_model.enc_p.encoder_ssl.attn_layers.1.emb_rel_k
-vq_model.enc_p.encoder_ssl.attn_layers.1.emb_rel_v
-vq_model.enc_p.encoder_ssl.attn_layers.2.conv_k.bias
-vq_model.enc_p.encoder_ssl.attn_layers.2.conv_k.weight
-vq_model.enc_p.encoder_ssl.attn_layers.2.conv_o.bias
-vq_model.enc_p.encoder_ssl.attn_layers.2.conv_o.weight
-vq_model.enc_p.encoder_ssl.attn_layers.2.conv_q.bias
-vq_model.enc_p.encoder_ssl.attn_layers.2.conv_q.weight
-vq_model.enc_p.encoder_ssl.attn_layers.2.conv_v.bias
-vq_model.enc_p.encoder_ssl.attn_layers.2.conv_v.weight
-vq_model.enc_p.encoder_ssl.attn_layers.2.emb_rel_k
-vq_model.enc_p.encoder_ssl.attn_layers.2.emb_rel_v
-vq_model.enc_p.encoder_ssl.ffn_layers.0.conv_1.bias
-vq_model.enc_p.encoder_ssl.ffn_layers.0.conv_1.weight
-vq_model.enc_p.encoder_ssl.ffn_layers.0.conv_2.bias
-vq_model.enc_p.encoder_ssl.ffn_layers.0.conv_2.weight
-vq_model.enc_p.encoder_ssl.ffn_layers.1.conv_1.bias
-vq_model.enc_p.encoder_ssl.ffn_layers.1.conv_1.weight
-vq_model.enc_p.encoder_ssl.ffn_layers.1.conv_2.bias
-vq_model.enc_p.encoder_ssl.ffn_layers.1.conv_2.weight
-vq_model.enc_p.encoder_ssl.ffn_layers.2.conv_1.bias
-vq_model.enc_p.encoder_ssl.ffn_layers.2.conv_1.weight
-vq_model.enc_p.encoder_ssl.ffn_layers.2.conv_2.bias
-vq_model.enc_p.encoder_ssl.ffn_layers.2.conv_2.weight
-vq_model.enc_p.encoder_ssl.norm_layers_1.0.beta
-vq_model.enc_p.encoder_ssl.norm_layers_1.0.gamma
-vq_model.enc_p.encoder_ssl.norm_layers_1.1.beta
-vq_model.enc_p.encoder_ssl.norm_layers_1.1.gamma
-vq_model.enc_p.encoder_ssl.norm_layers_1.2.beta
-vq_model.enc_p.encoder_ssl.norm_layers_1.2.gamma
-vq_model.enc_p.encoder_ssl.norm_layers_2.0.beta
-vq_model.enc_p.encoder_ssl.norm_layers_2.0.gamma
-vq_model.enc_p.encoder_ssl.norm_layers_2.1.beta
-vq_model.enc_p.encoder_ssl.norm_layers_2.1.gamma
-vq_model.enc_p.encoder_ssl.norm_layers_2.2.beta
-vq_model.enc_p.encoder_ssl.norm_layers_2.2.gamma
-vq_model.enc_p.encoder_text.attn_layers.0.conv_k.bias
-vq_model.enc_p.encoder_text.attn_layers.0.conv_k.weight
-vq_model.enc_p.encoder_text.attn_layers.0.conv_o.bias
-vq_model.enc_p.encoder_text.attn_layers.0.conv_o.weight
-vq_model.enc_p.encoder_text.attn_layers.0.conv_q.bias
-vq_model.enc_p.encoder_text.attn_layers.0.conv_q.weight
-vq_model.enc_p.encoder_text.attn_layers.0.conv_v.bias
-vq_model.enc_p.encoder_text.attn_layers.0.conv_v.weight
-vq_model.enc_p.encoder_text.attn_layers.0.emb_rel_k
-vq_model.enc_p.encoder_text.attn_layers.0.emb_rel_v
-vq_model.enc_p.encoder_text.attn_layers.1.conv_k.bias
-vq_model.enc_p.encoder_text.attn_layers.1.conv_k.weight
-vq_model.enc_p.encoder_text.attn_layers.1.conv_o.bias
-vq_model.enc_p.encoder_text.attn_layers.1.conv_o.weight
-vq_model.enc_p.encoder_text.attn_layers.1.conv_q.bias
-vq_model.enc_p.encoder_text.attn_layers.1.conv_q.weight
-vq_model.enc_p.encoder_text.attn_layers.1.conv_v.bias
-vq_model.enc_p.encoder_text.attn_layers.1.conv_v.weight
-vq_model.enc_p.encoder_text.attn_layers.1.emb_rel_k
-vq_model.enc_p.encoder_text.attn_layers.1.emb_rel_v
-vq_model.enc_p.encoder_text.attn_layers.2.conv_k.bias
-vq_model.enc_p.encoder_text.attn_layers.2.conv_k.weight
-vq_model.enc_p.encoder_text.attn_layers.2.conv_o.bias
-vq_model.enc_p.encoder_text.attn_layers.2.conv_o.weight
-vq_model.enc_p.encoder_text.attn_layers.2.conv_q.bias
-vq_model.enc_p.encoder_text.attn_layers.2.conv_q.weight
-vq_model.enc_p.encoder_text.attn_layers.2.conv_v.bias
-vq_model.enc_p.encoder_text.attn_layers.2.conv_v.weight
-vq_model.enc_p.encoder_text.attn_layers.2.emb_rel_k
-vq_model.enc_p.encoder_text.attn_layers.2.emb_rel_v
-vq_model.enc_p.encoder_text.attn_layers.3.conv_k.bias
-vq_model.enc_p.encoder_text.attn_layers.3.conv_k.weight
-vq_model.enc_p.encoder_text.attn_layers.3.conv_o.bias
-vq_model.enc_p.encoder_text.attn_layers.3.conv_o.weight
-vq_model.enc_p.encoder_text.attn_layers.3.conv_q.bias
-vq_model.enc_p.encoder_text.attn_layers.3.conv_q.weight
-vq_model.enc_p.encoder_text.attn_layers.3.conv_v.bias
-vq_model.enc_p.encoder_text.attn_layers.3.conv_v.weight
-vq_model.enc_p.encoder_text.attn_layers.3.emb_rel_k
-vq_model.enc_p.encoder_text.attn_layers.3.emb_rel_v
-vq_model.enc_p.encoder_text.attn_layers.4.conv_k.bias
-vq_model.enc_p.encoder_text.attn_layers.4.conv_k.weight
-vq_model.enc_p.encoder_text.attn_layers.4.conv_o.bias
-vq_model.enc_p.encoder_text.attn_layers.4.conv_o.weight
-vq_model.enc_p.encoder_text.attn_layers.4.conv_q.bias
-vq_model.enc_p.encoder_text.attn_layers.4.conv_q.weight
-vq_model.enc_p.encoder_text.attn_layers.4.conv_v.bias
-vq_model.enc_p.encoder_text.attn_layers.4.conv_v.weight
-vq_model.enc_p.encoder_text.attn_layers.4.emb_rel_k
-vq_model.enc_p.encoder_text.attn_layers.4.emb_rel_v
-vq_model.enc_p.encoder_text.attn_layers.5.conv_k.bias
-vq_model.enc_p.encoder_text.attn_layers.5.conv_k.weight
-vq_model.enc_p.encoder_text.attn_layers.5.conv_o.bias
-vq_model.enc_p.encoder_text.attn_layers.5.conv_o.weight
-vq_model.enc_p.encoder_text.attn_layers.5.conv_q.bias
-vq_model.enc_p.encoder_text.attn_layers.5.conv_q.weight
-vq_model.enc_p.encoder_text.attn_layers.5.conv_v.bias
-vq_model.enc_p.encoder_text.attn_layers.5.conv_v.weight
-vq_model.enc_p.encoder_text.attn_layers.5.emb_rel_k
-vq_model.enc_p.encoder_text.attn_layers.5.emb_rel_v
-vq_model.enc_p.encoder_text.ffn_layers.0.conv_1.bias
-vq_model.enc_p.encoder_text.ffn_layers.0.conv_1.weight
-vq_model.enc_p.encoder_text.ffn_layers.0.conv_2.bias
-vq_model.enc_p.encoder_text.ffn_layers.0.conv_2.weight
-vq_model.enc_p.encoder_text.ffn_layers.1.conv_1.bias
-vq_model.enc_p.encoder_text.ffn_layers.1.conv_1.weight
-vq_model.enc_p.encoder_text.ffn_layers.1.conv_2.bias
-vq_model.enc_p.encoder_text.ffn_layers.1.conv_2.weight
-vq_model.enc_p.encoder_text.ffn_layers.2.conv_1.bias
-vq_model.enc_p.encoder_text.ffn_layers.2.conv_1.weight
-vq_model.enc_p.encoder_text.ffn_layers.2.conv_2.bias
-vq_model.enc_p.encoder_text.ffn_layers.2.conv_2.weight
-vq_model.enc_p.encoder_text.ffn_layers.3.conv_1.bias
-vq_model.enc_p.encoder_text.ffn_layers.3.conv_1.weight
-vq_model.enc_p.encoder_text.ffn_layers.3.conv_2.bias
-vq_model.enc_p.encoder_text.ffn_layers.3.conv_2.weight
-vq_model.enc_p.encoder_text.ffn_layers.4.conv_1.bias
-vq_model.enc_p.encoder_text.ffn_layers.4.conv_1.weight
-vq_model.enc_p.encoder_text.ffn_layers.4.conv_2.bias
-vq_model.enc_p.encoder_text.ffn_layers.4.conv_2.weight
-vq_model.enc_p.encoder_text.ffn_layers.5.conv_1.bias
-vq_model.enc_p.encoder_text.ffn_layers.5.conv_1.weight
-vq_model.enc_p.encoder_text.ffn_layers.5.conv_2.bias
-vq_model.enc_p.encoder_text.ffn_layers.5.conv_2.weight
-vq_model.enc_p.encoder_text.norm_layers_1.0.beta
-vq_model.enc_p.encoder_text.norm_layers_1.0.gamma
-vq_model.enc_p.encoder_text.norm_layers_1.1.beta
-vq_model.enc_p.encoder_text.norm_layers_1.1.gamma
-vq_model.enc_p.encoder_text.norm_layers_1.2.beta
-vq_model.enc_p.encoder_text.norm_layers_1.2.gamma
-vq_model.enc_p.encoder_text.norm_layers_1.3.beta
-vq_model.enc_p.encoder_text.norm_layers_1.3.gamma
-vq_model.enc_p.encoder_text.norm_layers_1.4.beta
-vq_model.enc_p.encoder_text.norm_layers_1.4.gamma
-vq_model.enc_p.encoder_text.norm_layers_1.5.beta
-vq_model.enc_p.encoder_text.norm_layers_1.5.gamma
-vq_model.enc_p.encoder_text.norm_layers_2.0.beta
-vq_model.enc_p.encoder_text.norm_layers_2.0.gamma
-vq_model.enc_p.encoder_text.norm_layers_2.1.beta
-vq_model.enc_p.encoder_text.norm_layers_2.1.gamma
-vq_model.enc_p.encoder_text.norm_layers_2.2.beta
-vq_model.enc_p.encoder_text.norm_layers_2.2.gamma
-vq_model.enc_p.encoder_text.norm_layers_2.3.beta
-vq_model.enc_p.encoder_text.norm_layers_2.3.gamma
-vq_model.enc_p.encoder_text.norm_layers_2.4.beta
-vq_model.enc_p.encoder_text.norm_layers_2.4.gamma
-vq_model.enc_p.encoder_text.norm_layers_2.5.beta
-vq_model.enc_p.encoder_text.norm_layers_2.5.gamma
-vq_model.enc_p.mrte.c_post.bias
-vq_model.enc_p.mrte.c_post.weight
-vq_model.enc_p.mrte.c_pre.bias
-vq_model.enc_p.mrte.c_pre.weight
-vq_model.enc_p.mrte.cross_attention.conv_k.bias
-vq_model.enc_p.mrte.cross_attention.conv_k.weight
-vq_model.enc_p.mrte.cross_attention.conv_o.bias
-vq_model.enc_p.mrte.cross_attention.conv_o.weight
-vq_model.enc_p.mrte.cross_attention.conv_q.bias
-vq_model.enc_p.mrte.cross_attention.conv_q.weight
-vq_model.enc_p.mrte.cross_attention.conv_v.bias
-vq_model.enc_p.mrte.cross_attention.conv_v.weight
-vq_model.enc_p.mrte.text_pre.bias
-vq_model.enc_p.mrte.text_pre.weight
-vq_model.enc_p.proj.bias
-vq_model.enc_p.proj.weight
-vq_model.enc_p.ssl_proj.bias
-vq_model.enc_p.ssl_proj.weight
-vq_model.enc_p.text_embedding.weight
-vq_model.flow.flows.0.enc.cond_layer.bias
-vq_model.flow.flows.0.enc.cond_layer.weight_g
-vq_model.flow.flows.0.enc.cond_layer.weight_v
-vq_model.flow.flows.0.enc.in_layers.0.bias
-vq_model.flow.flows.0.enc.in_layers.0.weight_g
-vq_model.flow.flows.0.enc.in_layers.0.weight_v
-vq_model.flow.flows.0.enc.in_layers.1.bias
-vq_model.flow.flows.0.enc.in_layers.1.weight_g
-vq_model.flow.flows.0.enc.in_layers.1.weight_v
-vq_model.flow.flows.0.enc.in_layers.2.bias
-vq_model.flow.flows.0.enc.in_layers.2.weight_g
-vq_model.flow.flows.0.enc.in_layers.2.weight_v
-vq_model.flow.flows.0.enc.in_layers.3.bias
-vq_model.flow.flows.0.enc.in_layers.3.weight_g
-vq_model.flow.flows.0.enc.in_layers.3.weight_v
-vq_model.flow.flows.0.enc.res_skip_layers.0.bias
-vq_model.flow.flows.0.enc.res_skip_layers.0.weight_g
-vq_model.flow.flows.0.enc.res_skip_layers.0.weight_v
-vq_model.flow.flows.0.enc.res_skip_layers.1.bias
-vq_model.flow.flows.0.enc.res_skip_layers.1.weight_g
-vq_model.flow.flows.0.enc.res_skip_layers.1.weight_v
-vq_model.flow.flows.0.enc.res_skip_layers.2.bias
-vq_model.flow.flows.0.enc.res_skip_layers.2.weight_g
-vq_model.flow.flows.0.enc.res_skip_layers.2.weight_v
-vq_model.flow.flows.0.enc.res_skip_layers.3.bias
-vq_model.flow.flows.0.enc.res_skip_layers.3.weight_g
-vq_model.flow.flows.0.enc.res_skip_layers.3.weight_v
-vq_model.flow.flows.0.post.bias
-vq_model.flow.flows.0.post.weight
-vq_model.flow.flows.0.pre.bias
-vq_model.flow.flows.0.pre.weight
-vq_model.flow.flows.2.enc.cond_layer.bias
-vq_model.flow.flows.2.enc.cond_layer.weight_g
-vq_model.flow.flows.2.enc.cond_layer.weight_v
-vq_model.flow.flows.2.enc.in_layers.0.bias
-vq_model.flow.flows.2.enc.in_layers.0.weight_g
-vq_model.flow.flows.2.enc.in_layers.0.weight_v
-vq_model.flow.flows.2.enc.in_layers.1.bias
-vq_model.flow.flows.2.enc.in_layers.1.weight_g
-vq_model.flow.flows.2.enc.in_layers.1.weight_v
-vq_model.flow.flows.2.enc.in_layers.2.bias
-vq_model.flow.flows.2.enc.in_layers.2.weight_g
-vq_model.flow.flows.2.enc.in_layers.2.weight_v
-vq_model.flow.flows.2.enc.in_layers.3.bias
-vq_model.flow.flows.2.enc.in_layers.3.weight_g
-vq_model.flow.flows.2.enc.in_layers.3.weight_v
-vq_model.flow.flows.2.enc.res_skip_layers.0.bias
-vq_model.flow.flows.2.enc.res_skip_layers.0.weight_g
-vq_model.flow.flows.2.enc.res_skip_layers.0.weight_v
-vq_model.flow.flows.2.enc.res_skip_layers.1.bias
-vq_model.flow.flows.2.enc.res_skip_layers.1.weight_g
-vq_model.flow.flows.2.enc.res_skip_layers.1.weight_v
-vq_model.flow.flows.2.enc.res_skip_layers.2.bias
-vq_model.flow.flows.2.enc.res_skip_layers.2.weight_g
-vq_model.flow.flows.2.enc.res_skip_layers.2.weight_v
-vq_model.flow.flows.2.enc.res_skip_layers.3.bias
-vq_model.flow.flows.2.enc.res_skip_layers.3.weight_g
-vq_model.flow.flows.2.enc.res_skip_layers.3.weight_v
-vq_model.flow.flows.2.post.bias
-vq_model.flow.flows.2.post.weight
-vq_model.flow.flows.2.pre.bias
-vq_model.flow.flows.2.pre.weight
-vq_model.flow.flows.4.enc.cond_layer.bias
-vq_model.flow.flows.4.enc.cond_layer.weight_g
-vq_model.flow.flows.4.enc.cond_layer.weight_v
-vq_model.flow.flows.4.enc.in_layers.0.bias
-vq_model.flow.flows.4.enc.in_layers.0.weight_g
-vq_model.flow.flows.4.enc.in_layers.0.weight_v
-vq_model.flow.flows.4.enc.in_layers.1.bias
-vq_model.flow.flows.4.enc.in_layers.1.weight_g
-vq_model.flow.flows.4.enc.in_layers.1.weight_v
-vq_model.flow.flows.4.enc.in_layers.2.bias
-vq_model.flow.flows.4.enc.in_layers.2.weight_g
-vq_model.flow.flows.4.enc.in_layers.2.weight_v
-vq_model.flow.flows.4.enc.in_layers.3.bias
-vq_model.flow.flows.4.enc.in_layers.3.weight_g
-vq_model.flow.flows.4.enc.in_layers.3.weight_v
-vq_model.flow.flows.4.enc.res_skip_layers.0.bias
-vq_model.flow.flows.4.enc.res_skip_layers.0.weight_g
-vq_model.flow.flows.4.enc.res_skip_layers.0.weight_v
-vq_model.flow.flows.4.enc.res_skip_layers.1.bias
-vq_model.flow.flows.4.enc.res_skip_layers.1.weight_g
-vq_model.flow.flows.4.enc.res_skip_layers.1.weight_v
-vq_model.flow.flows.4.enc.res_skip_layers.2.bias
-vq_model.flow.flows.4.enc.res_skip_layers.2.weight_g
-vq_model.flow.flows.4.enc.res_skip_layers.2.weight_v
-vq_model.flow.flows.4.enc.res_skip_layers.3.bias
-vq_model.flow.flows.4.enc.res_skip_layers.3.weight_g
-vq_model.flow.flows.4.enc.res_skip_layers.3.weight_v
-vq_model.flow.flows.4.post.bias
-vq_model.flow.flows.4.post.weight
-vq_model.flow.flows.4.pre.bias
-vq_model.flow.flows.4.pre.weight
-vq_model.flow.flows.6.enc.cond_layer.bias
-vq_model.flow.flows.6.enc.cond_layer.weight_g
-vq_model.flow.flows.6.enc.cond_layer.weight_v
-vq_model.flow.flows.6.enc.in_layers.0.bias
-vq_model.flow.flows.6.enc.in_layers.0.weight_g
-vq_model.flow.flows.6.enc.in_layers.0.weight_v
-vq_model.flow.flows.6.enc.in_layers.1.bias
-vq_model.flow.flows.6.enc.in_layers.1.weight_g
-vq_model.flow.flows.6.enc.in_layers.1.weight_v
-vq_model.flow.flows.6.enc.in_layers.2.bias
-vq_model.flow.flows.6.enc.in_layers.2.weight_g
-vq_model.flow.flows.6.enc.in_layers.2.weight_v
-vq_model.flow.flows.6.enc.in_layers.3.bias
-vq_model.flow.flows.6.enc.in_layers.3.weight_g
-vq_model.flow.flows.6.enc.in_layers.3.weight_v
-vq_model.flow.flows.6.enc.res_skip_layers.0.bias
-vq_model.flow.flows.6.enc.res_skip_layers.0.weight_g
-vq_model.flow.flows.6.enc.res_skip_layers.0.weight_v
-vq_model.flow.flows.6.enc.res_skip_layers.1.bias
-vq_model.flow.flows.6.enc.res_skip_layers.1.weight_g
-vq_model.flow.flows.6.enc.res_skip_layers.1.weight_v
-vq_model.flow.flows.6.enc.res_skip_layers.2.bias
-vq_model.flow.flows.6.enc.res_skip_layers.2.weight_g
-vq_model.flow.flows.6.enc.res_skip_layers.2.weight_v
-vq_model.flow.flows.6.enc.res_skip_layers.3.bias
-vq_model.flow.flows.6.enc.res_skip_layers.3.weight_g
-vq_model.flow.flows.6.enc.res_skip_layers.3.weight_v
-vq_model.flow.flows.6.post.bias
-vq_model.flow.flows.6.post.weight
-vq_model.flow.flows.6.pre.bias
-vq_model.flow.flows.6.pre.weight
-vq_model.quantizer.vq.layers.0._codebook.embed
-vq_model.ref_enc.fc.fc.bias
-vq_model.ref_enc.fc.fc.weight
-vq_model.ref_enc.slf_attn.fc.bias
-vq_model.ref_enc.slf_attn.fc.weight
-vq_model.ref_enc.slf_attn.w_ks.bias
-vq_model.ref_enc.slf_attn.w_ks.weight
-vq_model.ref_enc.slf_attn.w_qs.bias
-vq_model.ref_enc.slf_attn.w_qs.weight
-vq_model.ref_enc.slf_attn.w_vs.bias
-vq_model.ref_enc.slf_attn.w_vs.weight
-vq_model.ref_enc.spectral.0.fc.bias
-vq_model.ref_enc.spectral.0.fc.weight
-vq_model.ref_enc.spectral.3.fc.bias
-vq_model.ref_enc.spectral.3.fc.weight
-vq_model.ref_enc.temporal.0.conv1.conv.bias
-vq_model.ref_enc.temporal.0.conv1.conv.weight
-vq_model.ref_enc.temporal.1.conv1.conv.bias
-vq_model.ref_enc.temporal.1.conv1.conv.weight
+vq_model.dec.cond.bias
+vq_model.dec.cond.weight
+vq_model.dec.conv_post.weight
+vq_model.dec.conv_pre.bias
+vq_model.dec.conv_pre.weight
+vq_model.dec.resblocks.0.convs1.0.bias
+vq_model.dec.resblocks.0.convs1.0.weight_g
+vq_model.dec.resblocks.0.convs1.0.weight_v
+vq_model.dec.resblocks.0.convs1.1.bias
+vq_model.dec.resblocks.0.convs1.1.weight_g
+vq_model.dec.resblocks.0.convs1.1.weight_v
+vq_model.dec.resblocks.0.convs1.2.bias
+vq_model.dec.resblocks.0.convs1.2.weight_g
+vq_model.dec.resblocks.0.convs1.2.weight_v
+vq_model.dec.resblocks.0.convs2.0.bias
+vq_model.dec.resblocks.0.convs2.0.weight_g
+vq_model.dec.resblocks.0.convs2.0.weight_v
+vq_model.dec.resblocks.0.convs2.1.bias
+vq_model.dec.resblocks.0.convs2.1.weight_g
+vq_model.dec.resblocks.0.convs2.1.weight_v
+vq_model.dec.resblocks.0.convs2.2.bias
+vq_model.dec.resblocks.0.convs2.2.weight_g
+vq_model.dec.resblocks.0.convs2.2.weight_v
+vq_model.dec.resblocks.1.convs1.0.bias
+vq_model.dec.resblocks.1.convs1.0.weight_g
+vq_model.dec.resblocks.1.convs1.0.weight_v
+vq_model.dec.resblocks.1.convs1.1.bias
+vq_model.dec.resblocks.1.convs1.1.weight_g
+vq_model.dec.resblocks.1.convs1.1.weight_v
+vq_model.dec.resblocks.1.convs1.2.bias
+vq_model.dec.resblocks.1.convs1.2.weight_g
+vq_model.dec.resblocks.1.convs1.2.weight_v
+vq_model.dec.resblocks.1.convs2.0.bias
+vq_model.dec.resblocks.1.convs2.0.weight_g
+vq_model.dec.resblocks.1.convs2.0.weight_v
+vq_model.dec.resblocks.1.convs2.1.bias
+vq_model.dec.resblocks.1.convs2.1.weight_g
+vq_model.dec.resblocks.1.convs2.1.weight_v
+vq_model.dec.resblocks.1.convs2.2.bias
+vq_model.dec.resblocks.1.convs2.2.weight_g
+vq_model.dec.resblocks.1.convs2.2.weight_v
+vq_model.dec.resblocks.10.convs1.0.bias
+vq_model.dec.resblocks.10.convs1.0.weight_g
+vq_model.dec.resblocks.10.convs1.0.weight_v
+vq_model.dec.resblocks.10.convs1.1.bias
+vq_model.dec.resblocks.10.convs1.1.weight_g
+vq_model.dec.resblocks.10.convs1.1.weight_v
+vq_model.dec.resblocks.10.convs1.2.bias
+vq_model.dec.resblocks.10.convs1.2.weight_g
+vq_model.dec.resblocks.10.convs1.2.weight_v
+vq_model.dec.resblocks.10.convs2.0.bias
+vq_model.dec.resblocks.10.convs2.0.weight_g
+vq_model.dec.resblocks.10.convs2.0.weight_v
+vq_model.dec.resblocks.10.convs2.1.bias
+vq_model.dec.resblocks.10.convs2.1.weight_g
+vq_model.dec.resblocks.10.convs2.1.weight_v
+vq_model.dec.resblocks.10.convs2.2.bias
+vq_model.dec.resblocks.10.convs2.2.weight_g
+vq_model.dec.resblocks.10.convs2.2.weight_v
+vq_model.dec.resblocks.11.convs1.0.bias
+vq_model.dec.resblocks.11.convs1.0.weight_g
+vq_model.dec.resblocks.11.convs1.0.weight_v
+vq_model.dec.resblocks.11.convs1.1.bias
+vq_model.dec.resblocks.11.convs1.1.weight_g
+vq_model.dec.resblocks.11.convs1.1.weight_v
+vq_model.dec.resblocks.11.convs1.2.bias
+vq_model.dec.resblocks.11.convs1.2.weight_g
+vq_model.dec.resblocks.11.convs1.2.weight_v
+vq_model.dec.resblocks.11.convs2.0.bias
+vq_model.dec.resblocks.11.convs2.0.weight_g
+vq_model.dec.resblocks.11.convs2.0.weight_v
+vq_model.dec.resblocks.11.convs2.1.bias
+vq_model.dec.resblocks.11.convs2.1.weight_g
+vq_model.dec.resblocks.11.convs2.1.weight_v
+vq_model.dec.resblocks.11.convs2.2.bias
+vq_model.dec.resblocks.11.convs2.2.weight_g
+vq_model.dec.resblocks.11.convs2.2.weight_v
+vq_model.dec.resblocks.12.convs1.0.bias
+vq_model.dec.resblocks.12.convs1.0.weight_g
+vq_model.dec.resblocks.12.convs1.0.weight_v
+vq_model.dec.resblocks.12.convs1.1.bias
+vq_model.dec.resblocks.12.convs1.1.weight_g
+vq_model.dec.resblocks.12.convs1.1.weight_v
+vq_model.dec.resblocks.12.convs1.2.bias
+vq_model.dec.resblocks.12.convs1.2.weight_g
+vq_model.dec.resblocks.12.convs1.2.weight_v
+vq_model.dec.resblocks.12.convs2.0.bias
+vq_model.dec.resblocks.12.convs2.0.weight_g
+vq_model.dec.resblocks.12.convs2.0.weight_v
+vq_model.dec.resblocks.12.convs2.1.bias
+vq_model.dec.resblocks.12.convs2.1.weight_g
+vq_model.dec.resblocks.12.convs2.1.weight_v
+vq_model.dec.resblocks.12.convs2.2.bias
+vq_model.dec.resblocks.12.convs2.2.weight_g
+vq_model.dec.resblocks.12.convs2.2.weight_v
+vq_model.dec.resblocks.13.convs1.0.bias
+vq_model.dec.resblocks.13.convs1.0.weight_g
+vq_model.dec.resblocks.13.convs1.0.weight_v
+vq_model.dec.resblocks.13.convs1.1.bias
+vq_model.dec.resblocks.13.convs1.1.weight_g
+vq_model.dec.resblocks.13.convs1.1.weight_v
+vq_model.dec.resblocks.13.convs1.2.bias
+vq_model.dec.resblocks.13.convs1.2.weight_g
+vq_model.dec.resblocks.13.convs1.2.weight_v
+vq_model.dec.resblocks.13.convs2.0.bias
+vq_model.dec.resblocks.13.convs2.0.weight_g
+vq_model.dec.resblocks.13.convs2.0.weight_v
+vq_model.dec.resblocks.13.convs2.1.bias
+vq_model.dec.resblocks.13.convs2.1.weight_g
+vq_model.dec.resblocks.13.convs2.1.weight_v
+vq_model.dec.resblocks.13.convs2.2.bias
+vq_model.dec.resblocks.13.convs2.2.weight_g
+vq_model.dec.resblocks.13.convs2.2.weight_v
+vq_model.dec.resblocks.14.convs1.0.bias
+vq_model.dec.resblocks.14.convs1.0.weight_g
+vq_model.dec.resblocks.14.convs1.0.weight_v
+vq_model.dec.resblocks.14.convs1.1.bias
+vq_model.dec.resblocks.14.convs1.1.weight_g
+vq_model.dec.resblocks.14.convs1.1.weight_v
+vq_model.dec.resblocks.14.convs1.2.bias
+vq_model.dec.resblocks.14.convs1.2.weight_g
+vq_model.dec.resblocks.14.convs1.2.weight_v
+vq_model.dec.resblocks.14.convs2.0.bias
+vq_model.dec.resblocks.14.convs2.0.weight_g
+vq_model.dec.resblocks.14.convs2.0.weight_v
+vq_model.dec.resblocks.14.convs2.1.bias
+vq_model.dec.resblocks.14.convs2.1.weight_g
+vq_model.dec.resblocks.14.convs2.1.weight_v
+vq_model.dec.resblocks.14.convs2.2.bias
+vq_model.dec.resblocks.14.convs2.2.weight_g
+vq_model.dec.resblocks.14.convs2.2.weight_v
+vq_model.dec.resblocks.2.convs1.0.bias
+vq_model.dec.resblocks.2.convs1.0.weight_g
+vq_model.dec.resblocks.2.convs1.0.weight_v
+vq_model.dec.resblocks.2.convs1.1.bias
+vq_model.dec.resblocks.2.convs1.1.weight_g
+vq_model.dec.resblocks.2.convs1.1.weight_v
+vq_model.dec.resblocks.2.convs1.2.bias
+vq_model.dec.resblocks.2.convs1.2.weight_g
+vq_model.dec.resblocks.2.convs1.2.weight_v
+vq_model.dec.resblocks.2.convs2.0.bias
+vq_model.dec.resblocks.2.convs2.0.weight_g
+vq_model.dec.resblocks.2.convs2.0.weight_v
+vq_model.dec.resblocks.2.convs2.1.bias
+vq_model.dec.resblocks.2.convs2.1.weight_g
+vq_model.dec.resblocks.2.convs2.1.weight_v
+vq_model.dec.resblocks.2.convs2.2.bias
+vq_model.dec.resblocks.2.convs2.2.weight_g
+vq_model.dec.resblocks.2.convs2.2.weight_v
+vq_model.dec.resblocks.3.convs1.0.bias
+vq_model.dec.resblocks.3.convs1.0.weight_g
+vq_model.dec.resblocks.3.convs1.0.weight_v
+vq_model.dec.resblocks.3.convs1.1.bias
+vq_model.dec.resblocks.3.convs1.1.weight_g
+vq_model.dec.resblocks.3.convs1.1.weight_v
+vq_model.dec.resblocks.3.convs1.2.bias
+vq_model.dec.resblocks.3.convs1.2.weight_g
+vq_model.dec.resblocks.3.convs1.2.weight_v
+vq_model.dec.resblocks.3.convs2.0.bias
+vq_model.dec.resblocks.3.convs2.0.weight_g
+vq_model.dec.resblocks.3.convs2.0.weight_v
+vq_model.dec.resblocks.3.convs2.1.bias
+vq_model.dec.resblocks.3.convs2.1.weight_g
+vq_model.dec.resblocks.3.convs2.1.weight_v
+vq_model.dec.resblocks.3.convs2.2.bias
+vq_model.dec.resblocks.3.convs2.2.weight_g
+vq_model.dec.resblocks.3.convs2.2.weight_v
+vq_model.dec.resblocks.4.convs1.0.bias
+vq_model.dec.resblocks.4.convs1.0.weight_g
+vq_model.dec.resblocks.4.convs1.0.weight_v
+vq_model.dec.resblocks.4.convs1.1.bias
+vq_model.dec.resblocks.4.convs1.1.weight_g
+vq_model.dec.resblocks.4.convs1.1.weight_v
+vq_model.dec.resblocks.4.convs1.2.bias
+vq_model.dec.resblocks.4.convs1.2.weight_g
+vq_model.dec.resblocks.4.convs1.2.weight_v
+vq_model.dec.resblocks.4.convs2.0.bias
+vq_model.dec.resblocks.4.convs2.0.weight_g
+vq_model.dec.resblocks.4.convs2.0.weight_v
+vq_model.dec.resblocks.4.convs2.1.bias
+vq_model.dec.resblocks.4.convs2.1.weight_g
+vq_model.dec.resblocks.4.convs2.1.weight_v
+vq_model.dec.resblocks.4.convs2.2.bias
+vq_model.dec.resblocks.4.convs2.2.weight_g
+vq_model.dec.resblocks.4.convs2.2.weight_v
+vq_model.dec.resblocks.5.convs1.0.bias
+vq_model.dec.resblocks.5.convs1.0.weight_g
+vq_model.dec.resblocks.5.convs1.0.weight_v
+vq_model.dec.resblocks.5.convs1.1.bias
+vq_model.dec.resblocks.5.convs1.1.weight_g
+vq_model.dec.resblocks.5.convs1.1.weight_v
+vq_model.dec.resblocks.5.convs1.2.bias
+vq_model.dec.resblocks.5.convs1.2.weight_g
+vq_model.dec.resblocks.5.convs1.2.weight_v
+vq_model.dec.resblocks.5.convs2.0.bias
+vq_model.dec.resblocks.5.convs2.0.weight_g
+vq_model.dec.resblocks.5.convs2.0.weight_v
+vq_model.dec.resblocks.5.convs2.1.bias
+vq_model.dec.resblocks.5.convs2.1.weight_g
+vq_model.dec.resblocks.5.convs2.1.weight_v
+vq_model.dec.resblocks.5.convs2.2.bias
+vq_model.dec.resblocks.5.convs2.2.weight_g
+vq_model.dec.resblocks.5.convs2.2.weight_v
+vq_model.dec.resblocks.6.convs1.0.bias
+vq_model.dec.resblocks.6.convs1.0.weight_g
+vq_model.dec.resblocks.6.convs1.0.weight_v
+vq_model.dec.resblocks.6.convs1.1.bias
+vq_model.dec.resblocks.6.convs1.1.weight_g
+vq_model.dec.resblocks.6.convs1.1.weight_v
+vq_model.dec.resblocks.6.convs1.2.bias
+vq_model.dec.resblocks.6.convs1.2.weight_g
+vq_model.dec.resblocks.6.convs1.2.weight_v
+vq_model.dec.resblocks.6.convs2.0.bias
+vq_model.dec.resblocks.6.convs2.0.weight_g
+vq_model.dec.resblocks.6.convs2.0.weight_v
+vq_model.dec.resblocks.6.convs2.1.bias
+vq_model.dec.resblocks.6.convs2.1.weight_g
+vq_model.dec.resblocks.6.convs2.1.weight_v
+vq_model.dec.resblocks.6.convs2.2.bias
+vq_model.dec.resblocks.6.convs2.2.weight_g
+vq_model.dec.resblocks.6.convs2.2.weight_v
+vq_model.dec.resblocks.7.convs1.0.bias
+vq_model.dec.resblocks.7.convs1.0.weight_g
+vq_model.dec.resblocks.7.convs1.0.weight_v
+vq_model.dec.resblocks.7.convs1.1.bias
+vq_model.dec.resblocks.7.convs1.1.weight_g
+vq_model.dec.resblocks.7.convs1.1.weight_v
+vq_model.dec.resblocks.7.convs1.2.bias
+vq_model.dec.resblocks.7.convs1.2.weight_g
+vq_model.dec.resblocks.7.convs1.2.weight_v
+vq_model.dec.resblocks.7.convs2.0.bias
+vq_model.dec.resblocks.7.convs2.0.weight_g
+vq_model.dec.resblocks.7.convs2.0.weight_v
+vq_model.dec.resblocks.7.convs2.1.bias
+vq_model.dec.resblocks.7.convs2.1.weight_g
+vq_model.dec.resblocks.7.convs2.1.weight_v
+vq_model.dec.resblocks.7.convs2.2.bias
+vq_model.dec.resblocks.7.convs2.2.weight_g
+vq_model.dec.resblocks.7.convs2.2.weight_v
+vq_model.dec.resblocks.8.convs1.0.bias
+vq_model.dec.resblocks.8.convs1.0.weight_g
+vq_model.dec.resblocks.8.convs1.0.weight_v
+vq_model.dec.resblocks.8.convs1.1.bias
+vq_model.dec.resblocks.8.convs1.1.weight_g
+vq_model.dec.resblocks.8.convs1.1.weight_v
+vq_model.dec.resblocks.8.convs1.2.bias
+vq_model.dec.resblocks.8.convs1.2.weight_g
+vq_model.dec.resblocks.8.convs1.2.weight_v
+vq_model.dec.resblocks.8.convs2.0.bias
+vq_model.dec.resblocks.8.convs2.0.weight_g
+vq_model.dec.resblocks.8.convs2.0.weight_v
+vq_model.dec.resblocks.8.convs2.1.bias
+vq_model.dec.resblocks.8.convs2.1.weight_g
+vq_model.dec.resblocks.8.convs2.1.weight_v
+vq_model.dec.resblocks.8.convs2.2.bias
+vq_model.dec.resblocks.8.convs2.2.weight_g
+vq_model.dec.resblocks.8.convs2.2.weight_v
+vq_model.dec.resblocks.9.convs1.0.bias
+vq_model.dec.resblocks.9.convs1.0.weight_g
+vq_model.dec.resblocks.9.convs1.0.weight_v
+vq_model.dec.resblocks.9.convs1.1.bias
+vq_model.dec.resblocks.9.convs1.1.weight_g
+vq_model.dec.resblocks.9.convs1.1.weight_v
+vq_model.dec.resblocks.9.convs1.2.bias
+vq_model.dec.resblocks.9.convs1.2.weight_g
+vq_model.dec.resblocks.9.convs1.2.weight_v
+vq_model.dec.resblocks.9.convs2.0.bias
+vq_model.dec.resblocks.9.convs2.0.weight_g
+vq_model.dec.resblocks.9.convs2.0.weight_v
+vq_model.dec.resblocks.9.convs2.1.bias
+vq_model.dec.resblocks.9.convs2.1.weight_g
+vq_model.dec.resblocks.9.convs2.1.weight_v
+vq_model.dec.resblocks.9.convs2.2.bias
+vq_model.dec.resblocks.9.convs2.2.weight_g
+vq_model.dec.resblocks.9.convs2.2.weight_v
+vq_model.dec.ups.0.bias
+vq_model.dec.ups.0.weight_g
+vq_model.dec.ups.0.weight_v
+vq_model.dec.ups.1.bias
+vq_model.dec.ups.1.weight_g
+vq_model.dec.ups.1.weight_v
+vq_model.dec.ups.2.bias
+vq_model.dec.ups.2.weight_g
+vq_model.dec.ups.2.weight_v
+vq_model.dec.ups.3.bias
+vq_model.dec.ups.3.weight_g
+vq_model.dec.ups.3.weight_v
+vq_model.dec.ups.4.bias
+vq_model.dec.ups.4.weight_g
+vq_model.dec.ups.4.weight_v
+vq_model.enc_p.encoder2.attn_layers.0.conv_k.bias
+vq_model.enc_p.encoder2.attn_layers.0.conv_k.weight
+vq_model.enc_p.encoder2.attn_layers.0.conv_o.bias
+vq_model.enc_p.encoder2.attn_layers.0.conv_o.weight
+vq_model.enc_p.encoder2.attn_layers.0.conv_q.bias
+vq_model.enc_p.encoder2.attn_layers.0.conv_q.weight
+vq_model.enc_p.encoder2.attn_layers.0.conv_v.bias
+vq_model.enc_p.encoder2.attn_layers.0.conv_v.weight
+vq_model.enc_p.encoder2.attn_layers.0.emb_rel_k
+vq_model.enc_p.encoder2.attn_layers.0.emb_rel_v
+vq_model.enc_p.encoder2.attn_layers.1.conv_k.bias
+vq_model.enc_p.encoder2.attn_layers.1.conv_k.weight
+vq_model.enc_p.encoder2.attn_layers.1.conv_o.bias
+vq_model.enc_p.encoder2.attn_layers.1.conv_o.weight
+vq_model.enc_p.encoder2.attn_layers.1.conv_q.bias
+vq_model.enc_p.encoder2.attn_layers.1.conv_q.weight
+vq_model.enc_p.encoder2.attn_layers.1.conv_v.bias
+vq_model.enc_p.encoder2.attn_layers.1.conv_v.weight
+vq_model.enc_p.encoder2.attn_layers.1.emb_rel_k
+vq_model.enc_p.encoder2.attn_layers.1.emb_rel_v
+vq_model.enc_p.encoder2.attn_layers.2.conv_k.bias
+vq_model.enc_p.encoder2.attn_layers.2.conv_k.weight
+vq_model.enc_p.encoder2.attn_layers.2.conv_o.bias
+vq_model.enc_p.encoder2.attn_layers.2.conv_o.weight
+vq_model.enc_p.encoder2.attn_layers.2.conv_q.bias
+vq_model.enc_p.encoder2.attn_layers.2.conv_q.weight
+vq_model.enc_p.encoder2.attn_layers.2.conv_v.bias
+vq_model.enc_p.encoder2.attn_layers.2.conv_v.weight
+vq_model.enc_p.encoder2.attn_layers.2.emb_rel_k
+vq_model.enc_p.encoder2.attn_layers.2.emb_rel_v
+vq_model.enc_p.encoder2.ffn_layers.0.conv_1.bias
+vq_model.enc_p.encoder2.ffn_layers.0.conv_1.weight
+vq_model.enc_p.encoder2.ffn_layers.0.conv_2.bias
+vq_model.enc_p.encoder2.ffn_layers.0.conv_2.weight
+vq_model.enc_p.encoder2.ffn_layers.1.conv_1.bias
+vq_model.enc_p.encoder2.ffn_layers.1.conv_1.weight
+vq_model.enc_p.encoder2.ffn_layers.1.conv_2.bias
+vq_model.enc_p.encoder2.ffn_layers.1.conv_2.weight
+vq_model.enc_p.encoder2.ffn_layers.2.conv_1.bias
+vq_model.enc_p.encoder2.ffn_layers.2.conv_1.weight
+vq_model.enc_p.encoder2.ffn_layers.2.conv_2.bias
+vq_model.enc_p.encoder2.ffn_layers.2.conv_2.weight
+vq_model.enc_p.encoder2.norm_layers_1.0.beta
+vq_model.enc_p.encoder2.norm_layers_1.0.gamma
+vq_model.enc_p.encoder2.norm_layers_1.1.beta
+vq_model.enc_p.encoder2.norm_layers_1.1.gamma
+vq_model.enc_p.encoder2.norm_layers_1.2.beta
+vq_model.enc_p.encoder2.norm_layers_1.2.gamma
+vq_model.enc_p.encoder2.norm_layers_2.0.beta
+vq_model.enc_p.encoder2.norm_layers_2.0.gamma
+vq_model.enc_p.encoder2.norm_layers_2.1.beta
+vq_model.enc_p.encoder2.norm_layers_2.1.gamma
+vq_model.enc_p.encoder2.norm_layers_2.2.beta
+vq_model.enc_p.encoder2.norm_layers_2.2.gamma
+vq_model.enc_p.encoder_ssl.attn_layers.0.conv_k.bias
+vq_model.enc_p.encoder_ssl.attn_layers.0.conv_k.weight
+vq_model.enc_p.encoder_ssl.attn_layers.0.conv_o.bias
+vq_model.enc_p.encoder_ssl.attn_layers.0.conv_o.weight
+vq_model.enc_p.encoder_ssl.attn_layers.0.conv_q.bias
+vq_model.enc_p.encoder_ssl.attn_layers.0.conv_q.weight
+vq_model.enc_p.encoder_ssl.attn_layers.0.conv_v.bias
+vq_model.enc_p.encoder_ssl.attn_layers.0.conv_v.weight
+vq_model.enc_p.encoder_ssl.attn_layers.0.emb_rel_k
+vq_model.enc_p.encoder_ssl.attn_layers.0.emb_rel_v
+vq_model.enc_p.encoder_ssl.attn_layers.1.conv_k.bias
+vq_model.enc_p.encoder_ssl.attn_layers.1.conv_k.weight
+vq_model.enc_p.encoder_ssl.attn_layers.1.conv_o.bias
+vq_model.enc_p.encoder_ssl.attn_layers.1.conv_o.weight
+vq_model.enc_p.encoder_ssl.attn_layers.1.conv_q.bias
+vq_model.enc_p.encoder_ssl.attn_layers.1.conv_q.weight
+vq_model.enc_p.encoder_ssl.attn_layers.1.conv_v.bias
+vq_model.enc_p.encoder_ssl.attn_layers.1.conv_v.weight
+vq_model.enc_p.encoder_ssl.attn_layers.1.emb_rel_k
+vq_model.enc_p.encoder_ssl.attn_layers.1.emb_rel_v
+vq_model.enc_p.encoder_ssl.attn_layers.2.conv_k.bias
+vq_model.enc_p.encoder_ssl.attn_layers.2.conv_k.weight
+vq_model.enc_p.encoder_ssl.attn_layers.2.conv_o.bias
+vq_model.enc_p.encoder_ssl.attn_layers.2.conv_o.weight
+vq_model.enc_p.encoder_ssl.attn_layers.2.conv_q.bias
+vq_model.enc_p.encoder_ssl.attn_layers.2.conv_q.weight
+vq_model.enc_p.encoder_ssl.attn_layers.2.conv_v.bias
+vq_model.enc_p.encoder_ssl.attn_layers.2.conv_v.weight
+vq_model.enc_p.encoder_ssl.attn_layers.2.emb_rel_k
+vq_model.enc_p.encoder_ssl.attn_layers.2.emb_rel_v
+vq_model.enc_p.encoder_ssl.ffn_layers.0.conv_1.bias
+vq_model.enc_p.encoder_ssl.ffn_layers.0.conv_1.weight
+vq_model.enc_p.encoder_ssl.ffn_layers.0.conv_2.bias
+vq_model.enc_p.encoder_ssl.ffn_layers.0.conv_2.weight
+vq_model.enc_p.encoder_ssl.ffn_layers.1.conv_1.bias
+vq_model.enc_p.encoder_ssl.ffn_layers.1.conv_1.weight
+vq_model.enc_p.encoder_ssl.ffn_layers.1.conv_2.bias
+vq_model.enc_p.encoder_ssl.ffn_layers.1.conv_2.weight
+vq_model.enc_p.encoder_ssl.ffn_layers.2.conv_1.bias
+vq_model.enc_p.encoder_ssl.ffn_layers.2.conv_1.weight
+vq_model.enc_p.encoder_ssl.ffn_layers.2.conv_2.bias
+vq_model.enc_p.encoder_ssl.ffn_layers.2.conv_2.weight
+vq_model.enc_p.encoder_ssl.norm_layers_1.0.beta
+vq_model.enc_p.encoder_ssl.norm_layers_1.0.gamma
+vq_model.enc_p.encoder_ssl.norm_layers_1.1.beta
+vq_model.enc_p.encoder_ssl.norm_layers_1.1.gamma
+vq_model.enc_p.encoder_ssl.norm_layers_1.2.beta
+vq_model.enc_p.encoder_ssl.norm_layers_1.2.gamma
+vq_model.enc_p.encoder_ssl.norm_layers_2.0.beta
+vq_model.enc_p.encoder_ssl.norm_layers_2.0.gamma
+vq_model.enc_p.encoder_ssl.norm_layers_2.1.beta
+vq_model.enc_p.encoder_ssl.norm_layers_2.1.gamma
+vq_model.enc_p.encoder_ssl.norm_layers_2.2.beta
+vq_model.enc_p.encoder_ssl.norm_layers_2.2.gamma
+vq_model.enc_p.encoder_text.attn_layers.0.conv_k.bias
+vq_model.enc_p.encoder_text.attn_layers.0.conv_k.weight
+vq_model.enc_p.encoder_text.attn_layers.0.conv_o.bias
+vq_model.enc_p.encoder_text.attn_layers.0.conv_o.weight
+vq_model.enc_p.encoder_text.attn_layers.0.conv_q.bias
+vq_model.enc_p.encoder_text.attn_layers.0.conv_q.weight
+vq_model.enc_p.encoder_text.attn_layers.0.conv_v.bias
+vq_model.enc_p.encoder_text.attn_layers.0.conv_v.weight
+vq_model.enc_p.encoder_text.attn_layers.0.emb_rel_k
+vq_model.enc_p.encoder_text.attn_layers.0.emb_rel_v
+vq_model.enc_p.encoder_text.attn_layers.1.conv_k.bias
+vq_model.enc_p.encoder_text.attn_layers.1.conv_k.weight
+vq_model.enc_p.encoder_text.attn_layers.1.conv_o.bias
+vq_model.enc_p.encoder_text.attn_layers.1.conv_o.weight
+vq_model.enc_p.encoder_text.attn_layers.1.conv_q.bias
+vq_model.enc_p.encoder_text.attn_layers.1.conv_q.weight
+vq_model.enc_p.encoder_text.attn_layers.1.conv_v.bias
+vq_model.enc_p.encoder_text.attn_layers.1.conv_v.weight
+vq_model.enc_p.encoder_text.attn_layers.1.emb_rel_k
+vq_model.enc_p.encoder_text.attn_layers.1.emb_rel_v
+vq_model.enc_p.encoder_text.attn_layers.2.conv_k.bias
+vq_model.enc_p.encoder_text.attn_layers.2.conv_k.weight
+vq_model.enc_p.encoder_text.attn_layers.2.conv_o.bias
+vq_model.enc_p.encoder_text.attn_layers.2.conv_o.weight
+vq_model.enc_p.encoder_text.attn_layers.2.conv_q.bias
+vq_model.enc_p.encoder_text.attn_layers.2.conv_q.weight
+vq_model.enc_p.encoder_text.attn_layers.2.conv_v.bias
+vq_model.enc_p.encoder_text.attn_layers.2.conv_v.weight
+vq_model.enc_p.encoder_text.attn_layers.2.emb_rel_k
+vq_model.enc_p.encoder_text.attn_layers.2.emb_rel_v
+vq_model.enc_p.encoder_text.attn_layers.3.conv_k.bias
+vq_model.enc_p.encoder_text.attn_layers.3.conv_k.weight
+vq_model.enc_p.encoder_text.attn_layers.3.conv_o.bias
+vq_model.enc_p.encoder_text.attn_layers.3.conv_o.weight
+vq_model.enc_p.encoder_text.attn_layers.3.conv_q.bias
+vq_model.enc_p.encoder_text.attn_layers.3.conv_q.weight
+vq_model.enc_p.encoder_text.attn_layers.3.conv_v.bias
+vq_model.enc_p.encoder_text.attn_layers.3.conv_v.weight
+vq_model.enc_p.encoder_text.attn_layers.3.emb_rel_k
+vq_model.enc_p.encoder_text.attn_layers.3.emb_rel_v
+vq_model.enc_p.encoder_text.attn_layers.4.conv_k.bias
+vq_model.enc_p.encoder_text.attn_layers.4.conv_k.weight
+vq_model.enc_p.encoder_text.attn_layers.4.conv_o.bias
+vq_model.enc_p.encoder_text.attn_layers.4.conv_o.weight
+vq_model.enc_p.encoder_text.attn_layers.4.conv_q.bias
+vq_model.enc_p.encoder_text.attn_layers.4.conv_q.weight
+vq_model.enc_p.encoder_text.attn_layers.4.conv_v.bias
+vq_model.enc_p.encoder_text.attn_layers.4.conv_v.weight
+vq_model.enc_p.encoder_text.attn_layers.4.emb_rel_k
+vq_model.enc_p.encoder_text.attn_layers.4.emb_rel_v
+vq_model.enc_p.encoder_text.attn_layers.5.conv_k.bias
+vq_model.enc_p.encoder_text.attn_layers.5.conv_k.weight
+vq_model.enc_p.encoder_text.attn_layers.5.conv_o.bias
+vq_model.enc_p.encoder_text.attn_layers.5.conv_o.weight
+vq_model.enc_p.encoder_text.attn_layers.5.conv_q.bias
+vq_model.enc_p.encoder_text.attn_layers.5.conv_q.weight
+vq_model.enc_p.encoder_text.attn_layers.5.conv_v.bias
+vq_model.enc_p.encoder_text.attn_layers.5.conv_v.weight
+vq_model.enc_p.encoder_text.attn_layers.5.emb_rel_k
+vq_model.enc_p.encoder_text.attn_layers.5.emb_rel_v
+vq_model.enc_p.encoder_text.ffn_layers.0.conv_1.bias
+vq_model.enc_p.encoder_text.ffn_layers.0.conv_1.weight
+vq_model.enc_p.encoder_text.ffn_layers.0.conv_2.bias
+vq_model.enc_p.encoder_text.ffn_layers.0.conv_2.weight
+vq_model.enc_p.encoder_text.ffn_layers.1.conv_1.bias
+vq_model.enc_p.encoder_text.ffn_layers.1.conv_1.weight
+vq_model.enc_p.encoder_text.ffn_layers.1.conv_2.bias
+vq_model.enc_p.encoder_text.ffn_layers.1.conv_2.weight
+vq_model.enc_p.encoder_text.ffn_layers.2.conv_1.bias
+vq_model.enc_p.encoder_text.ffn_layers.2.conv_1.weight
+vq_model.enc_p.encoder_text.ffn_layers.2.conv_2.bias
+vq_model.enc_p.encoder_text.ffn_layers.2.conv_2.weight
+vq_model.enc_p.encoder_text.ffn_layers.3.conv_1.bias
+vq_model.enc_p.encoder_text.ffn_layers.3.conv_1.weight
+vq_model.enc_p.encoder_text.ffn_layers.3.conv_2.bias
+vq_model.enc_p.encoder_text.ffn_layers.3.conv_2.weight
+vq_model.enc_p.encoder_text.ffn_layers.4.conv_1.bias
+vq_model.enc_p.encoder_text.ffn_layers.4.conv_1.weight
+vq_model.enc_p.encoder_text.ffn_layers.4.conv_2.bias
+vq_model.enc_p.encoder_text.ffn_layers.4.conv_2.weight
+vq_model.enc_p.encoder_text.ffn_layers.5.conv_1.bias
+vq_model.enc_p.encoder_text.ffn_layers.5.conv_1.weight
+vq_model.enc_p.encoder_text.ffn_layers.5.conv_2.bias
+vq_model.enc_p.encoder_text.ffn_layers.5.conv_2.weight
+vq_model.enc_p.encoder_text.norm_layers_1.0.beta
+vq_model.enc_p.encoder_text.norm_layers_1.0.gamma
+vq_model.enc_p.encoder_text.norm_layers_1.1.beta
+vq_model.enc_p.encoder_text.norm_layers_1.1.gamma
+vq_model.enc_p.encoder_text.norm_layers_1.2.beta
+vq_model.enc_p.encoder_text.norm_layers_1.2.gamma
+vq_model.enc_p.encoder_text.norm_layers_1.3.beta
+vq_model.enc_p.encoder_text.norm_layers_1.3.gamma
+vq_model.enc_p.encoder_text.norm_layers_1.4.beta
+vq_model.enc_p.encoder_text.norm_layers_1.4.gamma
+vq_model.enc_p.encoder_text.norm_layers_1.5.beta
+vq_model.enc_p.encoder_text.norm_layers_1.5.gamma
+vq_model.enc_p.encoder_text.norm_layers_2.0.beta
+vq_model.enc_p.encoder_text.norm_layers_2.0.gamma
+vq_model.enc_p.encoder_text.norm_layers_2.1.beta
+vq_model.enc_p.encoder_text.norm_layers_2.1.gamma
+vq_model.enc_p.encoder_text.norm_layers_2.2.beta
+vq_model.enc_p.encoder_text.norm_layers_2.2.gamma
+vq_model.enc_p.encoder_text.norm_layers_2.3.beta
+vq_model.enc_p.encoder_text.norm_layers_2.3.gamma
+vq_model.enc_p.encoder_text.norm_layers_2.4.beta
+vq_model.enc_p.encoder_text.norm_layers_2.4.gamma
+vq_model.enc_p.encoder_text.norm_layers_2.5.beta
+vq_model.enc_p.encoder_text.norm_layers_2.5.gamma
+vq_model.enc_p.mrte.c_post.bias
+vq_model.enc_p.mrte.c_post.weight
+vq_model.enc_p.mrte.c_pre.bias
+vq_model.enc_p.mrte.c_pre.weight
+vq_model.enc_p.mrte.cross_attention.conv_k.bias
+vq_model.enc_p.mrte.cross_attention.conv_k.weight
+vq_model.enc_p.mrte.cross_attention.conv_o.bias
+vq_model.enc_p.mrte.cross_attention.conv_o.weight
+vq_model.enc_p.mrte.cross_attention.conv_q.bias
+vq_model.enc_p.mrte.cross_attention.conv_q.weight
+vq_model.enc_p.mrte.cross_attention.conv_v.bias
+vq_model.enc_p.mrte.cross_attention.conv_v.weight
+vq_model.enc_p.mrte.text_pre.bias
+vq_model.enc_p.mrte.text_pre.weight
+vq_model.enc_p.proj.bias
+vq_model.enc_p.proj.weight
+vq_model.enc_p.ssl_proj.bias
+vq_model.enc_p.ssl_proj.weight
+vq_model.enc_p.text_embedding.weight
+vq_model.flow.flows.0.enc.cond_layer.bias
+vq_model.flow.flows.0.enc.cond_layer.weight_g
+vq_model.flow.flows.0.enc.cond_layer.weight_v
+vq_model.flow.flows.0.enc.in_layers.0.bias
+vq_model.flow.flows.0.enc.in_layers.0.weight_g
+vq_model.flow.flows.0.enc.in_layers.0.weight_v
+vq_model.flow.flows.0.enc.in_layers.1.bias
+vq_model.flow.flows.0.enc.in_layers.1.weight_g
+vq_model.flow.flows.0.enc.in_layers.1.weight_v
+vq_model.flow.flows.0.enc.in_layers.2.bias
+vq_model.flow.flows.0.enc.in_layers.2.weight_g
+vq_model.flow.flows.0.enc.in_layers.2.weight_v
+vq_model.flow.flows.0.enc.in_layers.3.bias
+vq_model.flow.flows.0.enc.in_layers.3.weight_g
+vq_model.flow.flows.0.enc.in_layers.3.weight_v
+vq_model.flow.flows.0.enc.res_skip_layers.0.bias
+vq_model.flow.flows.0.enc.res_skip_layers.0.weight_g
+vq_model.flow.flows.0.enc.res_skip_layers.0.weight_v
+vq_model.flow.flows.0.enc.res_skip_layers.1.bias
+vq_model.flow.flows.0.enc.res_skip_layers.1.weight_g
+vq_model.flow.flows.0.enc.res_skip_layers.1.weight_v
+vq_model.flow.flows.0.enc.res_skip_layers.2.bias
+vq_model.flow.flows.0.enc.res_skip_layers.2.weight_g
+vq_model.flow.flows.0.enc.res_skip_layers.2.weight_v
+vq_model.flow.flows.0.enc.res_skip_layers.3.bias
+vq_model.flow.flows.0.enc.res_skip_layers.3.weight_g
+vq_model.flow.flows.0.enc.res_skip_layers.3.weight_v
+vq_model.flow.flows.0.post.bias
+vq_model.flow.flows.0.post.weight
+vq_model.flow.flows.0.pre.bias
+vq_model.flow.flows.0.pre.weight
+vq_model.flow.flows.2.enc.cond_layer.bias
+vq_model.flow.flows.2.enc.cond_layer.weight_g
+vq_model.flow.flows.2.enc.cond_layer.weight_v
+vq_model.flow.flows.2.enc.in_layers.0.bias
+vq_model.flow.flows.2.enc.in_layers.0.weight_g
+vq_model.flow.flows.2.enc.in_layers.0.weight_v
+vq_model.flow.flows.2.enc.in_layers.1.bias
+vq_model.flow.flows.2.enc.in_layers.1.weight_g
+vq_model.flow.flows.2.enc.in_layers.1.weight_v
+vq_model.flow.flows.2.enc.in_layers.2.bias
+vq_model.flow.flows.2.enc.in_layers.2.weight_g
+vq_model.flow.flows.2.enc.in_layers.2.weight_v
+vq_model.flow.flows.2.enc.in_layers.3.bias
+vq_model.flow.flows.2.enc.in_layers.3.weight_g
+vq_model.flow.flows.2.enc.in_layers.3.weight_v
+vq_model.flow.flows.2.enc.res_skip_layers.0.bias
+vq_model.flow.flows.2.enc.res_skip_layers.0.weight_g
+vq_model.flow.flows.2.enc.res_skip_layers.0.weight_v
+vq_model.flow.flows.2.enc.res_skip_layers.1.bias
+vq_model.flow.flows.2.enc.res_skip_layers.1.weight_g
+vq_model.flow.flows.2.enc.res_skip_layers.1.weight_v
+vq_model.flow.flows.2.enc.res_skip_layers.2.bias
+vq_model.flow.flows.2.enc.res_skip_layers.2.weight_g
+vq_model.flow.flows.2.enc.res_skip_layers.2.weight_v
+vq_model.flow.flows.2.enc.res_skip_layers.3.bias
+vq_model.flow.flows.2.enc.res_skip_layers.3.weight_g
+vq_model.flow.flows.2.enc.res_skip_layers.3.weight_v
+vq_model.flow.flows.2.post.bias
+vq_model.flow.flows.2.post.weight
+vq_model.flow.flows.2.pre.bias
+vq_model.flow.flows.2.pre.weight
+vq_model.flow.flows.4.enc.cond_layer.bias
+vq_model.flow.flows.4.enc.cond_layer.weight_g
+vq_model.flow.flows.4.enc.cond_layer.weight_v
+vq_model.flow.flows.4.enc.in_layers.0.bias
+vq_model.flow.flows.4.enc.in_layers.0.weight_g
+vq_model.flow.flows.4.enc.in_layers.0.weight_v
+vq_model.flow.flows.4.enc.in_layers.1.bias
+vq_model.flow.flows.4.enc.in_layers.1.weight_g
+vq_model.flow.flows.4.enc.in_layers.1.weight_v
+vq_model.flow.flows.4.enc.in_layers.2.bias
+vq_model.flow.flows.4.enc.in_layers.2.weight_g
+vq_model.flow.flows.4.enc.in_layers.2.weight_v
+vq_model.flow.flows.4.enc.in_layers.3.bias
+vq_model.flow.flows.4.enc.in_layers.3.weight_g
+vq_model.flow.flows.4.enc.in_layers.3.weight_v
+vq_model.flow.flows.4.enc.res_skip_layers.0.bias
+vq_model.flow.flows.4.enc.res_skip_layers.0.weight_g
+vq_model.flow.flows.4.enc.res_skip_layers.0.weight_v
+vq_model.flow.flows.4.enc.res_skip_layers.1.bias
+vq_model.flow.flows.4.enc.res_skip_layers.1.weight_g
+vq_model.flow.flows.4.enc.res_skip_layers.1.weight_v
+vq_model.flow.flows.4.enc.res_skip_layers.2.bias
+vq_model.flow.flows.4.enc.res_skip_layers.2.weight_g
+vq_model.flow.flows.4.enc.res_skip_layers.2.weight_v
+vq_model.flow.flows.4.enc.res_skip_layers.3.bias
+vq_model.flow.flows.4.enc.res_skip_layers.3.weight_g
+vq_model.flow.flows.4.enc.res_skip_layers.3.weight_v
+vq_model.flow.flows.4.post.bias
+vq_model.flow.flows.4.post.weight
+vq_model.flow.flows.4.pre.bias
+vq_model.flow.flows.4.pre.weight
+vq_model.flow.flows.6.enc.cond_layer.bias
+vq_model.flow.flows.6.enc.cond_layer.weight_g
+vq_model.flow.flows.6.enc.cond_layer.weight_v
+vq_model.flow.flows.6.enc.in_layers.0.bias
+vq_model.flow.flows.6.enc.in_layers.0.weight_g
+vq_model.flow.flows.6.enc.in_layers.0.weight_v
+vq_model.flow.flows.6.enc.in_layers.1.bias
+vq_model.flow.flows.6.enc.in_layers.1.weight_g
+vq_model.flow.flows.6.enc.in_layers.1.weight_v
+vq_model.flow.flows.6.enc.in_layers.2.bias
+vq_model.flow.flows.6.enc.in_layers.2.weight_g
+vq_model.flow.flows.6.enc.in_layers.2.weight_v
+vq_model.flow.flows.6.enc.in_layers.3.bias
+vq_model.flow.flows.6.enc.in_layers.3.weight_g
+vq_model.flow.flows.6.enc.in_layers.3.weight_v
+vq_model.flow.flows.6.enc.res_skip_layers.0.bias
+vq_model.flow.flows.6.enc.res_skip_layers.0.weight_g
+vq_model.flow.flows.6.enc.res_skip_layers.0.weight_v
+vq_model.flow.flows.6.enc.res_skip_layers.1.bias
+vq_model.flow.flows.6.enc.res_skip_layers.1.weight_g
+vq_model.flow.flows.6.enc.res_skip_layers.1.weight_v
+vq_model.flow.flows.6.enc.res_skip_layers.2.bias
+vq_model.flow.flows.6.enc.res_skip_layers.2.weight_g
+vq_model.flow.flows.6.enc.res_skip_layers.2.weight_v
+vq_model.flow.flows.6.enc.res_skip_layers.3.bias
+vq_model.flow.flows.6.enc.res_skip_layers.3.weight_g
+vq_model.flow.flows.6.enc.res_skip_layers.3.weight_v
+vq_model.flow.flows.6.post.bias
+vq_model.flow.flows.6.post.weight
+vq_model.flow.flows.6.pre.bias
+vq_model.flow.flows.6.pre.weight
+vq_model.quantizer.vq.layers.0._codebook.embed
+vq_model.ref_enc.fc.fc.bias
+vq_model.ref_enc.fc.fc.weight
+vq_model.ref_enc.slf_attn.fc.bias
+vq_model.ref_enc.slf_attn.fc.weight
+vq_model.ref_enc.slf_attn.w_ks.bias
+vq_model.ref_enc.slf_attn.w_ks.weight
+vq_model.ref_enc.slf_attn.w_qs.bias
+vq_model.ref_enc.slf_attn.w_qs.weight
+vq_model.ref_enc.slf_attn.w_vs.bias
+vq_model.ref_enc.slf_attn.w_vs.weight
+vq_model.ref_enc.spectral.0.fc.bias
+vq_model.ref_enc.spectral.0.fc.weight
+vq_model.ref_enc.spectral.3.fc.bias
+vq_model.ref_enc.spectral.3.fc.weight
+vq_model.ref_enc.temporal.0.conv1.conv.bias
+vq_model.ref_enc.temporal.0.conv1.conv.weight
+vq_model.ref_enc.temporal.1.conv1.conv.bias
+vq_model.ref_enc.temporal.1.conv1.conv.weight
diff --git a/genie_tts/Data/v2ProPlus/Keys/prompt_encoder_weights.txt b/genie_tts/Data/v2ProPlus/Keys/prompt_encoder_weights.txt
index 8ae9345f6b061d27362302611f0fe08c8091eb29..fbb01e9ef646f2c9472bda37d78cfd004724dd4a 100644
--- a/genie_tts/Data/v2ProPlus/Keys/prompt_encoder_weights.txt
+++ b/genie_tts/Data/v2ProPlus/Keys/prompt_encoder_weights.txt
@@ -1,23 +1,23 @@
-ref_enc.spectral.0.fc.weight
-ref_enc.spectral.0.fc.bias
-ref_enc.spectral.3.fc.weight
-ref_enc.spectral.3.fc.bias
-ref_enc.temporal.0.conv1.conv.weight
-ref_enc.temporal.0.conv1.conv.bias
-ref_enc.temporal.1.conv1.conv.weight
-ref_enc.temporal.1.conv1.conv.bias
-ref_enc.slf_attn.w_qs.weight
-ref_enc.slf_attn.w_qs.bias
-ref_enc.slf_attn.w_ks.weight
-ref_enc.slf_attn.w_ks.bias
-ref_enc.slf_attn.w_vs.weight
-ref_enc.slf_attn.w_vs.bias
-ref_enc.slf_attn.fc.weight
-ref_enc.slf_attn.fc.bias
-ref_enc.fc.fc.weight
-ref_enc.fc.fc.bias
-sv_emb.weight
-sv_emb.bias
-ge_to512.weight
-ge_to512.bias
-prelu.weight
+ref_enc.spectral.0.fc.weight
+ref_enc.spectral.0.fc.bias
+ref_enc.spectral.3.fc.weight
+ref_enc.spectral.3.fc.bias
+ref_enc.temporal.0.conv1.conv.weight
+ref_enc.temporal.0.conv1.conv.bias
+ref_enc.temporal.1.conv1.conv.weight
+ref_enc.temporal.1.conv1.conv.bias
+ref_enc.slf_attn.w_qs.weight
+ref_enc.slf_attn.w_qs.bias
+ref_enc.slf_attn.w_ks.weight
+ref_enc.slf_attn.w_ks.bias
+ref_enc.slf_attn.w_vs.weight
+ref_enc.slf_attn.w_vs.bias
+ref_enc.slf_attn.fc.weight
+ref_enc.slf_attn.fc.bias
+ref_enc.fc.fc.weight
+ref_enc.fc.fc.bias
+sv_emb.weight
+sv_emb.bias
+ge_to512.weight
+ge_to512.bias
+prelu.weight
diff --git a/genie_tts/Data/v2ProPlus/Keys/vits_weights.txt b/genie_tts/Data/v2ProPlus/Keys/vits_weights.txt
index 4ed033cab8dfa1bd065ed90ae62e8e02b093c03b..7b6bc49c812af16ebf0035b2abbad5cc17609dfc 100644
--- a/genie_tts/Data/v2ProPlus/Keys/vits_weights.txt
+++ b/genie_tts/Data/v2ProPlus/Keys/vits_weights.txt
@@ -1,650 +1,650 @@
-vq_model.enc_p.ssl_proj.weight
-vq_model.enc_p.ssl_proj.bias
-vq_model.enc_p.encoder_ssl.attn_layers.0.emb_rel_k
-vq_model.enc_p.encoder_ssl.attn_layers.0.emb_rel_v
-vq_model.enc_p.encoder_ssl.attn_layers.0.conv_q.weight
-vq_model.enc_p.encoder_ssl.attn_layers.0.conv_q.bias
-vq_model.enc_p.encoder_ssl.attn_layers.0.conv_k.weight
-vq_model.enc_p.encoder_ssl.attn_layers.0.conv_k.bias
-vq_model.enc_p.encoder_ssl.attn_layers.0.conv_v.weight
-vq_model.enc_p.encoder_ssl.attn_layers.0.conv_v.bias
-vq_model.enc_p.encoder_ssl.attn_layers.0.conv_o.weight
-vq_model.enc_p.encoder_ssl.attn_layers.0.conv_o.bias
-vq_model.enc_p.encoder_ssl.attn_layers.1.emb_rel_k
-vq_model.enc_p.encoder_ssl.attn_layers.1.emb_rel_v
-vq_model.enc_p.encoder_ssl.attn_layers.1.conv_q.weight
-vq_model.enc_p.encoder_ssl.attn_layers.1.conv_q.bias
-vq_model.enc_p.encoder_ssl.attn_layers.1.conv_k.weight
-vq_model.enc_p.encoder_ssl.attn_layers.1.conv_k.bias
-vq_model.enc_p.encoder_ssl.attn_layers.1.conv_v.weight
-vq_model.enc_p.encoder_ssl.attn_layers.1.conv_v.bias
-vq_model.enc_p.encoder_ssl.attn_layers.1.conv_o.weight
-vq_model.enc_p.encoder_ssl.attn_layers.1.conv_o.bias
-vq_model.enc_p.encoder_ssl.attn_layers.2.emb_rel_k
-vq_model.enc_p.encoder_ssl.attn_layers.2.emb_rel_v
-vq_model.enc_p.encoder_ssl.attn_layers.2.conv_q.weight
-vq_model.enc_p.encoder_ssl.attn_layers.2.conv_q.bias
-vq_model.enc_p.encoder_ssl.attn_layers.2.conv_k.weight
-vq_model.enc_p.encoder_ssl.attn_layers.2.conv_k.bias
-vq_model.enc_p.encoder_ssl.attn_layers.2.conv_v.weight
-vq_model.enc_p.encoder_ssl.attn_layers.2.conv_v.bias
-vq_model.enc_p.encoder_ssl.attn_layers.2.conv_o.weight
-vq_model.enc_p.encoder_ssl.attn_layers.2.conv_o.bias
-vq_model.enc_p.encoder_ssl.norm_layers_1.0.gamma
-vq_model.enc_p.encoder_ssl.norm_layers_1.0.beta
-vq_model.enc_p.encoder_ssl.norm_layers_1.1.gamma
-vq_model.enc_p.encoder_ssl.norm_layers_1.1.beta
-vq_model.enc_p.encoder_ssl.norm_layers_1.2.gamma
-vq_model.enc_p.encoder_ssl.norm_layers_1.2.beta
-vq_model.enc_p.encoder_ssl.ffn_layers.0.conv_1.weight
-vq_model.enc_p.encoder_ssl.ffn_layers.0.conv_1.bias
-vq_model.enc_p.encoder_ssl.ffn_layers.0.conv_2.weight
-vq_model.enc_p.encoder_ssl.ffn_layers.0.conv_2.bias
-vq_model.enc_p.encoder_ssl.ffn_layers.1.conv_1.weight
-vq_model.enc_p.encoder_ssl.ffn_layers.1.conv_1.bias
-vq_model.enc_p.encoder_ssl.ffn_layers.1.conv_2.weight
-vq_model.enc_p.encoder_ssl.ffn_layers.1.conv_2.bias
-vq_model.enc_p.encoder_ssl.ffn_layers.2.conv_1.weight
-vq_model.enc_p.encoder_ssl.ffn_layers.2.conv_1.bias
-vq_model.enc_p.encoder_ssl.ffn_layers.2.conv_2.weight
-vq_model.enc_p.encoder_ssl.ffn_layers.2.conv_2.bias
-vq_model.enc_p.encoder_ssl.norm_layers_2.0.gamma
-vq_model.enc_p.encoder_ssl.norm_layers_2.0.beta
-vq_model.enc_p.encoder_ssl.norm_layers_2.1.gamma
-vq_model.enc_p.encoder_ssl.norm_layers_2.1.beta
-vq_model.enc_p.encoder_ssl.norm_layers_2.2.gamma
-vq_model.enc_p.encoder_ssl.norm_layers_2.2.beta
-vq_model.enc_p.encoder_text.attn_layers.0.emb_rel_k
-vq_model.enc_p.encoder_text.attn_layers.0.emb_rel_v
-vq_model.enc_p.encoder_text.attn_layers.0.conv_q.weight
-vq_model.enc_p.encoder_text.attn_layers.0.conv_q.bias
-vq_model.enc_p.encoder_text.attn_layers.0.conv_k.weight
-vq_model.enc_p.encoder_text.attn_layers.0.conv_k.bias
-vq_model.enc_p.encoder_text.attn_layers.0.conv_v.weight
-vq_model.enc_p.encoder_text.attn_layers.0.conv_v.bias
-vq_model.enc_p.encoder_text.attn_layers.0.conv_o.weight
-vq_model.enc_p.encoder_text.attn_layers.0.conv_o.bias
-vq_model.enc_p.encoder_text.attn_layers.1.emb_rel_k
-vq_model.enc_p.encoder_text.attn_layers.1.emb_rel_v
-vq_model.enc_p.encoder_text.attn_layers.1.conv_q.weight
-vq_model.enc_p.encoder_text.attn_layers.1.conv_q.bias
-vq_model.enc_p.encoder_text.attn_layers.1.conv_k.weight
-vq_model.enc_p.encoder_text.attn_layers.1.conv_k.bias
-vq_model.enc_p.encoder_text.attn_layers.1.conv_v.weight
-vq_model.enc_p.encoder_text.attn_layers.1.conv_v.bias
-vq_model.enc_p.encoder_text.attn_layers.1.conv_o.weight
-vq_model.enc_p.encoder_text.attn_layers.1.conv_o.bias
-vq_model.enc_p.encoder_text.attn_layers.2.emb_rel_k
-vq_model.enc_p.encoder_text.attn_layers.2.emb_rel_v
-vq_model.enc_p.encoder_text.attn_layers.2.conv_q.weight
-vq_model.enc_p.encoder_text.attn_layers.2.conv_q.bias
-vq_model.enc_p.encoder_text.attn_layers.2.conv_k.weight
-vq_model.enc_p.encoder_text.attn_layers.2.conv_k.bias
-vq_model.enc_p.encoder_text.attn_layers.2.conv_v.weight
-vq_model.enc_p.encoder_text.attn_layers.2.conv_v.bias
-vq_model.enc_p.encoder_text.attn_layers.2.conv_o.weight
-vq_model.enc_p.encoder_text.attn_layers.2.conv_o.bias
-vq_model.enc_p.encoder_text.attn_layers.3.emb_rel_k
-vq_model.enc_p.encoder_text.attn_layers.3.emb_rel_v
-vq_model.enc_p.encoder_text.attn_layers.3.conv_q.weight
-vq_model.enc_p.encoder_text.attn_layers.3.conv_q.bias
-vq_model.enc_p.encoder_text.attn_layers.3.conv_k.weight
-vq_model.enc_p.encoder_text.attn_layers.3.conv_k.bias
-vq_model.enc_p.encoder_text.attn_layers.3.conv_v.weight
-vq_model.enc_p.encoder_text.attn_layers.3.conv_v.bias
-vq_model.enc_p.encoder_text.attn_layers.3.conv_o.weight
-vq_model.enc_p.encoder_text.attn_layers.3.conv_o.bias
-vq_model.enc_p.encoder_text.attn_layers.4.emb_rel_k
-vq_model.enc_p.encoder_text.attn_layers.4.emb_rel_v
-vq_model.enc_p.encoder_text.attn_layers.4.conv_q.weight
-vq_model.enc_p.encoder_text.attn_layers.4.conv_q.bias
-vq_model.enc_p.encoder_text.attn_layers.4.conv_k.weight
-vq_model.enc_p.encoder_text.attn_layers.4.conv_k.bias
-vq_model.enc_p.encoder_text.attn_layers.4.conv_v.weight
-vq_model.enc_p.encoder_text.attn_layers.4.conv_v.bias
-vq_model.enc_p.encoder_text.attn_layers.4.conv_o.weight
-vq_model.enc_p.encoder_text.attn_layers.4.conv_o.bias
-vq_model.enc_p.encoder_text.attn_layers.5.emb_rel_k
-vq_model.enc_p.encoder_text.attn_layers.5.emb_rel_v
-vq_model.enc_p.encoder_text.attn_layers.5.conv_q.weight
-vq_model.enc_p.encoder_text.attn_layers.5.conv_q.bias
-vq_model.enc_p.encoder_text.attn_layers.5.conv_k.weight
-vq_model.enc_p.encoder_text.attn_layers.5.conv_k.bias
-vq_model.enc_p.encoder_text.attn_layers.5.conv_v.weight
-vq_model.enc_p.encoder_text.attn_layers.5.conv_v.bias
-vq_model.enc_p.encoder_text.attn_layers.5.conv_o.weight
-vq_model.enc_p.encoder_text.attn_layers.5.conv_o.bias
-vq_model.enc_p.encoder_text.norm_layers_1.0.gamma
-vq_model.enc_p.encoder_text.norm_layers_1.0.beta
-vq_model.enc_p.encoder_text.norm_layers_1.1.gamma
-vq_model.enc_p.encoder_text.norm_layers_1.1.beta
-vq_model.enc_p.encoder_text.norm_layers_1.2.gamma
-vq_model.enc_p.encoder_text.norm_layers_1.2.beta
-vq_model.enc_p.encoder_text.norm_layers_1.3.gamma
-vq_model.enc_p.encoder_text.norm_layers_1.3.beta
-vq_model.enc_p.encoder_text.norm_layers_1.4.gamma
-vq_model.enc_p.encoder_text.norm_layers_1.4.beta
-vq_model.enc_p.encoder_text.norm_layers_1.5.gamma
-vq_model.enc_p.encoder_text.norm_layers_1.5.beta
-vq_model.enc_p.encoder_text.ffn_layers.0.conv_1.weight
-vq_model.enc_p.encoder_text.ffn_layers.0.conv_1.bias
-vq_model.enc_p.encoder_text.ffn_layers.0.conv_2.weight
-vq_model.enc_p.encoder_text.ffn_layers.0.conv_2.bias
-vq_model.enc_p.encoder_text.ffn_layers.1.conv_1.weight
-vq_model.enc_p.encoder_text.ffn_layers.1.conv_1.bias
-vq_model.enc_p.encoder_text.ffn_layers.1.conv_2.weight
-vq_model.enc_p.encoder_text.ffn_layers.1.conv_2.bias
-vq_model.enc_p.encoder_text.ffn_layers.2.conv_1.weight
-vq_model.enc_p.encoder_text.ffn_layers.2.conv_1.bias
-vq_model.enc_p.encoder_text.ffn_layers.2.conv_2.weight
-vq_model.enc_p.encoder_text.ffn_layers.2.conv_2.bias
-vq_model.enc_p.encoder_text.ffn_layers.3.conv_1.weight
-vq_model.enc_p.encoder_text.ffn_layers.3.conv_1.bias
-vq_model.enc_p.encoder_text.ffn_layers.3.conv_2.weight
-vq_model.enc_p.encoder_text.ffn_layers.3.conv_2.bias
-vq_model.enc_p.encoder_text.ffn_layers.4.conv_1.weight
-vq_model.enc_p.encoder_text.ffn_layers.4.conv_1.bias
-vq_model.enc_p.encoder_text.ffn_layers.4.conv_2.weight
-vq_model.enc_p.encoder_text.ffn_layers.4.conv_2.bias
-vq_model.enc_p.encoder_text.ffn_layers.5.conv_1.weight
-vq_model.enc_p.encoder_text.ffn_layers.5.conv_1.bias
-vq_model.enc_p.encoder_text.ffn_layers.5.conv_2.weight
-vq_model.enc_p.encoder_text.ffn_layers.5.conv_2.bias
-vq_model.enc_p.encoder_text.norm_layers_2.0.gamma
-vq_model.enc_p.encoder_text.norm_layers_2.0.beta
-vq_model.enc_p.encoder_text.norm_layers_2.1.gamma
-vq_model.enc_p.encoder_text.norm_layers_2.1.beta
-vq_model.enc_p.encoder_text.norm_layers_2.2.gamma
-vq_model.enc_p.encoder_text.norm_layers_2.2.beta
-vq_model.enc_p.encoder_text.norm_layers_2.3.gamma
-vq_model.enc_p.encoder_text.norm_layers_2.3.beta
-vq_model.enc_p.encoder_text.norm_layers_2.4.gamma
-vq_model.enc_p.encoder_text.norm_layers_2.4.beta
-vq_model.enc_p.encoder_text.norm_layers_2.5.gamma
-vq_model.enc_p.encoder_text.norm_layers_2.5.beta
-vq_model.enc_p.text_embedding.weight
-vq_model.enc_p.mrte.cross_attention.conv_q.weight
-vq_model.enc_p.mrte.cross_attention.conv_q.bias
-vq_model.enc_p.mrte.cross_attention.conv_k.weight
-vq_model.enc_p.mrte.cross_attention.conv_k.bias
-vq_model.enc_p.mrte.cross_attention.conv_v.weight
-vq_model.enc_p.mrte.cross_attention.conv_v.bias
-vq_model.enc_p.mrte.cross_attention.conv_o.weight
-vq_model.enc_p.mrte.cross_attention.conv_o.bias
-vq_model.enc_p.mrte.c_pre.weight
-vq_model.enc_p.mrte.c_pre.bias
-vq_model.enc_p.mrte.text_pre.weight
-vq_model.enc_p.mrte.text_pre.bias
-vq_model.enc_p.mrte.c_post.weight
-vq_model.enc_p.mrte.c_post.bias
-vq_model.enc_p.encoder2.attn_layers.0.emb_rel_k
-vq_model.enc_p.encoder2.attn_layers.0.emb_rel_v
-vq_model.enc_p.encoder2.attn_layers.0.conv_q.weight
-vq_model.enc_p.encoder2.attn_layers.0.conv_q.bias
-vq_model.enc_p.encoder2.attn_layers.0.conv_k.weight
-vq_model.enc_p.encoder2.attn_layers.0.conv_k.bias
-vq_model.enc_p.encoder2.attn_layers.0.conv_v.weight
-vq_model.enc_p.encoder2.attn_layers.0.conv_v.bias
-vq_model.enc_p.encoder2.attn_layers.0.conv_o.weight
-vq_model.enc_p.encoder2.attn_layers.0.conv_o.bias
-vq_model.enc_p.encoder2.attn_layers.1.emb_rel_k
-vq_model.enc_p.encoder2.attn_layers.1.emb_rel_v
-vq_model.enc_p.encoder2.attn_layers.1.conv_q.weight
-vq_model.enc_p.encoder2.attn_layers.1.conv_q.bias
-vq_model.enc_p.encoder2.attn_layers.1.conv_k.weight
-vq_model.enc_p.encoder2.attn_layers.1.conv_k.bias
-vq_model.enc_p.encoder2.attn_layers.1.conv_v.weight
-vq_model.enc_p.encoder2.attn_layers.1.conv_v.bias
-vq_model.enc_p.encoder2.attn_layers.1.conv_o.weight
-vq_model.enc_p.encoder2.attn_layers.1.conv_o.bias
-vq_model.enc_p.encoder2.attn_layers.2.emb_rel_k
-vq_model.enc_p.encoder2.attn_layers.2.emb_rel_v
-vq_model.enc_p.encoder2.attn_layers.2.conv_q.weight
-vq_model.enc_p.encoder2.attn_layers.2.conv_q.bias
-vq_model.enc_p.encoder2.attn_layers.2.conv_k.weight
-vq_model.enc_p.encoder2.attn_layers.2.conv_k.bias
-vq_model.enc_p.encoder2.attn_layers.2.conv_v.weight
-vq_model.enc_p.encoder2.attn_layers.2.conv_v.bias
-vq_model.enc_p.encoder2.attn_layers.2.conv_o.weight
-vq_model.enc_p.encoder2.attn_layers.2.conv_o.bias
-vq_model.enc_p.encoder2.norm_layers_1.0.gamma
-vq_model.enc_p.encoder2.norm_layers_1.0.beta
-vq_model.enc_p.encoder2.norm_layers_1.1.gamma
-vq_model.enc_p.encoder2.norm_layers_1.1.beta
-vq_model.enc_p.encoder2.norm_layers_1.2.gamma
-vq_model.enc_p.encoder2.norm_layers_1.2.beta
-vq_model.enc_p.encoder2.ffn_layers.0.conv_1.weight
-vq_model.enc_p.encoder2.ffn_layers.0.conv_1.bias
-vq_model.enc_p.encoder2.ffn_layers.0.conv_2.weight
-vq_model.enc_p.encoder2.ffn_layers.0.conv_2.bias
-vq_model.enc_p.encoder2.ffn_layers.1.conv_1.weight
-vq_model.enc_p.encoder2.ffn_layers.1.conv_1.bias
-vq_model.enc_p.encoder2.ffn_layers.1.conv_2.weight
-vq_model.enc_p.encoder2.ffn_layers.1.conv_2.bias
-vq_model.enc_p.encoder2.ffn_layers.2.conv_1.weight
-vq_model.enc_p.encoder2.ffn_layers.2.conv_1.bias
-vq_model.enc_p.encoder2.ffn_layers.2.conv_2.weight
-vq_model.enc_p.encoder2.ffn_layers.2.conv_2.bias
-vq_model.enc_p.encoder2.norm_layers_2.0.gamma
-vq_model.enc_p.encoder2.norm_layers_2.0.beta
-vq_model.enc_p.encoder2.norm_layers_2.1.gamma
-vq_model.enc_p.encoder2.norm_layers_2.1.beta
-vq_model.enc_p.encoder2.norm_layers_2.2.gamma
-vq_model.enc_p.encoder2.norm_layers_2.2.beta
-vq_model.enc_p.proj.weight
-vq_model.enc_p.proj.bias
-vq_model.dec.conv_pre.weight
-vq_model.dec.conv_pre.bias
-vq_model.dec.ups.0.bias
-vq_model.dec.ups.0.weight_g
-vq_model.dec.ups.0.weight_v
-vq_model.dec.ups.1.bias
-vq_model.dec.ups.1.weight_g
-vq_model.dec.ups.1.weight_v
-vq_model.dec.ups.2.bias
-vq_model.dec.ups.2.weight_g
-vq_model.dec.ups.2.weight_v
-vq_model.dec.ups.3.bias
-vq_model.dec.ups.3.weight_g
-vq_model.dec.ups.3.weight_v
-vq_model.dec.ups.4.bias
-vq_model.dec.ups.4.weight_g
-vq_model.dec.ups.4.weight_v
-vq_model.dec.resblocks.0.convs1.0.bias
-vq_model.dec.resblocks.0.convs1.0.weight_g
-vq_model.dec.resblocks.0.convs1.0.weight_v
-vq_model.dec.resblocks.0.convs1.1.bias
-vq_model.dec.resblocks.0.convs1.1.weight_g
-vq_model.dec.resblocks.0.convs1.1.weight_v
-vq_model.dec.resblocks.0.convs1.2.bias
-vq_model.dec.resblocks.0.convs1.2.weight_g
-vq_model.dec.resblocks.0.convs1.2.weight_v
-vq_model.dec.resblocks.0.convs2.0.bias
-vq_model.dec.resblocks.0.convs2.0.weight_g
-vq_model.dec.resblocks.0.convs2.0.weight_v
-vq_model.dec.resblocks.0.convs2.1.bias
-vq_model.dec.resblocks.0.convs2.1.weight_g
-vq_model.dec.resblocks.0.convs2.1.weight_v
-vq_model.dec.resblocks.0.convs2.2.bias
-vq_model.dec.resblocks.0.convs2.2.weight_g
-vq_model.dec.resblocks.0.convs2.2.weight_v
-vq_model.dec.resblocks.1.convs1.0.bias
-vq_model.dec.resblocks.1.convs1.0.weight_g
-vq_model.dec.resblocks.1.convs1.0.weight_v
-vq_model.dec.resblocks.1.convs1.1.bias
-vq_model.dec.resblocks.1.convs1.1.weight_g
-vq_model.dec.resblocks.1.convs1.1.weight_v
-vq_model.dec.resblocks.1.convs1.2.bias
-vq_model.dec.resblocks.1.convs1.2.weight_g
-vq_model.dec.resblocks.1.convs1.2.weight_v
-vq_model.dec.resblocks.1.convs2.0.bias
-vq_model.dec.resblocks.1.convs2.0.weight_g
-vq_model.dec.resblocks.1.convs2.0.weight_v
-vq_model.dec.resblocks.1.convs2.1.bias
-vq_model.dec.resblocks.1.convs2.1.weight_g
-vq_model.dec.resblocks.1.convs2.1.weight_v
-vq_model.dec.resblocks.1.convs2.2.bias
-vq_model.dec.resblocks.1.convs2.2.weight_g
-vq_model.dec.resblocks.1.convs2.2.weight_v
-vq_model.dec.resblocks.2.convs1.0.bias
-vq_model.dec.resblocks.2.convs1.0.weight_g
-vq_model.dec.resblocks.2.convs1.0.weight_v
-vq_model.dec.resblocks.2.convs1.1.bias
-vq_model.dec.resblocks.2.convs1.1.weight_g
-vq_model.dec.resblocks.2.convs1.1.weight_v
-vq_model.dec.resblocks.2.convs1.2.bias
-vq_model.dec.resblocks.2.convs1.2.weight_g
-vq_model.dec.resblocks.2.convs1.2.weight_v
-vq_model.dec.resblocks.2.convs2.0.bias
-vq_model.dec.resblocks.2.convs2.0.weight_g
-vq_model.dec.resblocks.2.convs2.0.weight_v
-vq_model.dec.resblocks.2.convs2.1.bias
-vq_model.dec.resblocks.2.convs2.1.weight_g
-vq_model.dec.resblocks.2.convs2.1.weight_v
-vq_model.dec.resblocks.2.convs2.2.bias
-vq_model.dec.resblocks.2.convs2.2.weight_g
-vq_model.dec.resblocks.2.convs2.2.weight_v
-vq_model.dec.resblocks.3.convs1.0.bias
-vq_model.dec.resblocks.3.convs1.0.weight_g
-vq_model.dec.resblocks.3.convs1.0.weight_v
-vq_model.dec.resblocks.3.convs1.1.bias
-vq_model.dec.resblocks.3.convs1.1.weight_g
-vq_model.dec.resblocks.3.convs1.1.weight_v
-vq_model.dec.resblocks.3.convs1.2.bias
-vq_model.dec.resblocks.3.convs1.2.weight_g
-vq_model.dec.resblocks.3.convs1.2.weight_v
-vq_model.dec.resblocks.3.convs2.0.bias
-vq_model.dec.resblocks.3.convs2.0.weight_g
-vq_model.dec.resblocks.3.convs2.0.weight_v
-vq_model.dec.resblocks.3.convs2.1.bias
-vq_model.dec.resblocks.3.convs2.1.weight_g
-vq_model.dec.resblocks.3.convs2.1.weight_v
-vq_model.dec.resblocks.3.convs2.2.bias
-vq_model.dec.resblocks.3.convs2.2.weight_g
-vq_model.dec.resblocks.3.convs2.2.weight_v
-vq_model.dec.resblocks.4.convs1.0.bias
-vq_model.dec.resblocks.4.convs1.0.weight_g
-vq_model.dec.resblocks.4.convs1.0.weight_v
-vq_model.dec.resblocks.4.convs1.1.bias
-vq_model.dec.resblocks.4.convs1.1.weight_g
-vq_model.dec.resblocks.4.convs1.1.weight_v
-vq_model.dec.resblocks.4.convs1.2.bias
-vq_model.dec.resblocks.4.convs1.2.weight_g
-vq_model.dec.resblocks.4.convs1.2.weight_v
-vq_model.dec.resblocks.4.convs2.0.bias
-vq_model.dec.resblocks.4.convs2.0.weight_g
-vq_model.dec.resblocks.4.convs2.0.weight_v
-vq_model.dec.resblocks.4.convs2.1.bias
-vq_model.dec.resblocks.4.convs2.1.weight_g
-vq_model.dec.resblocks.4.convs2.1.weight_v
-vq_model.dec.resblocks.4.convs2.2.bias
-vq_model.dec.resblocks.4.convs2.2.weight_g
-vq_model.dec.resblocks.4.convs2.2.weight_v
-vq_model.dec.resblocks.5.convs1.0.bias
-vq_model.dec.resblocks.5.convs1.0.weight_g
-vq_model.dec.resblocks.5.convs1.0.weight_v
-vq_model.dec.resblocks.5.convs1.1.bias
-vq_model.dec.resblocks.5.convs1.1.weight_g
-vq_model.dec.resblocks.5.convs1.1.weight_v
-vq_model.dec.resblocks.5.convs1.2.bias
-vq_model.dec.resblocks.5.convs1.2.weight_g
-vq_model.dec.resblocks.5.convs1.2.weight_v
-vq_model.dec.resblocks.5.convs2.0.bias
-vq_model.dec.resblocks.5.convs2.0.weight_g
-vq_model.dec.resblocks.5.convs2.0.weight_v
-vq_model.dec.resblocks.5.convs2.1.bias
-vq_model.dec.resblocks.5.convs2.1.weight_g
-vq_model.dec.resblocks.5.convs2.1.weight_v
-vq_model.dec.resblocks.5.convs2.2.bias
-vq_model.dec.resblocks.5.convs2.2.weight_g
-vq_model.dec.resblocks.5.convs2.2.weight_v
-vq_model.dec.resblocks.6.convs1.0.bias
-vq_model.dec.resblocks.6.convs1.0.weight_g
-vq_model.dec.resblocks.6.convs1.0.weight_v
-vq_model.dec.resblocks.6.convs1.1.bias
-vq_model.dec.resblocks.6.convs1.1.weight_g
-vq_model.dec.resblocks.6.convs1.1.weight_v
-vq_model.dec.resblocks.6.convs1.2.bias
-vq_model.dec.resblocks.6.convs1.2.weight_g
-vq_model.dec.resblocks.6.convs1.2.weight_v
-vq_model.dec.resblocks.6.convs2.0.bias
-vq_model.dec.resblocks.6.convs2.0.weight_g
-vq_model.dec.resblocks.6.convs2.0.weight_v
-vq_model.dec.resblocks.6.convs2.1.bias
-vq_model.dec.resblocks.6.convs2.1.weight_g
-vq_model.dec.resblocks.6.convs2.1.weight_v
-vq_model.dec.resblocks.6.convs2.2.bias
-vq_model.dec.resblocks.6.convs2.2.weight_g
-vq_model.dec.resblocks.6.convs2.2.weight_v
-vq_model.dec.resblocks.7.convs1.0.bias
-vq_model.dec.resblocks.7.convs1.0.weight_g
-vq_model.dec.resblocks.7.convs1.0.weight_v
-vq_model.dec.resblocks.7.convs1.1.bias
-vq_model.dec.resblocks.7.convs1.1.weight_g
-vq_model.dec.resblocks.7.convs1.1.weight_v
-vq_model.dec.resblocks.7.convs1.2.bias
-vq_model.dec.resblocks.7.convs1.2.weight_g
-vq_model.dec.resblocks.7.convs1.2.weight_v
-vq_model.dec.resblocks.7.convs2.0.bias
-vq_model.dec.resblocks.7.convs2.0.weight_g
-vq_model.dec.resblocks.7.convs2.0.weight_v
-vq_model.dec.resblocks.7.convs2.1.bias
-vq_model.dec.resblocks.7.convs2.1.weight_g
-vq_model.dec.resblocks.7.convs2.1.weight_v
-vq_model.dec.resblocks.7.convs2.2.bias
-vq_model.dec.resblocks.7.convs2.2.weight_g
-vq_model.dec.resblocks.7.convs2.2.weight_v
-vq_model.dec.resblocks.8.convs1.0.bias
-vq_model.dec.resblocks.8.convs1.0.weight_g
-vq_model.dec.resblocks.8.convs1.0.weight_v
-vq_model.dec.resblocks.8.convs1.1.bias
-vq_model.dec.resblocks.8.convs1.1.weight_g
-vq_model.dec.resblocks.8.convs1.1.weight_v
-vq_model.dec.resblocks.8.convs1.2.bias
-vq_model.dec.resblocks.8.convs1.2.weight_g
-vq_model.dec.resblocks.8.convs1.2.weight_v
-vq_model.dec.resblocks.8.convs2.0.bias
-vq_model.dec.resblocks.8.convs2.0.weight_g
-vq_model.dec.resblocks.8.convs2.0.weight_v
-vq_model.dec.resblocks.8.convs2.1.bias
-vq_model.dec.resblocks.8.convs2.1.weight_g
-vq_model.dec.resblocks.8.convs2.1.weight_v
-vq_model.dec.resblocks.8.convs2.2.bias
-vq_model.dec.resblocks.8.convs2.2.weight_g
-vq_model.dec.resblocks.8.convs2.2.weight_v
-vq_model.dec.resblocks.9.convs1.0.bias
-vq_model.dec.resblocks.9.convs1.0.weight_g
-vq_model.dec.resblocks.9.convs1.0.weight_v
-vq_model.dec.resblocks.9.convs1.1.bias
-vq_model.dec.resblocks.9.convs1.1.weight_g
-vq_model.dec.resblocks.9.convs1.1.weight_v
-vq_model.dec.resblocks.9.convs1.2.bias
-vq_model.dec.resblocks.9.convs1.2.weight_g
-vq_model.dec.resblocks.9.convs1.2.weight_v
-vq_model.dec.resblocks.9.convs2.0.bias
-vq_model.dec.resblocks.9.convs2.0.weight_g
-vq_model.dec.resblocks.9.convs2.0.weight_v
-vq_model.dec.resblocks.9.convs2.1.bias
-vq_model.dec.resblocks.9.convs2.1.weight_g
-vq_model.dec.resblocks.9.convs2.1.weight_v
-vq_model.dec.resblocks.9.convs2.2.bias
-vq_model.dec.resblocks.9.convs2.2.weight_g
-vq_model.dec.resblocks.9.convs2.2.weight_v
-vq_model.dec.resblocks.10.convs1.0.bias
-vq_model.dec.resblocks.10.convs1.0.weight_g
-vq_model.dec.resblocks.10.convs1.0.weight_v
-vq_model.dec.resblocks.10.convs1.1.bias
-vq_model.dec.resblocks.10.convs1.1.weight_g
-vq_model.dec.resblocks.10.convs1.1.weight_v
-vq_model.dec.resblocks.10.convs1.2.bias
-vq_model.dec.resblocks.10.convs1.2.weight_g
-vq_model.dec.resblocks.10.convs1.2.weight_v
-vq_model.dec.resblocks.10.convs2.0.bias
-vq_model.dec.resblocks.10.convs2.0.weight_g
-vq_model.dec.resblocks.10.convs2.0.weight_v
-vq_model.dec.resblocks.10.convs2.1.bias
-vq_model.dec.resblocks.10.convs2.1.weight_g
-vq_model.dec.resblocks.10.convs2.1.weight_v
-vq_model.dec.resblocks.10.convs2.2.bias
-vq_model.dec.resblocks.10.convs2.2.weight_g
-vq_model.dec.resblocks.10.convs2.2.weight_v
-vq_model.dec.resblocks.11.convs1.0.bias
-vq_model.dec.resblocks.11.convs1.0.weight_g
-vq_model.dec.resblocks.11.convs1.0.weight_v
-vq_model.dec.resblocks.11.convs1.1.bias
-vq_model.dec.resblocks.11.convs1.1.weight_g
-vq_model.dec.resblocks.11.convs1.1.weight_v
-vq_model.dec.resblocks.11.convs1.2.bias
-vq_model.dec.resblocks.11.convs1.2.weight_g
-vq_model.dec.resblocks.11.convs1.2.weight_v
-vq_model.dec.resblocks.11.convs2.0.bias
-vq_model.dec.resblocks.11.convs2.0.weight_g
-vq_model.dec.resblocks.11.convs2.0.weight_v
-vq_model.dec.resblocks.11.convs2.1.bias
-vq_model.dec.resblocks.11.convs2.1.weight_g
-vq_model.dec.resblocks.11.convs2.1.weight_v
-vq_model.dec.resblocks.11.convs2.2.bias
-vq_model.dec.resblocks.11.convs2.2.weight_g
-vq_model.dec.resblocks.11.convs2.2.weight_v
-vq_model.dec.resblocks.12.convs1.0.bias
-vq_model.dec.resblocks.12.convs1.0.weight_g
-vq_model.dec.resblocks.12.convs1.0.weight_v
-vq_model.dec.resblocks.12.convs1.1.bias
-vq_model.dec.resblocks.12.convs1.1.weight_g
-vq_model.dec.resblocks.12.convs1.1.weight_v
-vq_model.dec.resblocks.12.convs1.2.bias
-vq_model.dec.resblocks.12.convs1.2.weight_g
-vq_model.dec.resblocks.12.convs1.2.weight_v
-vq_model.dec.resblocks.12.convs2.0.bias
-vq_model.dec.resblocks.12.convs2.0.weight_g
-vq_model.dec.resblocks.12.convs2.0.weight_v
-vq_model.dec.resblocks.12.convs2.1.bias
-vq_model.dec.resblocks.12.convs2.1.weight_g
-vq_model.dec.resblocks.12.convs2.1.weight_v
-vq_model.dec.resblocks.12.convs2.2.bias
-vq_model.dec.resblocks.12.convs2.2.weight_g
-vq_model.dec.resblocks.12.convs2.2.weight_v
-vq_model.dec.resblocks.13.convs1.0.bias
-vq_model.dec.resblocks.13.convs1.0.weight_g
-vq_model.dec.resblocks.13.convs1.0.weight_v
-vq_model.dec.resblocks.13.convs1.1.bias
-vq_model.dec.resblocks.13.convs1.1.weight_g
-vq_model.dec.resblocks.13.convs1.1.weight_v
-vq_model.dec.resblocks.13.convs1.2.bias
-vq_model.dec.resblocks.13.convs1.2.weight_g
-vq_model.dec.resblocks.13.convs1.2.weight_v
-vq_model.dec.resblocks.13.convs2.0.bias
-vq_model.dec.resblocks.13.convs2.0.weight_g
-vq_model.dec.resblocks.13.convs2.0.weight_v
-vq_model.dec.resblocks.13.convs2.1.bias
-vq_model.dec.resblocks.13.convs2.1.weight_g
-vq_model.dec.resblocks.13.convs2.1.weight_v
-vq_model.dec.resblocks.13.convs2.2.bias
-vq_model.dec.resblocks.13.convs2.2.weight_g
-vq_model.dec.resblocks.13.convs2.2.weight_v
-vq_model.dec.resblocks.14.convs1.0.bias
-vq_model.dec.resblocks.14.convs1.0.weight_g
-vq_model.dec.resblocks.14.convs1.0.weight_v
-vq_model.dec.resblocks.14.convs1.1.bias
-vq_model.dec.resblocks.14.convs1.1.weight_g
-vq_model.dec.resblocks.14.convs1.1.weight_v
-vq_model.dec.resblocks.14.convs1.2.bias
-vq_model.dec.resblocks.14.convs1.2.weight_g
-vq_model.dec.resblocks.14.convs1.2.weight_v
-vq_model.dec.resblocks.14.convs2.0.bias
-vq_model.dec.resblocks.14.convs2.0.weight_g
-vq_model.dec.resblocks.14.convs2.0.weight_v
-vq_model.dec.resblocks.14.convs2.1.bias
-vq_model.dec.resblocks.14.convs2.1.weight_g
-vq_model.dec.resblocks.14.convs2.1.weight_v
-vq_model.dec.resblocks.14.convs2.2.bias
-vq_model.dec.resblocks.14.convs2.2.weight_g
-vq_model.dec.resblocks.14.convs2.2.weight_v
-vq_model.dec.conv_post.weight
-vq_model.dec.cond.weight
-vq_model.dec.cond.bias
-vq_model.flow.flows.0.pre.weight
-vq_model.flow.flows.0.pre.bias
-vq_model.flow.flows.0.enc.in_layers.0.bias
-vq_model.flow.flows.0.enc.in_layers.0.weight_g
-vq_model.flow.flows.0.enc.in_layers.0.weight_v
-vq_model.flow.flows.0.enc.in_layers.1.bias
-vq_model.flow.flows.0.enc.in_layers.1.weight_g
-vq_model.flow.flows.0.enc.in_layers.1.weight_v
-vq_model.flow.flows.0.enc.in_layers.2.bias
-vq_model.flow.flows.0.enc.in_layers.2.weight_g
-vq_model.flow.flows.0.enc.in_layers.2.weight_v
-vq_model.flow.flows.0.enc.in_layers.3.bias
-vq_model.flow.flows.0.enc.in_layers.3.weight_g
-vq_model.flow.flows.0.enc.in_layers.3.weight_v
-vq_model.flow.flows.0.enc.res_skip_layers.0.bias
-vq_model.flow.flows.0.enc.res_skip_layers.0.weight_g
-vq_model.flow.flows.0.enc.res_skip_layers.0.weight_v
-vq_model.flow.flows.0.enc.res_skip_layers.1.bias
-vq_model.flow.flows.0.enc.res_skip_layers.1.weight_g
-vq_model.flow.flows.0.enc.res_skip_layers.1.weight_v
-vq_model.flow.flows.0.enc.res_skip_layers.2.bias
-vq_model.flow.flows.0.enc.res_skip_layers.2.weight_g
-vq_model.flow.flows.0.enc.res_skip_layers.2.weight_v
-vq_model.flow.flows.0.enc.res_skip_layers.3.bias
-vq_model.flow.flows.0.enc.res_skip_layers.3.weight_g
-vq_model.flow.flows.0.enc.res_skip_layers.3.weight_v
-vq_model.flow.flows.0.enc.cond_layer.bias
-vq_model.flow.flows.0.enc.cond_layer.weight_g
-vq_model.flow.flows.0.enc.cond_layer.weight_v
-vq_model.flow.flows.0.post.weight
-vq_model.flow.flows.0.post.bias
-vq_model.flow.flows.2.pre.weight
-vq_model.flow.flows.2.pre.bias
-vq_model.flow.flows.2.enc.in_layers.0.bias
-vq_model.flow.flows.2.enc.in_layers.0.weight_g
-vq_model.flow.flows.2.enc.in_layers.0.weight_v
-vq_model.flow.flows.2.enc.in_layers.1.bias
-vq_model.flow.flows.2.enc.in_layers.1.weight_g
-vq_model.flow.flows.2.enc.in_layers.1.weight_v
-vq_model.flow.flows.2.enc.in_layers.2.bias
-vq_model.flow.flows.2.enc.in_layers.2.weight_g
-vq_model.flow.flows.2.enc.in_layers.2.weight_v
-vq_model.flow.flows.2.enc.in_layers.3.bias
-vq_model.flow.flows.2.enc.in_layers.3.weight_g
-vq_model.flow.flows.2.enc.in_layers.3.weight_v
-vq_model.flow.flows.2.enc.res_skip_layers.0.bias
-vq_model.flow.flows.2.enc.res_skip_layers.0.weight_g
-vq_model.flow.flows.2.enc.res_skip_layers.0.weight_v
-vq_model.flow.flows.2.enc.res_skip_layers.1.bias
-vq_model.flow.flows.2.enc.res_skip_layers.1.weight_g
-vq_model.flow.flows.2.enc.res_skip_layers.1.weight_v
-vq_model.flow.flows.2.enc.res_skip_layers.2.bias
-vq_model.flow.flows.2.enc.res_skip_layers.2.weight_g
-vq_model.flow.flows.2.enc.res_skip_layers.2.weight_v
-vq_model.flow.flows.2.enc.res_skip_layers.3.bias
-vq_model.flow.flows.2.enc.res_skip_layers.3.weight_g
-vq_model.flow.flows.2.enc.res_skip_layers.3.weight_v
-vq_model.flow.flows.2.enc.cond_layer.bias
-vq_model.flow.flows.2.enc.cond_layer.weight_g
-vq_model.flow.flows.2.enc.cond_layer.weight_v
-vq_model.flow.flows.2.post.weight
-vq_model.flow.flows.2.post.bias
-vq_model.flow.flows.4.pre.weight
-vq_model.flow.flows.4.pre.bias
-vq_model.flow.flows.4.enc.in_layers.0.bias
-vq_model.flow.flows.4.enc.in_layers.0.weight_g
-vq_model.flow.flows.4.enc.in_layers.0.weight_v
-vq_model.flow.flows.4.enc.in_layers.1.bias
-vq_model.flow.flows.4.enc.in_layers.1.weight_g
-vq_model.flow.flows.4.enc.in_layers.1.weight_v
-vq_model.flow.flows.4.enc.in_layers.2.bias
-vq_model.flow.flows.4.enc.in_layers.2.weight_g
-vq_model.flow.flows.4.enc.in_layers.2.weight_v
-vq_model.flow.flows.4.enc.in_layers.3.bias
-vq_model.flow.flows.4.enc.in_layers.3.weight_g
-vq_model.flow.flows.4.enc.in_layers.3.weight_v
-vq_model.flow.flows.4.enc.res_skip_layers.0.bias
-vq_model.flow.flows.4.enc.res_skip_layers.0.weight_g
-vq_model.flow.flows.4.enc.res_skip_layers.0.weight_v
-vq_model.flow.flows.4.enc.res_skip_layers.1.bias
-vq_model.flow.flows.4.enc.res_skip_layers.1.weight_g
-vq_model.flow.flows.4.enc.res_skip_layers.1.weight_v
-vq_model.flow.flows.4.enc.res_skip_layers.2.bias
-vq_model.flow.flows.4.enc.res_skip_layers.2.weight_g
-vq_model.flow.flows.4.enc.res_skip_layers.2.weight_v
-vq_model.flow.flows.4.enc.res_skip_layers.3.bias
-vq_model.flow.flows.4.enc.res_skip_layers.3.weight_g
-vq_model.flow.flows.4.enc.res_skip_layers.3.weight_v
-vq_model.flow.flows.4.enc.cond_layer.bias
-vq_model.flow.flows.4.enc.cond_layer.weight_g
-vq_model.flow.flows.4.enc.cond_layer.weight_v
-vq_model.flow.flows.4.post.weight
-vq_model.flow.flows.4.post.bias
-vq_model.flow.flows.6.pre.weight
-vq_model.flow.flows.6.pre.bias
-vq_model.flow.flows.6.enc.in_layers.0.bias
-vq_model.flow.flows.6.enc.in_layers.0.weight_g
-vq_model.flow.flows.6.enc.in_layers.0.weight_v
-vq_model.flow.flows.6.enc.in_layers.1.bias
-vq_model.flow.flows.6.enc.in_layers.1.weight_g
-vq_model.flow.flows.6.enc.in_layers.1.weight_v
-vq_model.flow.flows.6.enc.in_layers.2.bias
-vq_model.flow.flows.6.enc.in_layers.2.weight_g
-vq_model.flow.flows.6.enc.in_layers.2.weight_v
-vq_model.flow.flows.6.enc.in_layers.3.bias
-vq_model.flow.flows.6.enc.in_layers.3.weight_g
-vq_model.flow.flows.6.enc.in_layers.3.weight_v
-vq_model.flow.flows.6.enc.res_skip_layers.0.bias
-vq_model.flow.flows.6.enc.res_skip_layers.0.weight_g
-vq_model.flow.flows.6.enc.res_skip_layers.0.weight_v
-vq_model.flow.flows.6.enc.res_skip_layers.1.bias
-vq_model.flow.flows.6.enc.res_skip_layers.1.weight_g
-vq_model.flow.flows.6.enc.res_skip_layers.1.weight_v
-vq_model.flow.flows.6.enc.res_skip_layers.2.bias
-vq_model.flow.flows.6.enc.res_skip_layers.2.weight_g
-vq_model.flow.flows.6.enc.res_skip_layers.2.weight_v
-vq_model.flow.flows.6.enc.res_skip_layers.3.bias
-vq_model.flow.flows.6.enc.res_skip_layers.3.weight_g
-vq_model.flow.flows.6.enc.res_skip_layers.3.weight_v
-vq_model.flow.flows.6.enc.cond_layer.bias
-vq_model.flow.flows.6.enc.cond_layer.weight_g
-vq_model.flow.flows.6.enc.cond_layer.weight_v
-vq_model.flow.flows.6.post.weight
-vq_model.flow.flows.6.post.bias
-vq_model.quantizer.vq.layers.0._codebook.embed
+vq_model.enc_p.ssl_proj.weight
+vq_model.enc_p.ssl_proj.bias
+vq_model.enc_p.encoder_ssl.attn_layers.0.emb_rel_k
+vq_model.enc_p.encoder_ssl.attn_layers.0.emb_rel_v
+vq_model.enc_p.encoder_ssl.attn_layers.0.conv_q.weight
+vq_model.enc_p.encoder_ssl.attn_layers.0.conv_q.bias
+vq_model.enc_p.encoder_ssl.attn_layers.0.conv_k.weight
+vq_model.enc_p.encoder_ssl.attn_layers.0.conv_k.bias
+vq_model.enc_p.encoder_ssl.attn_layers.0.conv_v.weight
+vq_model.enc_p.encoder_ssl.attn_layers.0.conv_v.bias
+vq_model.enc_p.encoder_ssl.attn_layers.0.conv_o.weight
+vq_model.enc_p.encoder_ssl.attn_layers.0.conv_o.bias
+vq_model.enc_p.encoder_ssl.attn_layers.1.emb_rel_k
+vq_model.enc_p.encoder_ssl.attn_layers.1.emb_rel_v
+vq_model.enc_p.encoder_ssl.attn_layers.1.conv_q.weight
+vq_model.enc_p.encoder_ssl.attn_layers.1.conv_q.bias
+vq_model.enc_p.encoder_ssl.attn_layers.1.conv_k.weight
+vq_model.enc_p.encoder_ssl.attn_layers.1.conv_k.bias
+vq_model.enc_p.encoder_ssl.attn_layers.1.conv_v.weight
+vq_model.enc_p.encoder_ssl.attn_layers.1.conv_v.bias
+vq_model.enc_p.encoder_ssl.attn_layers.1.conv_o.weight
+vq_model.enc_p.encoder_ssl.attn_layers.1.conv_o.bias
+vq_model.enc_p.encoder_ssl.attn_layers.2.emb_rel_k
+vq_model.enc_p.encoder_ssl.attn_layers.2.emb_rel_v
+vq_model.enc_p.encoder_ssl.attn_layers.2.conv_q.weight
+vq_model.enc_p.encoder_ssl.attn_layers.2.conv_q.bias
+vq_model.enc_p.encoder_ssl.attn_layers.2.conv_k.weight
+vq_model.enc_p.encoder_ssl.attn_layers.2.conv_k.bias
+vq_model.enc_p.encoder_ssl.attn_layers.2.conv_v.weight
+vq_model.enc_p.encoder_ssl.attn_layers.2.conv_v.bias
+vq_model.enc_p.encoder_ssl.attn_layers.2.conv_o.weight
+vq_model.enc_p.encoder_ssl.attn_layers.2.conv_o.bias
+vq_model.enc_p.encoder_ssl.norm_layers_1.0.gamma
+vq_model.enc_p.encoder_ssl.norm_layers_1.0.beta
+vq_model.enc_p.encoder_ssl.norm_layers_1.1.gamma
+vq_model.enc_p.encoder_ssl.norm_layers_1.1.beta
+vq_model.enc_p.encoder_ssl.norm_layers_1.2.gamma
+vq_model.enc_p.encoder_ssl.norm_layers_1.2.beta
+vq_model.enc_p.encoder_ssl.ffn_layers.0.conv_1.weight
+vq_model.enc_p.encoder_ssl.ffn_layers.0.conv_1.bias
+vq_model.enc_p.encoder_ssl.ffn_layers.0.conv_2.weight
+vq_model.enc_p.encoder_ssl.ffn_layers.0.conv_2.bias
+vq_model.enc_p.encoder_ssl.ffn_layers.1.conv_1.weight
+vq_model.enc_p.encoder_ssl.ffn_layers.1.conv_1.bias
+vq_model.enc_p.encoder_ssl.ffn_layers.1.conv_2.weight
+vq_model.enc_p.encoder_ssl.ffn_layers.1.conv_2.bias
+vq_model.enc_p.encoder_ssl.ffn_layers.2.conv_1.weight
+vq_model.enc_p.encoder_ssl.ffn_layers.2.conv_1.bias
+vq_model.enc_p.encoder_ssl.ffn_layers.2.conv_2.weight
+vq_model.enc_p.encoder_ssl.ffn_layers.2.conv_2.bias
+vq_model.enc_p.encoder_ssl.norm_layers_2.0.gamma
+vq_model.enc_p.encoder_ssl.norm_layers_2.0.beta
+vq_model.enc_p.encoder_ssl.norm_layers_2.1.gamma
+vq_model.enc_p.encoder_ssl.norm_layers_2.1.beta
+vq_model.enc_p.encoder_ssl.norm_layers_2.2.gamma
+vq_model.enc_p.encoder_ssl.norm_layers_2.2.beta
+vq_model.enc_p.encoder_text.attn_layers.0.emb_rel_k
+vq_model.enc_p.encoder_text.attn_layers.0.emb_rel_v
+vq_model.enc_p.encoder_text.attn_layers.0.conv_q.weight
+vq_model.enc_p.encoder_text.attn_layers.0.conv_q.bias
+vq_model.enc_p.encoder_text.attn_layers.0.conv_k.weight
+vq_model.enc_p.encoder_text.attn_layers.0.conv_k.bias
+vq_model.enc_p.encoder_text.attn_layers.0.conv_v.weight
+vq_model.enc_p.encoder_text.attn_layers.0.conv_v.bias
+vq_model.enc_p.encoder_text.attn_layers.0.conv_o.weight
+vq_model.enc_p.encoder_text.attn_layers.0.conv_o.bias
+vq_model.enc_p.encoder_text.attn_layers.1.emb_rel_k
+vq_model.enc_p.encoder_text.attn_layers.1.emb_rel_v
+vq_model.enc_p.encoder_text.attn_layers.1.conv_q.weight
+vq_model.enc_p.encoder_text.attn_layers.1.conv_q.bias
+vq_model.enc_p.encoder_text.attn_layers.1.conv_k.weight
+vq_model.enc_p.encoder_text.attn_layers.1.conv_k.bias
+vq_model.enc_p.encoder_text.attn_layers.1.conv_v.weight
+vq_model.enc_p.encoder_text.attn_layers.1.conv_v.bias
+vq_model.enc_p.encoder_text.attn_layers.1.conv_o.weight
+vq_model.enc_p.encoder_text.attn_layers.1.conv_o.bias
+vq_model.enc_p.encoder_text.attn_layers.2.emb_rel_k
+vq_model.enc_p.encoder_text.attn_layers.2.emb_rel_v
+vq_model.enc_p.encoder_text.attn_layers.2.conv_q.weight
+vq_model.enc_p.encoder_text.attn_layers.2.conv_q.bias
+vq_model.enc_p.encoder_text.attn_layers.2.conv_k.weight
+vq_model.enc_p.encoder_text.attn_layers.2.conv_k.bias
+vq_model.enc_p.encoder_text.attn_layers.2.conv_v.weight
+vq_model.enc_p.encoder_text.attn_layers.2.conv_v.bias
+vq_model.enc_p.encoder_text.attn_layers.2.conv_o.weight
+vq_model.enc_p.encoder_text.attn_layers.2.conv_o.bias
+vq_model.enc_p.encoder_text.attn_layers.3.emb_rel_k
+vq_model.enc_p.encoder_text.attn_layers.3.emb_rel_v
+vq_model.enc_p.encoder_text.attn_layers.3.conv_q.weight
+vq_model.enc_p.encoder_text.attn_layers.3.conv_q.bias
+vq_model.enc_p.encoder_text.attn_layers.3.conv_k.weight
+vq_model.enc_p.encoder_text.attn_layers.3.conv_k.bias
+vq_model.enc_p.encoder_text.attn_layers.3.conv_v.weight
+vq_model.enc_p.encoder_text.attn_layers.3.conv_v.bias
+vq_model.enc_p.encoder_text.attn_layers.3.conv_o.weight
+vq_model.enc_p.encoder_text.attn_layers.3.conv_o.bias
+vq_model.enc_p.encoder_text.attn_layers.4.emb_rel_k
+vq_model.enc_p.encoder_text.attn_layers.4.emb_rel_v
+vq_model.enc_p.encoder_text.attn_layers.4.conv_q.weight
+vq_model.enc_p.encoder_text.attn_layers.4.conv_q.bias
+vq_model.enc_p.encoder_text.attn_layers.4.conv_k.weight
+vq_model.enc_p.encoder_text.attn_layers.4.conv_k.bias
+vq_model.enc_p.encoder_text.attn_layers.4.conv_v.weight
+vq_model.enc_p.encoder_text.attn_layers.4.conv_v.bias
+vq_model.enc_p.encoder_text.attn_layers.4.conv_o.weight
+vq_model.enc_p.encoder_text.attn_layers.4.conv_o.bias
+vq_model.enc_p.encoder_text.attn_layers.5.emb_rel_k
+vq_model.enc_p.encoder_text.attn_layers.5.emb_rel_v
+vq_model.enc_p.encoder_text.attn_layers.5.conv_q.weight
+vq_model.enc_p.encoder_text.attn_layers.5.conv_q.bias
+vq_model.enc_p.encoder_text.attn_layers.5.conv_k.weight
+vq_model.enc_p.encoder_text.attn_layers.5.conv_k.bias
+vq_model.enc_p.encoder_text.attn_layers.5.conv_v.weight
+vq_model.enc_p.encoder_text.attn_layers.5.conv_v.bias
+vq_model.enc_p.encoder_text.attn_layers.5.conv_o.weight
+vq_model.enc_p.encoder_text.attn_layers.5.conv_o.bias
+vq_model.enc_p.encoder_text.norm_layers_1.0.gamma
+vq_model.enc_p.encoder_text.norm_layers_1.0.beta
+vq_model.enc_p.encoder_text.norm_layers_1.1.gamma
+vq_model.enc_p.encoder_text.norm_layers_1.1.beta
+vq_model.enc_p.encoder_text.norm_layers_1.2.gamma
+vq_model.enc_p.encoder_text.norm_layers_1.2.beta
+vq_model.enc_p.encoder_text.norm_layers_1.3.gamma
+vq_model.enc_p.encoder_text.norm_layers_1.3.beta
+vq_model.enc_p.encoder_text.norm_layers_1.4.gamma
+vq_model.enc_p.encoder_text.norm_layers_1.4.beta
+vq_model.enc_p.encoder_text.norm_layers_1.5.gamma
+vq_model.enc_p.encoder_text.norm_layers_1.5.beta
+vq_model.enc_p.encoder_text.ffn_layers.0.conv_1.weight
+vq_model.enc_p.encoder_text.ffn_layers.0.conv_1.bias
+vq_model.enc_p.encoder_text.ffn_layers.0.conv_2.weight
+vq_model.enc_p.encoder_text.ffn_layers.0.conv_2.bias
+vq_model.enc_p.encoder_text.ffn_layers.1.conv_1.weight
+vq_model.enc_p.encoder_text.ffn_layers.1.conv_1.bias
+vq_model.enc_p.encoder_text.ffn_layers.1.conv_2.weight
+vq_model.enc_p.encoder_text.ffn_layers.1.conv_2.bias
+vq_model.enc_p.encoder_text.ffn_layers.2.conv_1.weight
+vq_model.enc_p.encoder_text.ffn_layers.2.conv_1.bias
+vq_model.enc_p.encoder_text.ffn_layers.2.conv_2.weight
+vq_model.enc_p.encoder_text.ffn_layers.2.conv_2.bias
+vq_model.enc_p.encoder_text.ffn_layers.3.conv_1.weight
+vq_model.enc_p.encoder_text.ffn_layers.3.conv_1.bias
+vq_model.enc_p.encoder_text.ffn_layers.3.conv_2.weight
+vq_model.enc_p.encoder_text.ffn_layers.3.conv_2.bias
+vq_model.enc_p.encoder_text.ffn_layers.4.conv_1.weight
+vq_model.enc_p.encoder_text.ffn_layers.4.conv_1.bias
+vq_model.enc_p.encoder_text.ffn_layers.4.conv_2.weight
+vq_model.enc_p.encoder_text.ffn_layers.4.conv_2.bias
+vq_model.enc_p.encoder_text.ffn_layers.5.conv_1.weight
+vq_model.enc_p.encoder_text.ffn_layers.5.conv_1.bias
+vq_model.enc_p.encoder_text.ffn_layers.5.conv_2.weight
+vq_model.enc_p.encoder_text.ffn_layers.5.conv_2.bias
+vq_model.enc_p.encoder_text.norm_layers_2.0.gamma
+vq_model.enc_p.encoder_text.norm_layers_2.0.beta
+vq_model.enc_p.encoder_text.norm_layers_2.1.gamma
+vq_model.enc_p.encoder_text.norm_layers_2.1.beta
+vq_model.enc_p.encoder_text.norm_layers_2.2.gamma
+vq_model.enc_p.encoder_text.norm_layers_2.2.beta
+vq_model.enc_p.encoder_text.norm_layers_2.3.gamma
+vq_model.enc_p.encoder_text.norm_layers_2.3.beta
+vq_model.enc_p.encoder_text.norm_layers_2.4.gamma
+vq_model.enc_p.encoder_text.norm_layers_2.4.beta
+vq_model.enc_p.encoder_text.norm_layers_2.5.gamma
+vq_model.enc_p.encoder_text.norm_layers_2.5.beta
+vq_model.enc_p.text_embedding.weight
+vq_model.enc_p.mrte.cross_attention.conv_q.weight
+vq_model.enc_p.mrte.cross_attention.conv_q.bias
+vq_model.enc_p.mrte.cross_attention.conv_k.weight
+vq_model.enc_p.mrte.cross_attention.conv_k.bias
+vq_model.enc_p.mrte.cross_attention.conv_v.weight
+vq_model.enc_p.mrte.cross_attention.conv_v.bias
+vq_model.enc_p.mrte.cross_attention.conv_o.weight
+vq_model.enc_p.mrte.cross_attention.conv_o.bias
+vq_model.enc_p.mrte.c_pre.weight
+vq_model.enc_p.mrte.c_pre.bias
+vq_model.enc_p.mrte.text_pre.weight
+vq_model.enc_p.mrte.text_pre.bias
+vq_model.enc_p.mrte.c_post.weight
+vq_model.enc_p.mrte.c_post.bias
+vq_model.enc_p.encoder2.attn_layers.0.emb_rel_k
+vq_model.enc_p.encoder2.attn_layers.0.emb_rel_v
+vq_model.enc_p.encoder2.attn_layers.0.conv_q.weight
+vq_model.enc_p.encoder2.attn_layers.0.conv_q.bias
+vq_model.enc_p.encoder2.attn_layers.0.conv_k.weight
+vq_model.enc_p.encoder2.attn_layers.0.conv_k.bias
+vq_model.enc_p.encoder2.attn_layers.0.conv_v.weight
+vq_model.enc_p.encoder2.attn_layers.0.conv_v.bias
+vq_model.enc_p.encoder2.attn_layers.0.conv_o.weight
+vq_model.enc_p.encoder2.attn_layers.0.conv_o.bias
+vq_model.enc_p.encoder2.attn_layers.1.emb_rel_k
+vq_model.enc_p.encoder2.attn_layers.1.emb_rel_v
+vq_model.enc_p.encoder2.attn_layers.1.conv_q.weight
+vq_model.enc_p.encoder2.attn_layers.1.conv_q.bias
+vq_model.enc_p.encoder2.attn_layers.1.conv_k.weight
+vq_model.enc_p.encoder2.attn_layers.1.conv_k.bias
+vq_model.enc_p.encoder2.attn_layers.1.conv_v.weight
+vq_model.enc_p.encoder2.attn_layers.1.conv_v.bias
+vq_model.enc_p.encoder2.attn_layers.1.conv_o.weight
+vq_model.enc_p.encoder2.attn_layers.1.conv_o.bias
+vq_model.enc_p.encoder2.attn_layers.2.emb_rel_k
+vq_model.enc_p.encoder2.attn_layers.2.emb_rel_v
+vq_model.enc_p.encoder2.attn_layers.2.conv_q.weight
+vq_model.enc_p.encoder2.attn_layers.2.conv_q.bias
+vq_model.enc_p.encoder2.attn_layers.2.conv_k.weight
+vq_model.enc_p.encoder2.attn_layers.2.conv_k.bias
+vq_model.enc_p.encoder2.attn_layers.2.conv_v.weight
+vq_model.enc_p.encoder2.attn_layers.2.conv_v.bias
+vq_model.enc_p.encoder2.attn_layers.2.conv_o.weight
+vq_model.enc_p.encoder2.attn_layers.2.conv_o.bias
+vq_model.enc_p.encoder2.norm_layers_1.0.gamma
+vq_model.enc_p.encoder2.norm_layers_1.0.beta
+vq_model.enc_p.encoder2.norm_layers_1.1.gamma
+vq_model.enc_p.encoder2.norm_layers_1.1.beta
+vq_model.enc_p.encoder2.norm_layers_1.2.gamma
+vq_model.enc_p.encoder2.norm_layers_1.2.beta
+vq_model.enc_p.encoder2.ffn_layers.0.conv_1.weight
+vq_model.enc_p.encoder2.ffn_layers.0.conv_1.bias
+vq_model.enc_p.encoder2.ffn_layers.0.conv_2.weight
+vq_model.enc_p.encoder2.ffn_layers.0.conv_2.bias
+vq_model.enc_p.encoder2.ffn_layers.1.conv_1.weight
+vq_model.enc_p.encoder2.ffn_layers.1.conv_1.bias
+vq_model.enc_p.encoder2.ffn_layers.1.conv_2.weight
+vq_model.enc_p.encoder2.ffn_layers.1.conv_2.bias
+vq_model.enc_p.encoder2.ffn_layers.2.conv_1.weight
+vq_model.enc_p.encoder2.ffn_layers.2.conv_1.bias
+vq_model.enc_p.encoder2.ffn_layers.2.conv_2.weight
+vq_model.enc_p.encoder2.ffn_layers.2.conv_2.bias
+vq_model.enc_p.encoder2.norm_layers_2.0.gamma
+vq_model.enc_p.encoder2.norm_layers_2.0.beta
+vq_model.enc_p.encoder2.norm_layers_2.1.gamma
+vq_model.enc_p.encoder2.norm_layers_2.1.beta
+vq_model.enc_p.encoder2.norm_layers_2.2.gamma
+vq_model.enc_p.encoder2.norm_layers_2.2.beta
+vq_model.enc_p.proj.weight
+vq_model.enc_p.proj.bias
+vq_model.dec.conv_pre.weight
+vq_model.dec.conv_pre.bias
+vq_model.dec.ups.0.bias
+vq_model.dec.ups.0.weight_g
+vq_model.dec.ups.0.weight_v
+vq_model.dec.ups.1.bias
+vq_model.dec.ups.1.weight_g
+vq_model.dec.ups.1.weight_v
+vq_model.dec.ups.2.bias
+vq_model.dec.ups.2.weight_g
+vq_model.dec.ups.2.weight_v
+vq_model.dec.ups.3.bias
+vq_model.dec.ups.3.weight_g
+vq_model.dec.ups.3.weight_v
+vq_model.dec.ups.4.bias
+vq_model.dec.ups.4.weight_g
+vq_model.dec.ups.4.weight_v
+vq_model.dec.resblocks.0.convs1.0.bias
+vq_model.dec.resblocks.0.convs1.0.weight_g
+vq_model.dec.resblocks.0.convs1.0.weight_v
+vq_model.dec.resblocks.0.convs1.1.bias
+vq_model.dec.resblocks.0.convs1.1.weight_g
+vq_model.dec.resblocks.0.convs1.1.weight_v
+vq_model.dec.resblocks.0.convs1.2.bias
+vq_model.dec.resblocks.0.convs1.2.weight_g
+vq_model.dec.resblocks.0.convs1.2.weight_v
+vq_model.dec.resblocks.0.convs2.0.bias
+vq_model.dec.resblocks.0.convs2.0.weight_g
+vq_model.dec.resblocks.0.convs2.0.weight_v
+vq_model.dec.resblocks.0.convs2.1.bias
+vq_model.dec.resblocks.0.convs2.1.weight_g
+vq_model.dec.resblocks.0.convs2.1.weight_v
+vq_model.dec.resblocks.0.convs2.2.bias
+vq_model.dec.resblocks.0.convs2.2.weight_g
+vq_model.dec.resblocks.0.convs2.2.weight_v
+vq_model.dec.resblocks.1.convs1.0.bias
+vq_model.dec.resblocks.1.convs1.0.weight_g
+vq_model.dec.resblocks.1.convs1.0.weight_v
+vq_model.dec.resblocks.1.convs1.1.bias
+vq_model.dec.resblocks.1.convs1.1.weight_g
+vq_model.dec.resblocks.1.convs1.1.weight_v
+vq_model.dec.resblocks.1.convs1.2.bias
+vq_model.dec.resblocks.1.convs1.2.weight_g
+vq_model.dec.resblocks.1.convs1.2.weight_v
+vq_model.dec.resblocks.1.convs2.0.bias
+vq_model.dec.resblocks.1.convs2.0.weight_g
+vq_model.dec.resblocks.1.convs2.0.weight_v
+vq_model.dec.resblocks.1.convs2.1.bias
+vq_model.dec.resblocks.1.convs2.1.weight_g
+vq_model.dec.resblocks.1.convs2.1.weight_v
+vq_model.dec.resblocks.1.convs2.2.bias
+vq_model.dec.resblocks.1.convs2.2.weight_g
+vq_model.dec.resblocks.1.convs2.2.weight_v
+vq_model.dec.resblocks.2.convs1.0.bias
+vq_model.dec.resblocks.2.convs1.0.weight_g
+vq_model.dec.resblocks.2.convs1.0.weight_v
+vq_model.dec.resblocks.2.convs1.1.bias
+vq_model.dec.resblocks.2.convs1.1.weight_g
+vq_model.dec.resblocks.2.convs1.1.weight_v
+vq_model.dec.resblocks.2.convs1.2.bias
+vq_model.dec.resblocks.2.convs1.2.weight_g
+vq_model.dec.resblocks.2.convs1.2.weight_v
+vq_model.dec.resblocks.2.convs2.0.bias
+vq_model.dec.resblocks.2.convs2.0.weight_g
+vq_model.dec.resblocks.2.convs2.0.weight_v
+vq_model.dec.resblocks.2.convs2.1.bias
+vq_model.dec.resblocks.2.convs2.1.weight_g
+vq_model.dec.resblocks.2.convs2.1.weight_v
+vq_model.dec.resblocks.2.convs2.2.bias
+vq_model.dec.resblocks.2.convs2.2.weight_g
+vq_model.dec.resblocks.2.convs2.2.weight_v
+vq_model.dec.resblocks.3.convs1.0.bias
+vq_model.dec.resblocks.3.convs1.0.weight_g
+vq_model.dec.resblocks.3.convs1.0.weight_v
+vq_model.dec.resblocks.3.convs1.1.bias
+vq_model.dec.resblocks.3.convs1.1.weight_g
+vq_model.dec.resblocks.3.convs1.1.weight_v
+vq_model.dec.resblocks.3.convs1.2.bias
+vq_model.dec.resblocks.3.convs1.2.weight_g
+vq_model.dec.resblocks.3.convs1.2.weight_v
+vq_model.dec.resblocks.3.convs2.0.bias
+vq_model.dec.resblocks.3.convs2.0.weight_g
+vq_model.dec.resblocks.3.convs2.0.weight_v
+vq_model.dec.resblocks.3.convs2.1.bias
+vq_model.dec.resblocks.3.convs2.1.weight_g
+vq_model.dec.resblocks.3.convs2.1.weight_v
+vq_model.dec.resblocks.3.convs2.2.bias
+vq_model.dec.resblocks.3.convs2.2.weight_g
+vq_model.dec.resblocks.3.convs2.2.weight_v
+vq_model.dec.resblocks.4.convs1.0.bias
+vq_model.dec.resblocks.4.convs1.0.weight_g
+vq_model.dec.resblocks.4.convs1.0.weight_v
+vq_model.dec.resblocks.4.convs1.1.bias
+vq_model.dec.resblocks.4.convs1.1.weight_g
+vq_model.dec.resblocks.4.convs1.1.weight_v
+vq_model.dec.resblocks.4.convs1.2.bias
+vq_model.dec.resblocks.4.convs1.2.weight_g
+vq_model.dec.resblocks.4.convs1.2.weight_v
+vq_model.dec.resblocks.4.convs2.0.bias
+vq_model.dec.resblocks.4.convs2.0.weight_g
+vq_model.dec.resblocks.4.convs2.0.weight_v
+vq_model.dec.resblocks.4.convs2.1.bias
+vq_model.dec.resblocks.4.convs2.1.weight_g
+vq_model.dec.resblocks.4.convs2.1.weight_v
+vq_model.dec.resblocks.4.convs2.2.bias
+vq_model.dec.resblocks.4.convs2.2.weight_g
+vq_model.dec.resblocks.4.convs2.2.weight_v
+vq_model.dec.resblocks.5.convs1.0.bias
+vq_model.dec.resblocks.5.convs1.0.weight_g
+vq_model.dec.resblocks.5.convs1.0.weight_v
+vq_model.dec.resblocks.5.convs1.1.bias
+vq_model.dec.resblocks.5.convs1.1.weight_g
+vq_model.dec.resblocks.5.convs1.1.weight_v
+vq_model.dec.resblocks.5.convs1.2.bias
+vq_model.dec.resblocks.5.convs1.2.weight_g
+vq_model.dec.resblocks.5.convs1.2.weight_v
+vq_model.dec.resblocks.5.convs2.0.bias
+vq_model.dec.resblocks.5.convs2.0.weight_g
+vq_model.dec.resblocks.5.convs2.0.weight_v
+vq_model.dec.resblocks.5.convs2.1.bias
+vq_model.dec.resblocks.5.convs2.1.weight_g
+vq_model.dec.resblocks.5.convs2.1.weight_v
+vq_model.dec.resblocks.5.convs2.2.bias
+vq_model.dec.resblocks.5.convs2.2.weight_g
+vq_model.dec.resblocks.5.convs2.2.weight_v
+vq_model.dec.resblocks.6.convs1.0.bias
+vq_model.dec.resblocks.6.convs1.0.weight_g
+vq_model.dec.resblocks.6.convs1.0.weight_v
+vq_model.dec.resblocks.6.convs1.1.bias
+vq_model.dec.resblocks.6.convs1.1.weight_g
+vq_model.dec.resblocks.6.convs1.1.weight_v
+vq_model.dec.resblocks.6.convs1.2.bias
+vq_model.dec.resblocks.6.convs1.2.weight_g
+vq_model.dec.resblocks.6.convs1.2.weight_v
+vq_model.dec.resblocks.6.convs2.0.bias
+vq_model.dec.resblocks.6.convs2.0.weight_g
+vq_model.dec.resblocks.6.convs2.0.weight_v
+vq_model.dec.resblocks.6.convs2.1.bias
+vq_model.dec.resblocks.6.convs2.1.weight_g
+vq_model.dec.resblocks.6.convs2.1.weight_v
+vq_model.dec.resblocks.6.convs2.2.bias
+vq_model.dec.resblocks.6.convs2.2.weight_g
+vq_model.dec.resblocks.6.convs2.2.weight_v
+vq_model.dec.resblocks.7.convs1.0.bias
+vq_model.dec.resblocks.7.convs1.0.weight_g
+vq_model.dec.resblocks.7.convs1.0.weight_v
+vq_model.dec.resblocks.7.convs1.1.bias
+vq_model.dec.resblocks.7.convs1.1.weight_g
+vq_model.dec.resblocks.7.convs1.1.weight_v
+vq_model.dec.resblocks.7.convs1.2.bias
+vq_model.dec.resblocks.7.convs1.2.weight_g
+vq_model.dec.resblocks.7.convs1.2.weight_v
+vq_model.dec.resblocks.7.convs2.0.bias
+vq_model.dec.resblocks.7.convs2.0.weight_g
+vq_model.dec.resblocks.7.convs2.0.weight_v
+vq_model.dec.resblocks.7.convs2.1.bias
+vq_model.dec.resblocks.7.convs2.1.weight_g
+vq_model.dec.resblocks.7.convs2.1.weight_v
+vq_model.dec.resblocks.7.convs2.2.bias
+vq_model.dec.resblocks.7.convs2.2.weight_g
+vq_model.dec.resblocks.7.convs2.2.weight_v
+vq_model.dec.resblocks.8.convs1.0.bias
+vq_model.dec.resblocks.8.convs1.0.weight_g
+vq_model.dec.resblocks.8.convs1.0.weight_v
+vq_model.dec.resblocks.8.convs1.1.bias
+vq_model.dec.resblocks.8.convs1.1.weight_g
+vq_model.dec.resblocks.8.convs1.1.weight_v
+vq_model.dec.resblocks.8.convs1.2.bias
+vq_model.dec.resblocks.8.convs1.2.weight_g
+vq_model.dec.resblocks.8.convs1.2.weight_v
+vq_model.dec.resblocks.8.convs2.0.bias
+vq_model.dec.resblocks.8.convs2.0.weight_g
+vq_model.dec.resblocks.8.convs2.0.weight_v
+vq_model.dec.resblocks.8.convs2.1.bias
+vq_model.dec.resblocks.8.convs2.1.weight_g
+vq_model.dec.resblocks.8.convs2.1.weight_v
+vq_model.dec.resblocks.8.convs2.2.bias
+vq_model.dec.resblocks.8.convs2.2.weight_g
+vq_model.dec.resblocks.8.convs2.2.weight_v
+vq_model.dec.resblocks.9.convs1.0.bias
+vq_model.dec.resblocks.9.convs1.0.weight_g
+vq_model.dec.resblocks.9.convs1.0.weight_v
+vq_model.dec.resblocks.9.convs1.1.bias
+vq_model.dec.resblocks.9.convs1.1.weight_g
+vq_model.dec.resblocks.9.convs1.1.weight_v
+vq_model.dec.resblocks.9.convs1.2.bias
+vq_model.dec.resblocks.9.convs1.2.weight_g
+vq_model.dec.resblocks.9.convs1.2.weight_v
+vq_model.dec.resblocks.9.convs2.0.bias
+vq_model.dec.resblocks.9.convs2.0.weight_g
+vq_model.dec.resblocks.9.convs2.0.weight_v
+vq_model.dec.resblocks.9.convs2.1.bias
+vq_model.dec.resblocks.9.convs2.1.weight_g
+vq_model.dec.resblocks.9.convs2.1.weight_v
+vq_model.dec.resblocks.9.convs2.2.bias
+vq_model.dec.resblocks.9.convs2.2.weight_g
+vq_model.dec.resblocks.9.convs2.2.weight_v
+vq_model.dec.resblocks.10.convs1.0.bias
+vq_model.dec.resblocks.10.convs1.0.weight_g
+vq_model.dec.resblocks.10.convs1.0.weight_v
+vq_model.dec.resblocks.10.convs1.1.bias
+vq_model.dec.resblocks.10.convs1.1.weight_g
+vq_model.dec.resblocks.10.convs1.1.weight_v
+vq_model.dec.resblocks.10.convs1.2.bias
+vq_model.dec.resblocks.10.convs1.2.weight_g
+vq_model.dec.resblocks.10.convs1.2.weight_v
+vq_model.dec.resblocks.10.convs2.0.bias
+vq_model.dec.resblocks.10.convs2.0.weight_g
+vq_model.dec.resblocks.10.convs2.0.weight_v
+vq_model.dec.resblocks.10.convs2.1.bias
+vq_model.dec.resblocks.10.convs2.1.weight_g
+vq_model.dec.resblocks.10.convs2.1.weight_v
+vq_model.dec.resblocks.10.convs2.2.bias
+vq_model.dec.resblocks.10.convs2.2.weight_g
+vq_model.dec.resblocks.10.convs2.2.weight_v
+vq_model.dec.resblocks.11.convs1.0.bias
+vq_model.dec.resblocks.11.convs1.0.weight_g
+vq_model.dec.resblocks.11.convs1.0.weight_v
+vq_model.dec.resblocks.11.convs1.1.bias
+vq_model.dec.resblocks.11.convs1.1.weight_g
+vq_model.dec.resblocks.11.convs1.1.weight_v
+vq_model.dec.resblocks.11.convs1.2.bias
+vq_model.dec.resblocks.11.convs1.2.weight_g
+vq_model.dec.resblocks.11.convs1.2.weight_v
+vq_model.dec.resblocks.11.convs2.0.bias
+vq_model.dec.resblocks.11.convs2.0.weight_g
+vq_model.dec.resblocks.11.convs2.0.weight_v
+vq_model.dec.resblocks.11.convs2.1.bias
+vq_model.dec.resblocks.11.convs2.1.weight_g
+vq_model.dec.resblocks.11.convs2.1.weight_v
+vq_model.dec.resblocks.11.convs2.2.bias
+vq_model.dec.resblocks.11.convs2.2.weight_g
+vq_model.dec.resblocks.11.convs2.2.weight_v
+vq_model.dec.resblocks.12.convs1.0.bias
+vq_model.dec.resblocks.12.convs1.0.weight_g
+vq_model.dec.resblocks.12.convs1.0.weight_v
+vq_model.dec.resblocks.12.convs1.1.bias
+vq_model.dec.resblocks.12.convs1.1.weight_g
+vq_model.dec.resblocks.12.convs1.1.weight_v
+vq_model.dec.resblocks.12.convs1.2.bias
+vq_model.dec.resblocks.12.convs1.2.weight_g
+vq_model.dec.resblocks.12.convs1.2.weight_v
+vq_model.dec.resblocks.12.convs2.0.bias
+vq_model.dec.resblocks.12.convs2.0.weight_g
+vq_model.dec.resblocks.12.convs2.0.weight_v
+vq_model.dec.resblocks.12.convs2.1.bias
+vq_model.dec.resblocks.12.convs2.1.weight_g
+vq_model.dec.resblocks.12.convs2.1.weight_v
+vq_model.dec.resblocks.12.convs2.2.bias
+vq_model.dec.resblocks.12.convs2.2.weight_g
+vq_model.dec.resblocks.12.convs2.2.weight_v
+vq_model.dec.resblocks.13.convs1.0.bias
+vq_model.dec.resblocks.13.convs1.0.weight_g
+vq_model.dec.resblocks.13.convs1.0.weight_v
+vq_model.dec.resblocks.13.convs1.1.bias
+vq_model.dec.resblocks.13.convs1.1.weight_g
+vq_model.dec.resblocks.13.convs1.1.weight_v
+vq_model.dec.resblocks.13.convs1.2.bias
+vq_model.dec.resblocks.13.convs1.2.weight_g
+vq_model.dec.resblocks.13.convs1.2.weight_v
+vq_model.dec.resblocks.13.convs2.0.bias
+vq_model.dec.resblocks.13.convs2.0.weight_g
+vq_model.dec.resblocks.13.convs2.0.weight_v
+vq_model.dec.resblocks.13.convs2.1.bias
+vq_model.dec.resblocks.13.convs2.1.weight_g
+vq_model.dec.resblocks.13.convs2.1.weight_v
+vq_model.dec.resblocks.13.convs2.2.bias
+vq_model.dec.resblocks.13.convs2.2.weight_g
+vq_model.dec.resblocks.13.convs2.2.weight_v
+vq_model.dec.resblocks.14.convs1.0.bias
+vq_model.dec.resblocks.14.convs1.0.weight_g
+vq_model.dec.resblocks.14.convs1.0.weight_v
+vq_model.dec.resblocks.14.convs1.1.bias
+vq_model.dec.resblocks.14.convs1.1.weight_g
+vq_model.dec.resblocks.14.convs1.1.weight_v
+vq_model.dec.resblocks.14.convs1.2.bias
+vq_model.dec.resblocks.14.convs1.2.weight_g
+vq_model.dec.resblocks.14.convs1.2.weight_v
+vq_model.dec.resblocks.14.convs2.0.bias
+vq_model.dec.resblocks.14.convs2.0.weight_g
+vq_model.dec.resblocks.14.convs2.0.weight_v
+vq_model.dec.resblocks.14.convs2.1.bias
+vq_model.dec.resblocks.14.convs2.1.weight_g
+vq_model.dec.resblocks.14.convs2.1.weight_v
+vq_model.dec.resblocks.14.convs2.2.bias
+vq_model.dec.resblocks.14.convs2.2.weight_g
+vq_model.dec.resblocks.14.convs2.2.weight_v
+vq_model.dec.conv_post.weight
+vq_model.dec.cond.weight
+vq_model.dec.cond.bias
+vq_model.flow.flows.0.pre.weight
+vq_model.flow.flows.0.pre.bias
+vq_model.flow.flows.0.enc.in_layers.0.bias
+vq_model.flow.flows.0.enc.in_layers.0.weight_g
+vq_model.flow.flows.0.enc.in_layers.0.weight_v
+vq_model.flow.flows.0.enc.in_layers.1.bias
+vq_model.flow.flows.0.enc.in_layers.1.weight_g
+vq_model.flow.flows.0.enc.in_layers.1.weight_v
+vq_model.flow.flows.0.enc.in_layers.2.bias
+vq_model.flow.flows.0.enc.in_layers.2.weight_g
+vq_model.flow.flows.0.enc.in_layers.2.weight_v
+vq_model.flow.flows.0.enc.in_layers.3.bias
+vq_model.flow.flows.0.enc.in_layers.3.weight_g
+vq_model.flow.flows.0.enc.in_layers.3.weight_v
+vq_model.flow.flows.0.enc.res_skip_layers.0.bias
+vq_model.flow.flows.0.enc.res_skip_layers.0.weight_g
+vq_model.flow.flows.0.enc.res_skip_layers.0.weight_v
+vq_model.flow.flows.0.enc.res_skip_layers.1.bias
+vq_model.flow.flows.0.enc.res_skip_layers.1.weight_g
+vq_model.flow.flows.0.enc.res_skip_layers.1.weight_v
+vq_model.flow.flows.0.enc.res_skip_layers.2.bias
+vq_model.flow.flows.0.enc.res_skip_layers.2.weight_g
+vq_model.flow.flows.0.enc.res_skip_layers.2.weight_v
+vq_model.flow.flows.0.enc.res_skip_layers.3.bias
+vq_model.flow.flows.0.enc.res_skip_layers.3.weight_g
+vq_model.flow.flows.0.enc.res_skip_layers.3.weight_v
+vq_model.flow.flows.0.enc.cond_layer.bias
+vq_model.flow.flows.0.enc.cond_layer.weight_g
+vq_model.flow.flows.0.enc.cond_layer.weight_v
+vq_model.flow.flows.0.post.weight
+vq_model.flow.flows.0.post.bias
+vq_model.flow.flows.2.pre.weight
+vq_model.flow.flows.2.pre.bias
+vq_model.flow.flows.2.enc.in_layers.0.bias
+vq_model.flow.flows.2.enc.in_layers.0.weight_g
+vq_model.flow.flows.2.enc.in_layers.0.weight_v
+vq_model.flow.flows.2.enc.in_layers.1.bias
+vq_model.flow.flows.2.enc.in_layers.1.weight_g
+vq_model.flow.flows.2.enc.in_layers.1.weight_v
+vq_model.flow.flows.2.enc.in_layers.2.bias
+vq_model.flow.flows.2.enc.in_layers.2.weight_g
+vq_model.flow.flows.2.enc.in_layers.2.weight_v
+vq_model.flow.flows.2.enc.in_layers.3.bias
+vq_model.flow.flows.2.enc.in_layers.3.weight_g
+vq_model.flow.flows.2.enc.in_layers.3.weight_v
+vq_model.flow.flows.2.enc.res_skip_layers.0.bias
+vq_model.flow.flows.2.enc.res_skip_layers.0.weight_g
+vq_model.flow.flows.2.enc.res_skip_layers.0.weight_v
+vq_model.flow.flows.2.enc.res_skip_layers.1.bias
+vq_model.flow.flows.2.enc.res_skip_layers.1.weight_g
+vq_model.flow.flows.2.enc.res_skip_layers.1.weight_v
+vq_model.flow.flows.2.enc.res_skip_layers.2.bias
+vq_model.flow.flows.2.enc.res_skip_layers.2.weight_g
+vq_model.flow.flows.2.enc.res_skip_layers.2.weight_v
+vq_model.flow.flows.2.enc.res_skip_layers.3.bias
+vq_model.flow.flows.2.enc.res_skip_layers.3.weight_g
+vq_model.flow.flows.2.enc.res_skip_layers.3.weight_v
+vq_model.flow.flows.2.enc.cond_layer.bias
+vq_model.flow.flows.2.enc.cond_layer.weight_g
+vq_model.flow.flows.2.enc.cond_layer.weight_v
+vq_model.flow.flows.2.post.weight
+vq_model.flow.flows.2.post.bias
+vq_model.flow.flows.4.pre.weight
+vq_model.flow.flows.4.pre.bias
+vq_model.flow.flows.4.enc.in_layers.0.bias
+vq_model.flow.flows.4.enc.in_layers.0.weight_g
+vq_model.flow.flows.4.enc.in_layers.0.weight_v
+vq_model.flow.flows.4.enc.in_layers.1.bias
+vq_model.flow.flows.4.enc.in_layers.1.weight_g
+vq_model.flow.flows.4.enc.in_layers.1.weight_v
+vq_model.flow.flows.4.enc.in_layers.2.bias
+vq_model.flow.flows.4.enc.in_layers.2.weight_g
+vq_model.flow.flows.4.enc.in_layers.2.weight_v
+vq_model.flow.flows.4.enc.in_layers.3.bias
+vq_model.flow.flows.4.enc.in_layers.3.weight_g
+vq_model.flow.flows.4.enc.in_layers.3.weight_v
+vq_model.flow.flows.4.enc.res_skip_layers.0.bias
+vq_model.flow.flows.4.enc.res_skip_layers.0.weight_g
+vq_model.flow.flows.4.enc.res_skip_layers.0.weight_v
+vq_model.flow.flows.4.enc.res_skip_layers.1.bias
+vq_model.flow.flows.4.enc.res_skip_layers.1.weight_g
+vq_model.flow.flows.4.enc.res_skip_layers.1.weight_v
+vq_model.flow.flows.4.enc.res_skip_layers.2.bias
+vq_model.flow.flows.4.enc.res_skip_layers.2.weight_g
+vq_model.flow.flows.4.enc.res_skip_layers.2.weight_v
+vq_model.flow.flows.4.enc.res_skip_layers.3.bias
+vq_model.flow.flows.4.enc.res_skip_layers.3.weight_g
+vq_model.flow.flows.4.enc.res_skip_layers.3.weight_v
+vq_model.flow.flows.4.enc.cond_layer.bias
+vq_model.flow.flows.4.enc.cond_layer.weight_g
+vq_model.flow.flows.4.enc.cond_layer.weight_v
+vq_model.flow.flows.4.post.weight
+vq_model.flow.flows.4.post.bias
+vq_model.flow.flows.6.pre.weight
+vq_model.flow.flows.6.pre.bias
+vq_model.flow.flows.6.enc.in_layers.0.bias
+vq_model.flow.flows.6.enc.in_layers.0.weight_g
+vq_model.flow.flows.6.enc.in_layers.0.weight_v
+vq_model.flow.flows.6.enc.in_layers.1.bias
+vq_model.flow.flows.6.enc.in_layers.1.weight_g
+vq_model.flow.flows.6.enc.in_layers.1.weight_v
+vq_model.flow.flows.6.enc.in_layers.2.bias
+vq_model.flow.flows.6.enc.in_layers.2.weight_g
+vq_model.flow.flows.6.enc.in_layers.2.weight_v
+vq_model.flow.flows.6.enc.in_layers.3.bias
+vq_model.flow.flows.6.enc.in_layers.3.weight_g
+vq_model.flow.flows.6.enc.in_layers.3.weight_v
+vq_model.flow.flows.6.enc.res_skip_layers.0.bias
+vq_model.flow.flows.6.enc.res_skip_layers.0.weight_g
+vq_model.flow.flows.6.enc.res_skip_layers.0.weight_v
+vq_model.flow.flows.6.enc.res_skip_layers.1.bias
+vq_model.flow.flows.6.enc.res_skip_layers.1.weight_g
+vq_model.flow.flows.6.enc.res_skip_layers.1.weight_v
+vq_model.flow.flows.6.enc.res_skip_layers.2.bias
+vq_model.flow.flows.6.enc.res_skip_layers.2.weight_g
+vq_model.flow.flows.6.enc.res_skip_layers.2.weight_v
+vq_model.flow.flows.6.enc.res_skip_layers.3.bias
+vq_model.flow.flows.6.enc.res_skip_layers.3.weight_g
+vq_model.flow.flows.6.enc.res_skip_layers.3.weight_v
+vq_model.flow.flows.6.enc.cond_layer.bias
+vq_model.flow.flows.6.enc.cond_layer.weight_g
+vq_model.flow.flows.6.enc.cond_layer.weight_v
+vq_model.flow.flows.6.post.weight
+vq_model.flow.flows.6.post.bias
+vq_model.quantizer.vq.layers.0._codebook.embed
diff --git a/genie_tts/G2P/Chinese/CorrectPronunciation.py b/genie_tts/G2P/Chinese/CorrectPronunciation.py
index 33d2024fd9ccea293e44afad15a2819235848f44..804b934ebd49fa390cd93e40c67ae820d15a9c2c 100644
--- a/genie_tts/G2P/Chinese/CorrectPronunciation.py
+++ b/genie_tts/G2P/Chinese/CorrectPronunciation.py
@@ -1,50 +1,50 @@
-import os
-import pickle
-from typing import List, Dict, Any, Union
-
-from ...Core.Resources import Chinese_G2P_DIR
-
-# 常量定义
-DEFAULT_CACHE_PATH = os.path.join(Chinese_G2P_DIR, "polyphonic.pickle")
-
-
-class PolyphonicDictManager:
-    _data: Dict[str, Any] = {}
-
-    @classmethod
-    def get_data(cls, path: str = DEFAULT_CACHE_PATH) -> Dict[str, Any]:
-        if not cls._data:
-            with open(path, "rb") as f:
-                cls._data = pickle.load(f)
-        return cls._data
-
-
-def correct_pronunciation(word: str, word_pinyin: List[str]) -> Union[List[str], str]:
-    """
-        根据加载的字典修正发音，作为供外部程序调用的独立接口。
-        逻辑：优先查找整词修正，如果没有整词匹配，则遍历每个字符进行单字修正。
-
-        Input:
-            word (str): 原始中文字符串，例如 "银行"。
-            word_pinyins (List[str]): 当前预测的拼音列表，例如 ['yin2', 'xing2']。
-
-        Output:
-            Union[List[str], str]: 修正后的拼音列表或字符串。
-
-        Example:
-            # 字典包含整词 {'银行': ['yin2', 'hang2']}
-            result = correct_pronunciation("银行", ["yin2", "xing2"])
-            # Result: ["yin2", "hang2"]
-        """
-    pp_dict = PolyphonicDictManager.get_data()
-    new_word_pinyin = list(word_pinyin)
-    # 1. 尝试整词匹配
-    if new_pinyin := pp_dict.get(word):
-        return new_pinyin
-    # 2. 逐字修正
-    for idx, w in enumerate(word):
-        if idx >= len(new_word_pinyin):
-            break
-        if w_pinyin := pp_dict.get(w):
-            new_word_pinyin[idx] = w_pinyin[0]
-    return new_word_pinyin
+import os
+import pickle
+from typing import List, Dict, Any, Union
+
+from ...Core.Resources import Chinese_G2P_DIR
+
+# 常量定义
+DEFAULT_CACHE_PATH = os.path.join(Chinese_G2P_DIR, "polyphonic.pickle")
+
+
+class PolyphonicDictManager:
+    _data: Dict[str, Any] = {}
+
+    @classmethod
+    def get_data(cls, path: str = DEFAULT_CACHE_PATH) -> Dict[str, Any]:
+        if not cls._data:
+            with open(path, "rb") as f:
+                cls._data = pickle.load(f)
+        return cls._data
+
+
+def correct_pronunciation(word: str, word_pinyin: List[str]) -> Union[List[str], str]:
+    """
+        根据加载的字典修正发音，作为供外部程序调用的独立接口。
+        逻辑：优先查找整词修正，如果没有整词匹配，则遍历每个字符进行单字修正。
+
+        Input:
+            word (str): 原始中文字符串，例如 "银行"。
+            word_pinyins (List[str]): 当前预测的拼音列表，例如 ['yin2', 'xing2']。
+
+        Output:
+            Union[List[str], str]: 修正后的拼音列表或字符串。
+
+        Example:
+            # 字典包含整词 {'银行': ['yin2', 'hang2']}
+            result = correct_pronunciation("银行", ["yin2", "xing2"])
+            # Result: ["yin2", "hang2"]
+        """
+    pp_dict = PolyphonicDictManager.get_data()
+    new_word_pinyin = list(word_pinyin)
+    # 1. 尝试整词匹配
+    if new_pinyin := pp_dict.get(word):
+        return new_pinyin
+    # 2. 逐字修正
+    for idx, w in enumerate(word):
+        if idx >= len(new_word_pinyin):
+            break
+        if w_pinyin := pp_dict.get(w):
+            new_word_pinyin[idx] = w_pinyin[0]
+    return new_word_pinyin
diff --git a/genie_tts/G2P/Chinese/Erhua.py b/genie_tts/G2P/Chinese/Erhua.py
index 8ccc87b895563bc03e4943427219a2495094208b..8f165773fe4502163e624e73001a275813ff321e 100644
--- a/genie_tts/G2P/Chinese/Erhua.py
+++ b/genie_tts/G2P/Chinese/Erhua.py
@@ -1,49 +1,49 @@
-from typing import List, Tuple, Set
-
-
-class ErhuaProcessor:
-    """
-    处理中文G2P中的儿化音逻辑。
-    """
-
-    def __init__(self):
-        self.must_erhua: Set[str] = {
-            "小院儿", "胡同儿", "范儿", "老汉儿", "撒欢儿", "寻老礼儿", "妥妥儿", "媳妇儿"
-        }
-        self.not_erhua: Set[str] = {
-            "虐儿", "为儿", "护儿", "瞒儿", "救儿", "替儿", "有儿", "一儿", "我儿", "俺儿",
-            "妻儿", "拐儿", "聋儿", "乞儿", "患儿", "幼儿", "孤儿", "婴儿", "婴幼儿", "连体儿",
-            "脑瘫儿", "流浪儿", "体弱儿", "混血儿", "蜜雪儿", "舫儿", "祖儿", "美儿", "应采儿", "可儿",
-            "侄儿", "孙儿", "侄孙儿", "女儿", "男儿", "红孩儿", "花儿", "虫儿", "马儿", "鸟儿",
-            "猪儿", "猫儿", "狗儿", "少儿",
-        }
-
-    def merge_erhua(self, initials: List[str], finals: List[str], word: str, pos: str) -> Tuple[List[str], List[str]]:
-        # 1. 修正 er1 发音为 er2 (当'儿'在词尾且发音为er1时)
-        for i, phn in enumerate(finals):
-            if i == len(finals) - 1 and word[i] == "儿" and phn == "er1":
-                finals[i] = "er2"
-        # 2. 检查是否跳过儿化处理
-        if word not in self.must_erhua and (word in self.not_erhua or pos in {"a", "j", "nr"}):
-            return initials, finals
-        # 3. 长度校验 (处理如 "……" 等长度不一致的特殊符号情况)
-        if len(finals) != len(word):
-            return initials, finals
-        # 4. 执行儿化合并逻辑 (与前一个字发同音)
-        new_initials = []
-        new_finals = []
-        for i, phn in enumerate(finals):
-            # 判断是否需要合并儿化音
-            # 条件: 是最后一个字 + 是"儿" + 发音是er2/er5 + 后两字不在非儿化表中 + 前面已有韵母
-            if (
-                    i == len(finals) - 1
-                    and word[i] == "儿"
-                    and phn in {"er2", "er5"}
-                    and word[-2:] not in self.not_erhua
-                    and new_finals
-            ):
-                # 将 'er' 加上前一个字的声调
-                phn = "er" + new_finals[-1][-1]
-            new_initials.append(initials[i])
-            new_finals.append(phn)
-        return new_initials, new_finals
+from typing import List, Tuple, Set
+
+
+class ErhuaProcessor:
+    """
+    处理中文G2P中的儿化音逻辑。
+    """
+
+    def __init__(self):
+        self.must_erhua: Set[str] = {
+            "小院儿", "胡同儿", "范儿", "老汉儿", "撒欢儿", "寻老礼儿", "妥妥儿", "媳妇儿"
+        }
+        self.not_erhua: Set[str] = {
+            "虐儿", "为儿", "护儿", "瞒儿", "救儿", "替儿", "有儿", "一儿", "我儿", "俺儿",
+            "妻儿", "拐儿", "聋儿", "乞儿", "患儿", "幼儿", "孤儿", "婴儿", "婴幼儿", "连体儿",
+            "脑瘫儿", "流浪儿", "体弱儿", "混血儿", "蜜雪儿", "舫儿", "祖儿", "美儿", "应采儿", "可儿",
+            "侄儿", "孙儿", "侄孙儿", "女儿", "男儿", "红孩儿", "花儿", "虫儿", "马儿", "鸟儿",
+            "猪儿", "猫儿", "狗儿", "少儿",
+        }
+
+    def merge_erhua(self, initials: List[str], finals: List[str], word: str, pos: str) -> Tuple[List[str], List[str]]:
+        # 1. 修正 er1 发音为 er2 (当'儿'在词尾且发音为er1时)
+        for i, phn in enumerate(finals):
+            if i == len(finals) - 1 and word[i] == "儿" and phn == "er1":
+                finals[i] = "er2"
+        # 2. 检查是否跳过儿化处理
+        if word not in self.must_erhua and (word in self.not_erhua or pos in {"a", "j", "nr"}):
+            return initials, finals
+        # 3. 长度校验 (处理如 "……" 等长度不一致的特殊符号情况)
+        if len(finals) != len(word):
+            return initials, finals
+        # 4. 执行儿化合并逻辑 (与前一个字发同音)
+        new_initials = []
+        new_finals = []
+        for i, phn in enumerate(finals):
+            # 判断是否需要合并儿化音
+            # 条件: 是最后一个字 + 是"儿" + 发音是er2/er5 + 后两字不在非儿化表中 + 前面已有韵母
+            if (
+                    i == len(finals) - 1
+                    and word[i] == "儿"
+                    and phn in {"er2", "er5"}
+                    and word[-2:] not in self.not_erhua
+                    and new_finals
+            ):
+                # 将 'er' 加上前一个字的声调
+                phn = "er" + new_finals[-1][-1]
+            new_initials.append(initials[i])
+            new_finals.append(phn)
+        return new_initials, new_finals
diff --git a/genie_tts/G2P/Chinese/Normalization/__pycache__/__init__.cpython-311.pyc b/genie_tts/G2P/Chinese/Normalization/__pycache__/__init__.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..a28a5edbecd11f764acad8f84bfb1ae97dc05c7f
Binary files /dev/null and b/genie_tts/G2P/Chinese/Normalization/__pycache__/__init__.cpython-311.pyc differ
diff --git a/genie_tts/G2P/Chinese/Normalization/__pycache__/char_convert.cpython-311.pyc b/genie_tts/G2P/Chinese/Normalization/__pycache__/char_convert.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..8c423046a6d37b42a65b88ceda68c3adb8c752a6
Binary files /dev/null and b/genie_tts/G2P/Chinese/Normalization/__pycache__/char_convert.cpython-311.pyc differ
diff --git a/genie_tts/G2P/Chinese/Normalization/__pycache__/chronology.cpython-311.pyc b/genie_tts/G2P/Chinese/Normalization/__pycache__/chronology.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5ae25e3fcd6b370d61ac5da6a6b7761dcb7be676
Binary files /dev/null and b/genie_tts/G2P/Chinese/Normalization/__pycache__/chronology.cpython-311.pyc differ
diff --git a/genie_tts/G2P/Chinese/Normalization/__pycache__/constants.cpython-311.pyc b/genie_tts/G2P/Chinese/Normalization/__pycache__/constants.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f0827c54a2486be0ab595c072483466c64980ff8
Binary files /dev/null and b/genie_tts/G2P/Chinese/Normalization/__pycache__/constants.cpython-311.pyc differ
diff --git a/genie_tts/G2P/Chinese/Normalization/__pycache__/num.cpython-311.pyc b/genie_tts/G2P/Chinese/Normalization/__pycache__/num.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..acab141bcac1d0b822551759c8e4b4f63da27c90
Binary files /dev/null and b/genie_tts/G2P/Chinese/Normalization/__pycache__/num.cpython-311.pyc differ
diff --git a/genie_tts/G2P/Chinese/Normalization/__pycache__/phonecode.cpython-311.pyc b/genie_tts/G2P/Chinese/Normalization/__pycache__/phonecode.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..43e0328a0752c02a318f3809a8ab14caacd18805
Binary files /dev/null and b/genie_tts/G2P/Chinese/Normalization/__pycache__/phonecode.cpython-311.pyc differ
diff --git a/genie_tts/G2P/Chinese/Normalization/__pycache__/quantifier.cpython-311.pyc b/genie_tts/G2P/Chinese/Normalization/__pycache__/quantifier.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6d7d85454ecd3203bbaa0fa238a3533e6c86c67c
Binary files /dev/null and b/genie_tts/G2P/Chinese/Normalization/__pycache__/quantifier.cpython-311.pyc differ
diff --git a/genie_tts/G2P/Chinese/Normalization/__pycache__/text_normlization.cpython-311.pyc b/genie_tts/G2P/Chinese/Normalization/__pycache__/text_normlization.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..7ca61c154b193ec28deaaad1a00690e5efb27c25
Binary files /dev/null and b/genie_tts/G2P/Chinese/Normalization/__pycache__/text_normlization.cpython-311.pyc differ
diff --git a/genie_tts/G2P/Chinese/ToneSandhi.py b/genie_tts/G2P/Chinese/ToneSandhi.py
index e57622b21f3abbd13a3ada62c67c96e9aa67e57e..d9076e3c3f34121c04f42a3291084096a3e0b06b 100644
--- a/genie_tts/G2P/Chinese/ToneSandhi.py
+++ b/genie_tts/G2P/Chinese/ToneSandhi.py
@@ -1,354 +1,354 @@
-# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-中文拼音变调（Tone Sandhi）自动处理器
-"""
-
-from typing import List
-from typing import Tuple
-import jieba_fast as jieba
-from pypinyin import lazy_pinyin
-from pypinyin import Style
-
-
-class ToneSandhi:
-    def __init__(self):
-        self.must_neural_tone_words = {
-            "麻烦", "麻利", "鸳鸯", "高粱", "骨头", "骆驼", "马虎", "首饰", "馒头", "馄饨",
-            "风筝", "难为", "队伍", "阔气", "闺女", "门道", "锄头", "铺盖", "铃铛", "铁匠",
-            "钥匙", "里脊", "里头", "部分", "那么", "道士", "造化", "迷糊", "连累", "这么",
-            "这个", "运气", "过去", "软和", "转悠", "踏实", "跳蚤", "跟头", "趔趄", "财主",
-            "豆腐", "讲究", "记性", "记号", "认识", "规矩", "见识", "裁缝", "补丁", "衣裳",
-            "衣服", "衙门", "街坊", "行李", "行当", "蛤蟆", "蘑菇", "薄荷", "葫芦", "葡萄",
-            "萝卜", "荸荠", "苗条", "苗头", "苍蝇", "芝麻", "舒服", "舒坦", "舌头", "自在",
-            "膏药", "脾气", "脑袋", "脊梁", "能耐", "胳膊", "胭脂", "胡萝", "胡琴", "胡同",
-            "聪明", "耽误", "耽搁", "耷拉", "耳朵", "老爷", "老实", "老婆", "老头", "老太",
-            "翻腾", "罗嗦", "罐头", "编辑", "结实", "红火", "累赘", "糨糊", "糊涂", "精神",
-            "粮食", "簸箕", "篱笆", "算计", "算盘", "答应", "笤帚", "笑语", "笑话", "窟窿",
-            "窝囊", "窗户", "稳当", "稀罕", "称呼", "秧歌", "秀气", "秀才", "福气", "祖宗",
-            "砚台", "码头", "石榴", "石头", "石匠", "知识", "眼睛", "眯缝", "眨巴", "眉毛",
-            "相声", "盘算", "白净", "痢疾", "痛快", "疟疾", "疙瘩", "疏忽", "畜生", "生意",
-            "甘蔗", "琵琶", "琢磨", "琉璃", "玻璃", "玫瑰", "玄乎", "狐狸", "状元", "特务",
-            "牲口", "牙碜", "牌楼", "爽快", "爱人", "热闹", "烧饼", "烟筒", "烂糊", "点心",
-            "炊帚", "灯笼", "火候", "漂亮", "滑溜", "溜达", "温和", "清楚", "消息", "浪头",
-            "活泼", "比方", "正经", "欺负", "模糊", "槟榔", "棺材", "棒槌", "棉花", "核桃",
-            "栅栏", "柴火", "架势", "枕头", "枇杷", "机灵", "本事", "木头", "木匠", "朋友",
-            "月饼", "月亮", "暖和", "明白", "时候", "新鲜", "故事", "收拾", "收成", "提防",
-            "挖苦", "挑剔", "指甲", "指头", "拾掇", "拳头", "拨弄", "招牌", "招呼", "抬举",
-            "护士", "折腾", "扫帚", "打量", "打算", "打点", "打扮", "打听", "打发", "扎实",
-            "扁担", "戒指", "懒得", "意识", "意思", "情形", "悟性", "怪物", "思量", "怎么",
-            "念头", "念叨", "快活", "忙活", "志气", "心思", "得罪", "张罗", "弟兄", "开通",
-            "应酬", "庄稼", "干事", "帮手", "帐篷", "希罕", "师父", "师傅", "巴结", "巴掌",
-            "差事", "工夫", "岁数", "屁股", "尾巴", "少爷", "小气", "小伙", "将就", "对头",
-            "对付", "寡妇", "家伙", "客气", "实在", "官司", "学问", "学生", "字号", "嫁妆",
-            "媳妇", "媒人", "婆家", "娘家", "委屈", "姑娘", "姐夫", "妯娌", "妥当", "妖精",
-            "奴才", "女婿", "头发", "太阳", "大爷", "大方", "大意", "大夫", "多少", "多么",
-            "外甥", "壮实", "地道", "地方", "在乎", "困难", "嘴巴", "嘱咐", "嘟囔", "嘀咕",
-            "喜欢", "喇嘛", "喇叭", "商量", "唾沫", "哑巴", "哈欠", "哆嗦", "咳嗽", "和尚",
-            "告诉", "告示", "含糊", "吓唬", "后头", "名字", "名堂", "合同", "吆喝", "叫唤",
-            "口袋", "厚道", "厉害", "千斤", "包袱", "包涵", "匀称", "勤快", "动静", "动弹",
-            "功夫", "力气", "前头", "刺猬", "刺激", "别扭", "利落", "利索", "利害", "分析",
-            "出息", "凑合", "凉快", "冷战", "冤枉", "冒失", "养活", "关系", "先生", "兄弟",
-            "便宜", "使唤", "佩服", "作坊", "体面", "位置", "似的", "伙计", "休息", "什么",
-            "人家", "亲戚", "亲家", "交情", "云彩", "事情", "买卖", "主意", "丫头", "丧气",
-            "两口", "东西", "东家", "世故", "不由", "不在", "下水", "下巴", "上头", "上司",
-            "丈夫", "丈人", "一辈", "那个", "菩萨", "父亲", "母亲", "咕噜", "邋遢", "费用",
-            "冤家", "甜头", "介绍", "荒唐", "大人", "泥鳅", "幸福", "熟悉", "计划", "扑腾",
-            "蜡烛", "姥爷", "照顾", "喉咙", "吉他", "弄堂", "蚂蚱", "凤凰", "拖沓", "寒碜",
-            "糟蹋", "倒腾", "报复", "逻辑", "盘缠", "喽啰", "牢骚", "咖喱", "扫把", "惦记",
-        }
-        self.must_not_neural_tone_words = {
-            "男子", "女子", "分子", "原子", "量子", "莲子", "石子", "瓜子", "电子", "人人",
-            "虎虎", "幺幺", "干嘛", "学子", "哈哈", "数数", "袅袅", "局地", "以下", "娃哈哈",
-            "花花草草", "留得", "耕地", "想想", "熙熙", "攘攘", "卵子", "死死", "冉冉", "恳恳",
-            "佼佼", "吵吵", "打打", "考考", "整整", "莘莘", "落地", "算子", "家家户户", "青青",
-        }
-        self.punc = "：，；。？！“”‘’':,;.?!"
-
-    # the meaning of jieba pos tag: https://blog.csdn.net/weixin_44174352/article/details/113731041
-    # e.g.
-    # word: "家里"
-    # pos: "s"
-    # finals: ['ia1', 'i3']
-    def _neural_sandhi(self, word: str, pos: str, finals: List[str]) -> List[str]:
-        # reduplication words for n. and v. e.g. 奶奶, 试试, 旺旺
-        for j, item in enumerate(word):
-            if (
-                    j - 1 >= 0
-                    and item == word[j - 1]
-                    and pos[0] in {"n", "v", "a"}
-                    and word not in self.must_not_neural_tone_words
-            ):
-                finals[j] = finals[j][:-1] + "5"
-        ge_idx = word.find("个")
-        if len(word) >= 1 and word[-1] in "吧呢哈啊呐噻嘛吖嗨呐哦哒额滴哩哟喽啰耶喔诶":
-            finals[-1] = finals[-1][:-1] + "5"
-        elif len(word) >= 1 and word[-1] in "的地得":
-            finals[-1] = finals[-1][:-1] + "5"
-        # e.g. 走了, 看着, 去过
-        elif len(word) == 1 and word in "了着过" and pos in {"ul", "uz", "ug"}:
-            finals[-1] = finals[-1][:-1] + "5"
-        elif len(word) > 1 and word[-1] in "们子" and pos in {"r", "n"} and word not in self.must_not_neural_tone_words:
-            finals[-1] = finals[-1][:-1] + "5"
-        # e.g. 桌上, 地下, 家里
-        elif len(word) > 1 and word[-1] in "上下里" and pos in {"s", "l", "f"}:
-            finals[-1] = finals[-1][:-1] + "5"
-        # e.g. 上来, 下去
-        elif len(word) > 1 and word[-1] in "来去" and word[-2] in "上下进出回过起开":
-            finals[-1] = finals[-1][:-1] + "5"
-        # 个做量词
-        elif (
-                ge_idx >= 1 and (word[ge_idx - 1].isnumeric() or word[ge_idx - 1] in "几有两半多各整每做是")
-        ) or word == "个":
-            finals[ge_idx] = finals[ge_idx][:-1] + "5"
-        else:
-            if word in self.must_neural_tone_words or word[-2:] in self.must_neural_tone_words:
-                finals[-1] = finals[-1][:-1] + "5"
-
-        word_list = self._split_word(word)
-        finals_list = [finals[: len(word_list[0])], finals[len(word_list[0]):]]
-        for i, word in enumerate(word_list):
-            # conventional neural in Chinese
-            if word in self.must_neural_tone_words or word[-2:] in self.must_neural_tone_words:
-                finals_list[i][-1] = finals_list[i][-1][:-1] + "5"
-        finals = sum(finals_list, [])
-        return finals
-
-    @staticmethod
-    def _bu_sandhi(word: str, finals: List[str]) -> List[str]:
-        # e.g. 看不懂
-        if len(word) == 3 and word[1] == "不":
-            finals[1] = finals[1][:-1] + "5"
-        else:
-            for i, char in enumerate(word):
-                # "不" before tone4 should be bu2, e.g. 不怕
-                if char == "不" and i + 1 < len(word) and finals[i + 1][-1] == "4":
-                    finals[i] = finals[i][:-1] + "2"
-        return finals
-
-    def _yi_sandhi(self, word: str, finals: List[str]) -> List[str]:
-        # "一" in number sequences, e.g. 一零零, 二一零
-        if word.find("一") != -1 and all([item.isnumeric() for item in word if item != "一"]):
-            return finals
-        # "一" between reduplication words should be yi5, e.g. 看一看
-        elif len(word) == 3 and word[1] == "一" and word[0] == word[-1]:
-            finals[1] = finals[1][:-1] + "5"
-        # when "一" is ordinal word, it should be yi1
-        elif word.startswith("第一"):
-            finals[1] = finals[1][:-1] + "1"
-        else:
-            for i, char in enumerate(word):
-                if char == "一" and i + 1 < len(word):
-                    # "一" before tone4 should be yi2, e.g. 一段
-                    if finals[i + 1][-1] == "4":
-                        finals[i] = finals[i][:-1] + "2"
-                    # "一" before non-tone4 should be yi4, e.g. 一天
-                    else:
-                        # "一" 后面如果是标点，还读一声
-                        if word[i + 1] not in self.punc:
-                            finals[i] = finals[i][:-1] + "4"
-        return finals
-
-    @staticmethod
-    def _split_word(word: str) -> List[str]:
-        word_list = jieba.cut_for_search(word)
-        word_list = sorted(word_list, key=lambda i: len(i), reverse=False)
-        first_subword = word_list[0]
-        first_begin_idx = word.find(first_subword)
-        if first_begin_idx == 0:
-            second_subword = word[len(first_subword):]
-            new_word_list = [first_subword, second_subword]
-        else:
-            second_subword = word[: -len(first_subword)]
-            new_word_list = [second_subword, first_subword]
-        return new_word_list
-
-    def _three_sandhi(self, word: str, finals: List[str]) -> List[str]:
-        if len(word) == 2 and self._all_tone_three(finals):
-            finals[0] = finals[0][:-1] + "2"
-        elif len(word) == 3:
-            word_list = self._split_word(word)
-            if self._all_tone_three(finals):
-                #  disyllabic + monosyllabic, e.g. 蒙古/包
-                if len(word_list[0]) == 2:
-                    finals[0] = finals[0][:-1] + "2"
-                    finals[1] = finals[1][:-1] + "2"
-                #  monosyllabic + disyllabic, e.g. 纸/老虎
-                elif len(word_list[0]) == 1:
-                    finals[1] = finals[1][:-1] + "2"
-            else:
-                finals_list = [finals[: len(word_list[0])], finals[len(word_list[0]):]]
-                if len(finals_list) == 2:
-                    for i, sub in enumerate(finals_list):
-                        # e.g. 所有/人
-                        if self._all_tone_three(sub) and len(sub) == 2:
-                            finals_list[i][0] = finals_list[i][0][:-1] + "2"
-                        # e.g. 好/喜欢
-                        elif (
-                                i == 1
-                                and not self._all_tone_three(sub)
-                                and finals_list[i][0][-1] == "3"
-                                and finals_list[0][-1][-1] == "3"
-                        ):
-                            finals_list[0][-1] = finals_list[0][-1][:-1] + "2"
-                        finals = sum(finals_list, [])
-        # split idiom into two words whose length is 2
-        elif len(word) == 4:
-            finals_list = [finals[:2], finals[2:]]
-            finals = []
-            for sub in finals_list:
-                if self._all_tone_three(sub):
-                    sub[0] = sub[0][:-1] + "2"
-                finals += sub
-
-        return finals
-
-    @staticmethod
-    def _all_tone_three(finals: List[str]) -> bool:
-        # 增加 len(x) > 0 的判断，防止空字符串导致崩溃
-        return all(len(x) > 0 and x[-1] == "3" for x in finals)
-
-    @staticmethod
-    def _merge_bu(seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
-        new_seg = []
-        last_word = ""
-        for word, pos in seg:
-            if last_word == "不":
-                word = last_word + word
-            if word != "不":
-                new_seg.append((word, pos))
-            last_word = word[:]
-        if last_word == "不":
-            new_seg.append((last_word, "d"))
-        return new_seg
-
-    @staticmethod
-    def _merge_yi(seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
-        new_seg = []
-        i = 0
-        # function 1
-        while i < len(seg):
-            word, pos = seg[i]
-            merged = False
-            if i - 1 >= 0 and word == "一" and i + 1 < len(seg):
-                last = new_seg[-1] if new_seg else seg[i - 1]
-                if last[0] == seg[i + 1][0] and last[1] == "v" and seg[i + 1][1] == "v":
-                    combined = last[0] + "一" + seg[i + 1][0]
-                    new_seg[-1] = [combined, last[1]]
-                    i += 2
-                    merged = True
-            if not merged:
-                new_seg.append([word, pos])
-                i += 1
-        seg = new_seg
-        new_seg = []
-        # function 2
-        for word, pos in seg:
-            if new_seg and new_seg[-1][0] == "一":
-                new_seg[-1][0] = new_seg[-1][0] + word
-            else:
-                new_seg.append([word, pos])
-        return new_seg
-
-    # the first and the second words are all_tone_three
-    def _merge_continuous_three_tones(self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
-        new_seg = []
-        sub_finals_list = [
-            lazy_pinyin(word, neutral_tone_with_five=True, style=Style.FINALS_TONE3) for (word, pos) in seg
-        ]
-        assert len(sub_finals_list) == len(seg)
-        merge_last = [False] * len(seg)
-        for i, (word, pos) in enumerate(seg):
-            if (
-                    i - 1 >= 0
-                    and self._all_tone_three(sub_finals_list[i - 1])
-                    and self._all_tone_three(sub_finals_list[i])
-                    and not merge_last[i - 1]
-            ):
-                # if the last word is reduplication, not merge, because reduplication need to be _neural_sandhi
-                if not self._is_reduplication(seg[i - 1][0]) and len(seg[i - 1][0]) + len(seg[i][0]) <= 3:
-                    new_seg[-1][0] = new_seg[-1][0] + seg[i][0]
-                    merge_last[i] = True
-                else:
-                    new_seg.append([word, pos])
-            else:
-                new_seg.append([word, pos])
-
-        return new_seg
-
-    @staticmethod
-    def _is_reduplication(word: str) -> bool:
-        return len(word) == 2 and word[0] == word[1]
-
-    # the last char of first word and the first char of second word is tone_three
-    def _merge_continuous_three_tones_2(self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
-        new_seg = []
-        sub_finals_list = [
-            lazy_pinyin(word, neutral_tone_with_five=True, style=Style.FINALS_TONE3) for (word, pos) in seg
-        ]
-        assert len(sub_finals_list) == len(seg)
-        merge_last = [False] * len(seg)
-        for i, (word, pos) in enumerate(seg):
-            if (
-                    i - 1 >= 0
-                    and sub_finals_list[i - 1][-1][-1] == "3"
-                    and sub_finals_list[i][0][-1] == "3"
-                    and not merge_last[i - 1]
-            ):
-                # if the last word is reduplication, not merge, because reduplication need to be _neural_sandhi
-                if not self._is_reduplication(seg[i - 1][0]) and len(seg[i - 1][0]) + len(seg[i][0]) <= 3:
-                    new_seg[-1][0] = new_seg[-1][0] + seg[i][0]
-                    merge_last[i] = True
-                else:
-                    new_seg.append([word, pos])
-            else:
-                new_seg.append([word, pos])
-        return new_seg
-
-    @staticmethod
-    def _merge_er(seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
-        new_seg = []
-        for i, (word, pos) in enumerate(seg):
-            if i - 1 >= 0 and word == "儿" and seg[i - 1][0] != "#":
-                new_seg[-1][0] = new_seg[-1][0] + seg[i][0]
-            else:
-                new_seg.append([word, pos])
-        return new_seg
-
-    @staticmethod
-    def _merge_reduplication(seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
-        new_seg = []
-        for i, (word, pos) in enumerate(seg):
-            if new_seg and word == new_seg[-1][0]:
-                new_seg[-1][0] = new_seg[-1][0] + seg[i][0]
-            else:
-                new_seg.append([word, pos])
-        return new_seg
-
-    def pre_merge_for_modify(self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
-        seg = self._merge_bu(seg)
-        seg = self._merge_yi(seg)
-        seg = self._merge_reduplication(seg)
-        seg = self._merge_continuous_three_tones(seg)
-        seg = self._merge_continuous_three_tones_2(seg)
-        seg = self._merge_er(seg)
-        return seg
-
-    def modified_tone(self, word: str, pos: str, finals: List[str]) -> List[str]:
-        finals = self._bu_sandhi(word, finals)
-        finals = self._yi_sandhi(word, finals)
-        finals = self._neural_sandhi(word, pos, finals)
-        finals = self._three_sandhi(word, finals)
-        return finals
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+中文拼音变调（Tone Sandhi）自动处理器
+"""
+
+from typing import List
+from typing import Tuple
+import jieba_fast as jieba
+from pypinyin import lazy_pinyin
+from pypinyin import Style
+
+
+class ToneSandhi:
+    def __init__(self):
+        self.must_neural_tone_words = {
+            "麻烦", "麻利", "鸳鸯", "高粱", "骨头", "骆驼", "马虎", "首饰", "馒头", "馄饨",
+            "风筝", "难为", "队伍", "阔气", "闺女", "门道", "锄头", "铺盖", "铃铛", "铁匠",
+            "钥匙", "里脊", "里头", "部分", "那么", "道士", "造化", "迷糊", "连累", "这么",
+            "这个", "运气", "过去", "软和", "转悠", "踏实", "跳蚤", "跟头", "趔趄", "财主",
+            "豆腐", "讲究", "记性", "记号", "认识", "规矩", "见识", "裁缝", "补丁", "衣裳",
+            "衣服", "衙门", "街坊", "行李", "行当", "蛤蟆", "蘑菇", "薄荷", "葫芦", "葡萄",
+            "萝卜", "荸荠", "苗条", "苗头", "苍蝇", "芝麻", "舒服", "舒坦", "舌头", "自在",
+            "膏药", "脾气", "脑袋", "脊梁", "能耐", "胳膊", "胭脂", "胡萝", "胡琴", "胡同",
+            "聪明", "耽误", "耽搁", "耷拉", "耳朵", "老爷", "老实", "老婆", "老头", "老太",
+            "翻腾", "罗嗦", "罐头", "编辑", "结实", "红火", "累赘", "糨糊", "糊涂", "精神",
+            "粮食", "簸箕", "篱笆", "算计", "算盘", "答应", "笤帚", "笑语", "笑话", "窟窿",
+            "窝囊", "窗户", "稳当", "稀罕", "称呼", "秧歌", "秀气", "秀才", "福气", "祖宗",
+            "砚台", "码头", "石榴", "石头", "石匠", "知识", "眼睛", "眯缝", "眨巴", "眉毛",
+            "相声", "盘算", "白净", "痢疾", "痛快", "疟疾", "疙瘩", "疏忽", "畜生", "生意",
+            "甘蔗", "琵琶", "琢磨", "琉璃", "玻璃", "玫瑰", "玄乎", "狐狸", "状元", "特务",
+            "牲口", "牙碜", "牌楼", "爽快", "爱人", "热闹", "烧饼", "烟筒", "烂糊", "点心",
+            "炊帚", "灯笼", "火候", "漂亮", "滑溜", "溜达", "温和", "清楚", "消息", "浪头",
+            "活泼", "比方", "正经", "欺负", "模糊", "槟榔", "棺材", "棒槌", "棉花", "核桃",
+            "栅栏", "柴火", "架势", "枕头", "枇杷", "机灵", "本事", "木头", "木匠", "朋友",
+            "月饼", "月亮", "暖和", "明白", "时候", "新鲜", "故事", "收拾", "收成", "提防",
+            "挖苦", "挑剔", "指甲", "指头", "拾掇", "拳头", "拨弄", "招牌", "招呼", "抬举",
+            "护士", "折腾", "扫帚", "打量", "打算", "打点", "打扮", "打听", "打发", "扎实",
+            "扁担", "戒指", "懒得", "意识", "意思", "情形", "悟性", "怪物", "思量", "怎么",
+            "念头", "念叨", "快活", "忙活", "志气", "心思", "得罪", "张罗", "弟兄", "开通",
+            "应酬", "庄稼", "干事", "帮手", "帐篷", "希罕", "师父", "师傅", "巴结", "巴掌",
+            "差事", "工夫", "岁数", "屁股", "尾巴", "少爷", "小气", "小伙", "将就", "对头",
+            "对付", "寡妇", "家伙", "客气", "实在", "官司", "学问", "学生", "字号", "嫁妆",
+            "媳妇", "媒人", "婆家", "娘家", "委屈", "姑娘", "姐夫", "妯娌", "妥当", "妖精",
+            "奴才", "女婿", "头发", "太阳", "大爷", "大方", "大意", "大夫", "多少", "多么",
+            "外甥", "壮实", "地道", "地方", "在乎", "困难", "嘴巴", "嘱咐", "嘟囔", "嘀咕",
+            "喜欢", "喇嘛", "喇叭", "商量", "唾沫", "哑巴", "哈欠", "哆嗦", "咳嗽", "和尚",
+            "告诉", "告示", "含糊", "吓唬", "后头", "名字", "名堂", "合同", "吆喝", "叫唤",
+            "口袋", "厚道", "厉害", "千斤", "包袱", "包涵", "匀称", "勤快", "动静", "动弹",
+            "功夫", "力气", "前头", "刺猬", "刺激", "别扭", "利落", "利索", "利害", "分析",
+            "出息", "凑合", "凉快", "冷战", "冤枉", "冒失", "养活", "关系", "先生", "兄弟",
+            "便宜", "使唤", "佩服", "作坊", "体面", "位置", "似的", "伙计", "休息", "什么",
+            "人家", "亲戚", "亲家", "交情", "云彩", "事情", "买卖", "主意", "丫头", "丧气",
+            "两口", "东西", "东家", "世故", "不由", "不在", "下水", "下巴", "上头", "上司",
+            "丈夫", "丈人", "一辈", "那个", "菩萨", "父亲", "母亲", "咕噜", "邋遢", "费用",
+            "冤家", "甜头", "介绍", "荒唐", "大人", "泥鳅", "幸福", "熟悉", "计划", "扑腾",
+            "蜡烛", "姥爷", "照顾", "喉咙", "吉他", "弄堂", "蚂蚱", "凤凰", "拖沓", "寒碜",
+            "糟蹋", "倒腾", "报复", "逻辑", "盘缠", "喽啰", "牢骚", "咖喱", "扫把", "惦记",
+        }
+        self.must_not_neural_tone_words = {
+            "男子", "女子", "分子", "原子", "量子", "莲子", "石子", "瓜子", "电子", "人人",
+            "虎虎", "幺幺", "干嘛", "学子", "哈哈", "数数", "袅袅", "局地", "以下", "娃哈哈",
+            "花花草草", "留得", "耕地", "想想", "熙熙", "攘攘", "卵子", "死死", "冉冉", "恳恳",
+            "佼佼", "吵吵", "打打", "考考", "整整", "莘莘", "落地", "算子", "家家户户", "青青",
+        }
+        self.punc = "：，；。？！“”‘’':,;.?!"
+
+    # the meaning of jieba pos tag: https://blog.csdn.net/weixin_44174352/article/details/113731041
+    # e.g.
+    # word: "家里"
+    # pos: "s"
+    # finals: ['ia1', 'i3']
+    def _neural_sandhi(self, word: str, pos: str, finals: List[str]) -> List[str]:
+        # reduplication words for n. and v. e.g. 奶奶, 试试, 旺旺
+        for j, item in enumerate(word):
+            if (
+                    j - 1 >= 0
+                    and item == word[j - 1]
+                    and pos[0] in {"n", "v", "a"}
+                    and word not in self.must_not_neural_tone_words
+            ):
+                finals[j] = finals[j][:-1] + "5"
+        ge_idx = word.find("个")
+        if len(word) >= 1 and word[-1] in "吧呢哈啊呐噻嘛吖嗨呐哦哒额滴哩哟喽啰耶喔诶":
+            finals[-1] = finals[-1][:-1] + "5"
+        elif len(word) >= 1 and word[-1] in "的地得":
+            finals[-1] = finals[-1][:-1] + "5"
+        # e.g. 走了, 看着, 去过
+        elif len(word) == 1 and word in "了着过" and pos in {"ul", "uz", "ug"}:
+            finals[-1] = finals[-1][:-1] + "5"
+        elif len(word) > 1 and word[-1] in "们子" and pos in {"r", "n"} and word not in self.must_not_neural_tone_words:
+            finals[-1] = finals[-1][:-1] + "5"
+        # e.g. 桌上, 地下, 家里
+        elif len(word) > 1 and word[-1] in "上下里" and pos in {"s", "l", "f"}:
+            finals[-1] = finals[-1][:-1] + "5"
+        # e.g. 上来, 下去
+        elif len(word) > 1 and word[-1] in "来去" and word[-2] in "上下进出回过起开":
+            finals[-1] = finals[-1][:-1] + "5"
+        # 个做量词
+        elif (
+                ge_idx >= 1 and (word[ge_idx - 1].isnumeric() or word[ge_idx - 1] in "几有两半多各整每做是")
+        ) or word == "个":
+            finals[ge_idx] = finals[ge_idx][:-1] + "5"
+        else:
+            if word in self.must_neural_tone_words or word[-2:] in self.must_neural_tone_words:
+                finals[-1] = finals[-1][:-1] + "5"
+
+        word_list = self._split_word(word)
+        finals_list = [finals[: len(word_list[0])], finals[len(word_list[0]):]]
+        for i, word in enumerate(word_list):
+            # conventional neural in Chinese
+            if word in self.must_neural_tone_words or word[-2:] in self.must_neural_tone_words:
+                finals_list[i][-1] = finals_list[i][-1][:-1] + "5"
+        finals = sum(finals_list, [])
+        return finals
+
+    @staticmethod
+    def _bu_sandhi(word: str, finals: List[str]) -> List[str]:
+        # e.g. 看不懂
+        if len(word) == 3 and word[1] == "不":
+            finals[1] = finals[1][:-1] + "5"
+        else:
+            for i, char in enumerate(word):
+                # "不" before tone4 should be bu2, e.g. 不怕
+                if char == "不" and i + 1 < len(word) and finals[i + 1][-1] == "4":
+                    finals[i] = finals[i][:-1] + "2"
+        return finals
+
+    def _yi_sandhi(self, word: str, finals: List[str]) -> List[str]:
+        # "一" in number sequences, e.g. 一零零, 二一零
+        if word.find("一") != -1 and all([item.isnumeric() for item in word if item != "一"]):
+            return finals
+        # "一" between reduplication words should be yi5, e.g. 看一看
+        elif len(word) == 3 and word[1] == "一" and word[0] == word[-1]:
+            finals[1] = finals[1][:-1] + "5"
+        # when "一" is ordinal word, it should be yi1
+        elif word.startswith("第一"):
+            finals[1] = finals[1][:-1] + "1"
+        else:
+            for i, char in enumerate(word):
+                if char == "一" and i + 1 < len(word):
+                    # "一" before tone4 should be yi2, e.g. 一段
+                    if finals[i + 1][-1] == "4":
+                        finals[i] = finals[i][:-1] + "2"
+                    # "一" before non-tone4 should be yi4, e.g. 一天
+                    else:
+                        # "一" 后面如果是标点，还读一声
+                        if word[i + 1] not in self.punc:
+                            finals[i] = finals[i][:-1] + "4"
+        return finals
+
+    @staticmethod
+    def _split_word(word: str) -> List[str]:
+        word_list = jieba.cut_for_search(word)
+        word_list = sorted(word_list, key=lambda i: len(i), reverse=False)
+        first_subword = word_list[0]
+        first_begin_idx = word.find(first_subword)
+        if first_begin_idx == 0:
+            second_subword = word[len(first_subword):]
+            new_word_list = [first_subword, second_subword]
+        else:
+            second_subword = word[: -len(first_subword)]
+            new_word_list = [second_subword, first_subword]
+        return new_word_list
+
+    def _three_sandhi(self, word: str, finals: List[str]) -> List[str]:
+        if len(word) == 2 and self._all_tone_three(finals):
+            finals[0] = finals[0][:-1] + "2"
+        elif len(word) == 3:
+            word_list = self._split_word(word)
+            if self._all_tone_three(finals):
+                #  disyllabic + monosyllabic, e.g. 蒙古/包
+                if len(word_list[0]) == 2:
+                    finals[0] = finals[0][:-1] + "2"
+                    finals[1] = finals[1][:-1] + "2"
+                #  monosyllabic + disyllabic, e.g. 纸/老虎
+                elif len(word_list[0]) == 1:
+                    finals[1] = finals[1][:-1] + "2"
+            else:
+                finals_list = [finals[: len(word_list[0])], finals[len(word_list[0]):]]
+                if len(finals_list) == 2:
+                    for i, sub in enumerate(finals_list):
+                        # e.g. 所有/人
+                        if self._all_tone_three(sub) and len(sub) == 2:
+                            finals_list[i][0] = finals_list[i][0][:-1] + "2"
+                        # e.g. 好/喜欢
+                        elif (
+                                i == 1
+                                and not self._all_tone_three(sub)
+                                and finals_list[i][0][-1] == "3"
+                                and finals_list[0][-1][-1] == "3"
+                        ):
+                            finals_list[0][-1] = finals_list[0][-1][:-1] + "2"
+                        finals = sum(finals_list, [])
+        # split idiom into two words whose length is 2
+        elif len(word) == 4:
+            finals_list = [finals[:2], finals[2:]]
+            finals = []
+            for sub in finals_list:
+                if self._all_tone_three(sub):
+                    sub[0] = sub[0][:-1] + "2"
+                finals += sub
+
+        return finals
+
+    @staticmethod
+    def _all_tone_three(finals: List[str]) -> bool:
+        # 增加 len(x) > 0 的判断，防止空字符串导致崩溃
+        return all(len(x) > 0 and x[-1] == "3" for x in finals)
+
+    @staticmethod
+    def _merge_bu(seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        new_seg = []
+        last_word = ""
+        for word, pos in seg:
+            if last_word == "不":
+                word = last_word + word
+            if word != "不":
+                new_seg.append((word, pos))
+            last_word = word[:]
+        if last_word == "不":
+            new_seg.append((last_word, "d"))
+        return new_seg
+
+    @staticmethod
+    def _merge_yi(seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        new_seg = []
+        i = 0
+        # function 1
+        while i < len(seg):
+            word, pos = seg[i]
+            merged = False
+            if i - 1 >= 0 and word == "一" and i + 1 < len(seg):
+                last = new_seg[-1] if new_seg else seg[i - 1]
+                if last[0] == seg[i + 1][0] and last[1] == "v" and seg[i + 1][1] == "v":
+                    combined = last[0] + "一" + seg[i + 1][0]
+                    new_seg[-1] = [combined, last[1]]
+                    i += 2
+                    merged = True
+            if not merged:
+                new_seg.append([word, pos])
+                i += 1
+        seg = new_seg
+        new_seg = []
+        # function 2
+        for word, pos in seg:
+            if new_seg and new_seg[-1][0] == "一":
+                new_seg[-1][0] = new_seg[-1][0] + word
+            else:
+                new_seg.append([word, pos])
+        return new_seg
+
+    # the first and the second words are all_tone_three
+    def _merge_continuous_three_tones(self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        new_seg = []
+        sub_finals_list = [
+            lazy_pinyin(word, neutral_tone_with_five=True, style=Style.FINALS_TONE3) for (word, pos) in seg
+        ]
+        assert len(sub_finals_list) == len(seg)
+        merge_last = [False] * len(seg)
+        for i, (word, pos) in enumerate(seg):
+            if (
+                    i - 1 >= 0
+                    and self._all_tone_three(sub_finals_list[i - 1])
+                    and self._all_tone_three(sub_finals_list[i])
+                    and not merge_last[i - 1]
+            ):
+                # if the last word is reduplication, not merge, because reduplication need to be _neural_sandhi
+                if not self._is_reduplication(seg[i - 1][0]) and len(seg[i - 1][0]) + len(seg[i][0]) <= 3:
+                    new_seg[-1][0] = new_seg[-1][0] + seg[i][0]
+                    merge_last[i] = True
+                else:
+                    new_seg.append([word, pos])
+            else:
+                new_seg.append([word, pos])
+
+        return new_seg
+
+    @staticmethod
+    def _is_reduplication(word: str) -> bool:
+        return len(word) == 2 and word[0] == word[1]
+
+    # the last char of first word and the first char of second word is tone_three
+    def _merge_continuous_three_tones_2(self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        new_seg = []
+        sub_finals_list = [
+            lazy_pinyin(word, neutral_tone_with_five=True, style=Style.FINALS_TONE3) for (word, pos) in seg
+        ]
+        assert len(sub_finals_list) == len(seg)
+        merge_last = [False] * len(seg)
+        for i, (word, pos) in enumerate(seg):
+            if (
+                    i - 1 >= 0
+                    and sub_finals_list[i - 1][-1][-1] == "3"
+                    and sub_finals_list[i][0][-1] == "3"
+                    and not merge_last[i - 1]
+            ):
+                # if the last word is reduplication, not merge, because reduplication need to be _neural_sandhi
+                if not self._is_reduplication(seg[i - 1][0]) and len(seg[i - 1][0]) + len(seg[i][0]) <= 3:
+                    new_seg[-1][0] = new_seg[-1][0] + seg[i][0]
+                    merge_last[i] = True
+                else:
+                    new_seg.append([word, pos])
+            else:
+                new_seg.append([word, pos])
+        return new_seg
+
+    @staticmethod
+    def _merge_er(seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        new_seg = []
+        for i, (word, pos) in enumerate(seg):
+            if i - 1 >= 0 and word == "儿" and seg[i - 1][0] != "#":
+                new_seg[-1][0] = new_seg[-1][0] + seg[i][0]
+            else:
+                new_seg.append([word, pos])
+        return new_seg
+
+    @staticmethod
+    def _merge_reduplication(seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        new_seg = []
+        for i, (word, pos) in enumerate(seg):
+            if new_seg and word == new_seg[-1][0]:
+                new_seg[-1][0] = new_seg[-1][0] + seg[i][0]
+            else:
+                new_seg.append([word, pos])
+        return new_seg
+
+    def pre_merge_for_modify(self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        seg = self._merge_bu(seg)
+        seg = self._merge_yi(seg)
+        seg = self._merge_reduplication(seg)
+        seg = self._merge_continuous_three_tones(seg)
+        seg = self._merge_continuous_three_tones_2(seg)
+        seg = self._merge_er(seg)
+        return seg
+
+    def modified_tone(self, word: str, pos: str, finals: List[str]) -> List[str]:
+        finals = self._bu_sandhi(word, finals)
+        finals = self._yi_sandhi(word, finals)
+        finals = self._neural_sandhi(word, pos, finals)
+        finals = self._three_sandhi(word, finals)
+        return finals
diff --git a/genie_tts/G2P/Chinese/__pycache__/ChineseG2P.cpython-311.pyc b/genie_tts/G2P/Chinese/__pycache__/ChineseG2P.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..72ae0b8bdccc6a776abb8bbc5a0992f67269e729
Binary files /dev/null and b/genie_tts/G2P/Chinese/__pycache__/ChineseG2P.cpython-311.pyc differ
diff --git a/genie_tts/G2P/Chinese/__pycache__/CorrectPronunciation.cpython-311.pyc b/genie_tts/G2P/Chinese/__pycache__/CorrectPronunciation.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..894d795c181a32e670c41be3d08625ab0e1ba005
Binary files /dev/null and b/genie_tts/G2P/Chinese/__pycache__/CorrectPronunciation.cpython-311.pyc differ
diff --git a/genie_tts/G2P/Chinese/__pycache__/Erhua.cpython-311.pyc b/genie_tts/G2P/Chinese/__pycache__/Erhua.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3e57ad3d12df8ae6f6876a65123b45632c030bf5
Binary files /dev/null and b/genie_tts/G2P/Chinese/__pycache__/Erhua.cpython-311.pyc differ
diff --git a/genie_tts/G2P/Chinese/__pycache__/ToneSandhi.cpython-311.pyc b/genie_tts/G2P/Chinese/__pycache__/ToneSandhi.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..fb79eca42d06b99312b2d5afafb5e6ddbdef240d
Binary files /dev/null and b/genie_tts/G2P/Chinese/__pycache__/ToneSandhi.cpython-311.pyc differ
diff --git a/genie_tts/G2P/Chinese/__pycache__/__init__.cpython-311.pyc b/genie_tts/G2P/Chinese/__pycache__/__init__.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ffc32ed82ba66bf328b6478f2ec1459f10530179
Binary files /dev/null and b/genie_tts/G2P/Chinese/__pycache__/__init__.cpython-311.pyc differ
diff --git a/genie_tts/G2P/English/EnglishG2P.py b/genie_tts/G2P/English/EnglishG2P.py
index 5325385ce889018d128f9fec298de7836a0c5aeb..cad9edffdece9abce37005bc94480e8d2e2514ff 100644
--- a/genie_tts/G2P/English/EnglishG2P.py
+++ b/genie_tts/G2P/English/EnglishG2P.py
@@ -1,296 +1,296 @@
-import pickle
-import os
-import re
-from typing import List, Dict, Tuple
-
-import numpy as np
-import nltk
-from nltk.tokenize import TweetTokenizer
-from nltk import pos_tag
-
-from .Normalization import normalize
-from .WordSegment import segment_text
-from ..SymbolsV2 import symbols_v2, symbol_to_id_v2
-from ..SymbolsV2 import PUNCTUATION
-from ...Core.Resources import English_G2P_DIR
-
-# nltk 路径和分词器初始化
-nltk.data.path.append(English_G2P_DIR)
-word_tokenize = TweetTokenizer().tokenize
-
-# 路径定义
-CMU_DICT_PATH = os.path.join(English_G2P_DIR, "cmudict.rep")
-CMU_DICT_FAST_PATH = os.path.join(English_G2P_DIR, "cmudict-fast.rep")
-CMU_DICT_HOT_PATH = os.path.join(English_G2P_DIR, "engdict-hot.rep")
-CACHE_PATH = os.path.join(English_G2P_DIR, "engdict_cache.pickle")
-NAMECACHE_PATH = os.path.join(English_G2P_DIR, "namedict_cache.pickle")
-MODEL_PATH = os.path.join(English_G2P_DIR, "checkpoint20.npz")
-
-# 正则表达式和映射
-REP_MAP = {
-    "[;:：，；]": ",",
-    '["’]': "'",
-    "。": ".",
-    "！": "!",
-    "？": "?",
-}
-REP_MAP_PATTERN = re.compile("|".join(re.escape(p) for p in REP_MAP.keys()))
-PUNCTUATIONS_FOR_REGEX = "".join(re.escape(p) for p in PUNCTUATION)
-CONSECUTIVE_PUNCTUATION_PATTERN = re.compile(rf"([{PUNCTUATIONS_FOR_REGEX}\s])([{PUNCTUATIONS_FOR_REGEX}])+")
-
-
-# 辅助函数
-def _read_cmu_dict(file_path: str) -> Dict[str, List[str]]:
-    g2p_dict = {}
-    with open(file_path, 'r', encoding='utf-8') as f:
-        for line in f:
-            line = line.strip()
-            if not line or line.startswith(';;;'): continue
-            parts = re.split(r'\s+', line, maxsplit=1)
-            if len(parts) < 2: continue
-            word, pron_str = parts[0].lower(), parts[1]
-            pron = pron_str.split(" ")
-            word = re.sub(r'\(\d+\)$', '', word)
-            if word not in g2p_dict: g2p_dict[word] = [pron]
-    return g2p_dict
-
-
-def _load_and_cache_dict() -> Dict[str, List[List[str]]]:
-    with open(CACHE_PATH, "rb") as f:
-        g2p_dict = pickle.load(f)
-    hot_dict = _read_cmu_dict(CMU_DICT_HOT_PATH)
-    if hot_dict: g2p_dict.update(hot_dict)
-    return g2p_dict
-
-
-def replace_phs(phs: List[str]) -> List[str]:
-    rep_map = {"'": "-"}
-    phs_new = []
-    for ph in phs:
-        if ph in symbols_v2:
-            phs_new.append(ph)
-        elif ph in rep_map:
-            phs_new.append(rep_map[ph])
-    return phs_new
-
-
-def replace_consecutive_punctuation(text: str) -> str:
-    return CONSECUTIVE_PUNCTUATION_PATTERN.sub(r"\1", text)
-
-
-def text_normalize(text: str) -> str:
-    text = REP_MAP_PATTERN.sub(lambda x: REP_MAP[x.group()], text)
-    text = normalize(text)
-    text = replace_consecutive_punctuation(text)
-    return text
-
-
-class CleanG2p:
-    """
-    一个集成了神经网络预测功能的、独立的英文G2P转换器。
-    - 不再依赖 g2p_en 库，将模型推理逻辑直接内置。
-    - 依赖 numpy 库进行计算。
-    """
-
-    def __init__(self):
-        # 1. 初始化标准组件
-        self.cmu = _load_and_cache_dict()
-        self.namedict = self._load_name_dict()
-        for word in ["AE", "AI", "AR", "IOS", "HUD", "OS"]:
-            self.cmu.pop(word.lower(), None)
-        self._setup_homographs()
-
-        # 2. 初始化神经网络模型组件
-        self._setup_nn_components()
-        self._load_nn_model()
-
-    def _setup_nn_components(self):
-        """设置 G2P 神经网络所需的字母和音素表。"""
-        self.graphemes = ["<pad>", "<unk>", "</s>"] + list("abcdefghijklmnopqrstuvwxyz")
-        self.phonemes = ["<pad>", "<unk>", "<s>", "</s>"] + ['AA0', 'AA1', 'AA2', 'AE0', 'AE1', 'AE2', 'AH0', 'AH1',
-                                                             'AH2', 'AO0',
-                                                             'AO1', 'AO2', 'AW0', 'AW1', 'AW2', 'AY0', 'AY1', 'AY2',
-                                                             'B', 'CH', 'D', 'DH',
-                                                             'EH0', 'EH1', 'EH2', 'ER0', 'ER1', 'ER2', 'EY0', 'EY1',
-                                                             'EY2', 'F', 'G', 'HH',
-                                                             'IH0', 'IH1', 'IH2', 'IY0', 'IY1', 'IY2', 'JH', 'K', 'L',
-                                                             'M', 'N', 'NG', 'OW0', 'OW1',
-                                                             'OW2', 'OY0', 'OY1', 'OY2', 'P', 'R', 'S', 'SH', 'T', 'TH',
-                                                             'UH0', 'UH1', 'UH2', 'UW',
-                                                             'UW0', 'UW1', 'UW2', 'V', 'W', 'Y', 'Z', 'ZH']
-        self.g2idx = {g: idx for idx, g in enumerate(self.graphemes)}
-        self.idx2g = {idx: g for idx, g in enumerate(self.graphemes)}
-        self.p2idx = {p: idx for idx, p in enumerate(self.phonemes)}
-        self.idx2p = {idx: p for idx, p in enumerate(self.phonemes)}
-
-    def _load_nn_model(self):
-        """从 .npz 文件加载预训练的神经网络权重。"""
-        if not os.path.exists(MODEL_PATH):
-            raise FileNotFoundError(f"G2P model file not found at: {MODEL_PATH}. "
-                                    f"Please ensure 'checkpoint20.npz' is in the correct directory.")
-
-        variables = np.load(MODEL_PATH)
-        self.enc_emb = variables["enc_emb"]
-        self.enc_w_ih = variables["enc_w_ih"]
-        self.enc_w_hh = variables["enc_w_hh"]
-        self.enc_b_ih = variables["enc_b_ih"]
-        self.enc_b_hh = variables["enc_b_hh"]
-        self.dec_emb = variables["dec_emb"]
-        self.dec_w_ih = variables["dec_w_ih"]
-        self.dec_w_hh = variables["dec_w_hh"]
-        self.dec_b_ih = variables["dec_b_ih"]
-        self.dec_b_hh = variables["dec_b_hh"]
-        self.fc_w = variables["fc_w"]
-        self.fc_b = variables["fc_b"]
-        # logger.info("G2P neural network model loaded successfully.")
-
-    @staticmethod
-    def _sigmoid(x):
-        return 1 / (1 + np.exp(-x))
-
-    def _grucell(self, x, h, w_ih, w_hh, b_ih, b_hh):
-        rzn_ih = np.matmul(x, w_ih.T) + b_ih
-        rzn_hh = np.matmul(h, w_hh.T) + b_hh
-        rz_ih, n_ih = rzn_ih[:, :rzn_ih.shape[-1] * 2 // 3], rzn_ih[:, rzn_ih.shape[-1] * 2 // 3:]
-        rz_hh, n_hh = rzn_hh[:, :rzn_hh.shape[-1] * 2 // 3], rzn_hh[:, rzn_hh.shape[-1] * 2 // 3:]
-        rz = self._sigmoid(rz_ih + rz_hh)
-        r, z = np.split(rz, 2, -1)
-        n = np.tanh(n_ih + r * n_hh)
-        h = (1 - z) * n + z * h
-        return h
-
-    def _gru(self, x, steps, w_ih, w_hh, b_ih, b_hh, h0=None):
-        if h0 is None:
-            h0 = np.zeros((x.shape[0], w_hh.shape[1]), np.float32)
-        h = h0
-        outputs = np.zeros((x.shape[0], steps, w_hh.shape[1]), np.float32)
-        for t in range(steps):
-            h = self._grucell(x[:, t, :], h, w_ih, w_hh, b_ih, b_hh)
-            outputs[:, t, ::] = h
-        return outputs
-
-    def _encode(self, word: str) -> np.ndarray:
-        chars = list(word.lower()) + ["</s>"]
-        x = [self.g2idx.get(char, self.g2idx["<unk>"]) for char in chars]
-        x = np.take(self.enc_emb, np.expand_dims(x, 0), axis=0)
-        return x
-
-    def predict(self, word: str) -> List[str]:
-        """使用内置的神经网络模型预测单词的发音。"""
-        # Encoder
-        enc = self._encode(word)
-        enc = self._gru(enc, len(word) + 1, self.enc_w_ih, self.enc_w_hh,
-                        self.enc_b_ih, self.enc_b_hh, h0=np.zeros((1, self.enc_w_hh.shape[-1]), np.float32))
-        last_hidden = enc[:, -1, :]
-
-        # Decoder
-        dec = np.take(self.dec_emb, [self.p2idx["<s>"]], axis=0)  # Start with <s>
-        h = last_hidden
-        preds = []
-        for _ in range(20):  # Max steps
-            h = self._grucell(dec, h, self.dec_w_ih, self.dec_w_hh, self.dec_b_ih, self.dec_b_hh)
-            logits = np.matmul(h, self.fc_w.T) + self.fc_b
-            pred_idx = logits.argmax()
-            if pred_idx == self.p2idx["</s>"]: break
-            preds.append(pred_idx)
-            dec = np.take(self.dec_emb, [pred_idx], axis=0)
-
-        return [self.idx2p.get(idx, "<unk>") for idx in preds]
-
-    # --- 标准 G2P 逻辑 ---
-
-    @staticmethod
-    def _load_name_dict() -> Dict[str, List[List[str]]]:
-        if os.path.exists(NAMECACHE_PATH):
-            with open(NAMECACHE_PATH, "rb") as f: return pickle.load(f)
-        return {}
-
-    def _setup_homographs(self):
-        self.homograph2features: Dict[str, Tuple[List[str], List[str], str]] = {
-            "read": (["R", "EH1", "D"], ["R", "IY1", "D"], "VBD"),
-            "complex": (["K", "AH0", "M", "P", "L", "EH1", "K", "S"], ["K", "AA1", "M", "P", "L", "EH0", "K", "S"],
-                        "JJ"),
-            "lead": (["L", "IY1", "D"], ["L", "EH1", "D"], "NN"),
-            "presents": (["P", "R", "IY0", "Z", "EH1", "N", "T", "S"], ["P", "R", "EH1", "Z", "AH0", "N", "T", "S"],
-                         "VBZ"),
-        }
-
-    def __call__(self, text: str) -> List[str]:
-        original_words = word_tokenize(text)
-        normalized_text = text_normalize(text)
-        normalized_words = word_tokenize(normalized_text)
-
-        corrected_words = []
-        original_idx, normalized_idx = 0, 0
-        while original_idx < len(original_words) and normalized_idx < len(normalized_words):
-            if original_words[original_idx] == "I" and \
-                    " ".join(normalized_words[normalized_idx:normalized_idx + 2]) == "the first":
-                corrected_words.append("I")
-                original_idx += 1
-                normalized_idx += 2
-            else:
-                corrected_words.append(normalized_words[normalized_idx])
-                original_idx += 1
-                normalized_idx += 1
-        if normalized_idx < len(normalized_words):
-            corrected_words.extend(normalized_words[normalized_idx:])
-
-        if not corrected_words: return []
-
-        tokens = pos_tag(corrected_words)
-        prons = []
-        for o_word, pos in tokens:
-            word = o_word.lower()
-            if re.search("[a-z]", word) is None:
-                pron = [word]
-            elif word in self.homograph2features:
-                pron1, pron2, pos1 = self.homograph2features[word]
-                pron = pron1 if pos.startswith(pos1) else pron2
-            else:
-                pron = self._query_word(o_word)
-            prons.extend(pron)
-            prons.extend([" "])
-        return prons[:-1] if prons else []
-
-    def _query_word(self, o_word: str) -> List[str]:
-        word = o_word.lower()
-        if word in self.cmu:
-            if o_word == "A": return ["AH0"]
-            return self.cmu[word][0]
-        if o_word.istitle() and word in self.namedict:
-            return self.namedict[word][0]
-        if word.endswith("'s") and len(word) > 2:
-            base_pron = self._query_word(word[:-2])
-            if base_pron:
-                last_ph = base_pron[-1]
-                if last_ph in {"S", "Z", "SH", "ZH", "CH", "JH"}: return base_pron + ["AH0", "Z"]
-                if last_ph in {"P", "T", "K", "F", "TH"}: return base_pron + ["S"]
-                return base_pron + ["Z"]
-        if "-" in word and len(word) > 1:
-            parts = [p for p in word.split("-") if p]
-            if len(parts) > 1:
-                result = [ph for part in parts for ph in self._query_word(part)]
-                if result: return result
-        segments = segment_text(word)
-        if len(segments) > 1 and "".join(segments) == word:
-            result = [ph for segment in segments for ph in self._query_word(segment)]
-            if result: return result
-
-        return self.predict(o_word)
-
-
-_g2p_instance: CleanG2p = CleanG2p()
-
-
-def g2p(text: str) -> List[str]:
-    if _g2p_instance is None: raise RuntimeError("G2P model is not available.")
-    raw_phonemes = _g2p_instance(text)
-    undesired = {" ", "<pad>", "UW", "</s>", "<s>"}
-    phones = ["UNK" if ph == "<unk>" else ph for ph in raw_phonemes if ph not in undesired]
-    return replace_phs(phones)
-
-
-def english_to_phones(text: str) -> List[int]:
-    phones = g2p(text)
-    phones = [symbol_to_id_v2[ph] for ph in phones]
-    return phones
+import pickle
+import os
+import re
+from typing import List, Dict, Tuple
+
+import numpy as np
+import nltk
+from nltk.tokenize import TweetTokenizer
+from nltk import pos_tag
+
+from .Normalization import normalize
+from .WordSegment import segment_text
+from ..SymbolsV2 import symbols_v2, symbol_to_id_v2
+from ..SymbolsV2 import PUNCTUATION
+from ...Core.Resources import English_G2P_DIR
+
+# nltk 路径和分词器初始化
+nltk.data.path.append(English_G2P_DIR)
+word_tokenize = TweetTokenizer().tokenize
+
+# 路径定义
+CMU_DICT_PATH = os.path.join(English_G2P_DIR, "cmudict.rep")
+CMU_DICT_FAST_PATH = os.path.join(English_G2P_DIR, "cmudict-fast.rep")
+CMU_DICT_HOT_PATH = os.path.join(English_G2P_DIR, "engdict-hot.rep")
+CACHE_PATH = os.path.join(English_G2P_DIR, "engdict_cache.pickle")
+NAMECACHE_PATH = os.path.join(English_G2P_DIR, "namedict_cache.pickle")
+MODEL_PATH = os.path.join(English_G2P_DIR, "checkpoint20.npz")
+
+# 正则表达式和映射
+REP_MAP = {
+    "[;:：，；]": ",",
+    '["’]': "'",
+    "。": ".",
+    "！": "!",
+    "？": "?",
+}
+REP_MAP_PATTERN = re.compile("|".join(re.escape(p) for p in REP_MAP.keys()))
+PUNCTUATIONS_FOR_REGEX = "".join(re.escape(p) for p in PUNCTUATION)
+CONSECUTIVE_PUNCTUATION_PATTERN = re.compile(rf"([{PUNCTUATIONS_FOR_REGEX}\s])([{PUNCTUATIONS_FOR_REGEX}])+")
+
+
+# 辅助函数
+def _read_cmu_dict(file_path: str) -> Dict[str, List[str]]:
+    g2p_dict = {}
+    with open(file_path, 'r', encoding='utf-8') as f:
+        for line in f:
+            line = line.strip()
+            if not line or line.startswith(';;;'): continue
+            parts = re.split(r'\s+', line, maxsplit=1)
+            if len(parts) < 2: continue
+            word, pron_str = parts[0].lower(), parts[1]
+            pron = pron_str.split(" ")
+            word = re.sub(r'\(\d+\)$', '', word)
+            if word not in g2p_dict: g2p_dict[word] = [pron]
+    return g2p_dict
+
+
+def _load_and_cache_dict() -> Dict[str, List[List[str]]]:
+    with open(CACHE_PATH, "rb") as f:
+        g2p_dict = pickle.load(f)
+    hot_dict = _read_cmu_dict(CMU_DICT_HOT_PATH)
+    if hot_dict: g2p_dict.update(hot_dict)
+    return g2p_dict
+
+
+def replace_phs(phs: List[str]) -> List[str]:
+    rep_map = {"'": "-"}
+    phs_new = []
+    for ph in phs:
+        if ph in symbols_v2:
+            phs_new.append(ph)
+        elif ph in rep_map:
+            phs_new.append(rep_map[ph])
+    return phs_new
+
+
+def replace_consecutive_punctuation(text: str) -> str:
+    return CONSECUTIVE_PUNCTUATION_PATTERN.sub(r"\1", text)
+
+
+def text_normalize(text: str) -> str:
+    text = REP_MAP_PATTERN.sub(lambda x: REP_MAP[x.group()], text)
+    text = normalize(text)
+    text = replace_consecutive_punctuation(text)
+    return text
+
+
+class CleanG2p:
+    """
+    一个集成了神经网络预测功能的、独立的英文G2P转换器。
+    - 不再依赖 g2p_en 库，将模型推理逻辑直接内置。
+    - 依赖 numpy 库进行计算。
+    """
+
+    def __init__(self):
+        # 1. 初始化标准组件
+        self.cmu = _load_and_cache_dict()
+        self.namedict = self._load_name_dict()
+        for word in ["AE", "AI", "AR", "IOS", "HUD", "OS"]:
+            self.cmu.pop(word.lower(), None)
+        self._setup_homographs()
+
+        # 2. 初始化神经网络模型组件
+        self._setup_nn_components()
+        self._load_nn_model()
+
+    def _setup_nn_components(self):
+        """设置 G2P 神经网络所需的字母和音素表。"""
+        self.graphemes = ["<pad>", "<unk>", "</s>"] + list("abcdefghijklmnopqrstuvwxyz")
+        self.phonemes = ["<pad>", "<unk>", "<s>", "</s>"] + ['AA0', 'AA1', 'AA2', 'AE0', 'AE1', 'AE2', 'AH0', 'AH1',
+                                                             'AH2', 'AO0',
+                                                             'AO1', 'AO2', 'AW0', 'AW1', 'AW2', 'AY0', 'AY1', 'AY2',
+                                                             'B', 'CH', 'D', 'DH',
+                                                             'EH0', 'EH1', 'EH2', 'ER0', 'ER1', 'ER2', 'EY0', 'EY1',
+                                                             'EY2', 'F', 'G', 'HH',
+                                                             'IH0', 'IH1', 'IH2', 'IY0', 'IY1', 'IY2', 'JH', 'K', 'L',
+                                                             'M', 'N', 'NG', 'OW0', 'OW1',
+                                                             'OW2', 'OY0', 'OY1', 'OY2', 'P', 'R', 'S', 'SH', 'T', 'TH',
+                                                             'UH0', 'UH1', 'UH2', 'UW',
+                                                             'UW0', 'UW1', 'UW2', 'V', 'W', 'Y', 'Z', 'ZH']
+        self.g2idx = {g: idx for idx, g in enumerate(self.graphemes)}
+        self.idx2g = {idx: g for idx, g in enumerate(self.graphemes)}
+        self.p2idx = {p: idx for idx, p in enumerate(self.phonemes)}
+        self.idx2p = {idx: p for idx, p in enumerate(self.phonemes)}
+
+    def _load_nn_model(self):
+        """从 .npz 文件加载预训练的神经网络权重。"""
+        if not os.path.exists(MODEL_PATH):
+            raise FileNotFoundError(f"G2P model file not found at: {MODEL_PATH}. "
+                                    f"Please ensure 'checkpoint20.npz' is in the correct directory.")
+
+        variables = np.load(MODEL_PATH)
+        self.enc_emb = variables["enc_emb"]
+        self.enc_w_ih = variables["enc_w_ih"]
+        self.enc_w_hh = variables["enc_w_hh"]
+        self.enc_b_ih = variables["enc_b_ih"]
+        self.enc_b_hh = variables["enc_b_hh"]
+        self.dec_emb = variables["dec_emb"]
+        self.dec_w_ih = variables["dec_w_ih"]
+        self.dec_w_hh = variables["dec_w_hh"]
+        self.dec_b_ih = variables["dec_b_ih"]
+        self.dec_b_hh = variables["dec_b_hh"]
+        self.fc_w = variables["fc_w"]
+        self.fc_b = variables["fc_b"]
+        # logger.info("G2P neural network model loaded successfully.")
+
+    @staticmethod
+    def _sigmoid(x):
+        return 1 / (1 + np.exp(-x))
+
+    def _grucell(self, x, h, w_ih, w_hh, b_ih, b_hh):
+        rzn_ih = np.matmul(x, w_ih.T) + b_ih
+        rzn_hh = np.matmul(h, w_hh.T) + b_hh
+        rz_ih, n_ih = rzn_ih[:, :rzn_ih.shape[-1] * 2 // 3], rzn_ih[:, rzn_ih.shape[-1] * 2 // 3:]
+        rz_hh, n_hh = rzn_hh[:, :rzn_hh.shape[-1] * 2 // 3], rzn_hh[:, rzn_hh.shape[-1] * 2 // 3:]
+        rz = self._sigmoid(rz_ih + rz_hh)
+        r, z = np.split(rz, 2, -1)
+        n = np.tanh(n_ih + r * n_hh)
+        h = (1 - z) * n + z * h
+        return h
+
+    def _gru(self, x, steps, w_ih, w_hh, b_ih, b_hh, h0=None):
+        if h0 is None:
+            h0 = np.zeros((x.shape[0], w_hh.shape[1]), np.float32)
+        h = h0
+        outputs = np.zeros((x.shape[0], steps, w_hh.shape[1]), np.float32)
+        for t in range(steps):
+            h = self._grucell(x[:, t, :], h, w_ih, w_hh, b_ih, b_hh)
+            outputs[:, t, ::] = h
+        return outputs
+
+    def _encode(self, word: str) -> np.ndarray:
+        chars = list(word.lower()) + ["</s>"]
+        x = [self.g2idx.get(char, self.g2idx["<unk>"]) for char in chars]
+        x = np.take(self.enc_emb, np.expand_dims(x, 0), axis=0)
+        return x
+
+    def predict(self, word: str) -> List[str]:
+        """使用内置的神经网络模型预测单词的发音。"""
+        # Encoder
+        enc = self._encode(word)
+        enc = self._gru(enc, len(word) + 1, self.enc_w_ih, self.enc_w_hh,
+                        self.enc_b_ih, self.enc_b_hh, h0=np.zeros((1, self.enc_w_hh.shape[-1]), np.float32))
+        last_hidden = enc[:, -1, :]
+
+        # Decoder
+        dec = np.take(self.dec_emb, [self.p2idx["<s>"]], axis=0)  # Start with <s>
+        h = last_hidden
+        preds = []
+        for _ in range(20):  # Max steps
+            h = self._grucell(dec, h, self.dec_w_ih, self.dec_w_hh, self.dec_b_ih, self.dec_b_hh)
+            logits = np.matmul(h, self.fc_w.T) + self.fc_b
+            pred_idx = logits.argmax()
+            if pred_idx == self.p2idx["</s>"]: break
+            preds.append(pred_idx)
+            dec = np.take(self.dec_emb, [pred_idx], axis=0)
+
+        return [self.idx2p.get(idx, "<unk>") for idx in preds]
+
+    # --- 标准 G2P 逻辑 ---
+
+    @staticmethod
+    def _load_name_dict() -> Dict[str, List[List[str]]]:
+        if os.path.exists(NAMECACHE_PATH):
+            with open(NAMECACHE_PATH, "rb") as f: return pickle.load(f)
+        return {}
+
+    def _setup_homographs(self):
+        self.homograph2features: Dict[str, Tuple[List[str], List[str], str]] = {
+            "read": (["R", "EH1", "D"], ["R", "IY1", "D"], "VBD"),
+            "complex": (["K", "AH0", "M", "P", "L", "EH1", "K", "S"], ["K", "AA1", "M", "P", "L", "EH0", "K", "S"],
+                        "JJ"),
+            "lead": (["L", "IY1", "D"], ["L", "EH1", "D"], "NN"),
+            "presents": (["P", "R", "IY0", "Z", "EH1", "N", "T", "S"], ["P", "R", "EH1", "Z", "AH0", "N", "T", "S"],
+                         "VBZ"),
+        }
+
+    def __call__(self, text: str) -> List[str]:
+        original_words = word_tokenize(text)
+        normalized_text = text_normalize(text)
+        normalized_words = word_tokenize(normalized_text)
+
+        corrected_words = []
+        original_idx, normalized_idx = 0, 0
+        while original_idx < len(original_words) and normalized_idx < len(normalized_words):
+            if original_words[original_idx] == "I" and \
+                    " ".join(normalized_words[normalized_idx:normalized_idx + 2]) == "the first":
+                corrected_words.append("I")
+                original_idx += 1
+                normalized_idx += 2
+            else:
+                corrected_words.append(normalized_words[normalized_idx])
+                original_idx += 1
+                normalized_idx += 1
+        if normalized_idx < len(normalized_words):
+            corrected_words.extend(normalized_words[normalized_idx:])
+
+        if not corrected_words: return []
+
+        tokens = pos_tag(corrected_words)
+        prons = []
+        for o_word, pos in tokens:
+            word = o_word.lower()
+            if re.search("[a-z]", word) is None:
+                pron = [word]
+            elif word in self.homograph2features:
+                pron1, pron2, pos1 = self.homograph2features[word]
+                pron = pron1 if pos.startswith(pos1) else pron2
+            else:
+                pron = self._query_word(o_word)
+            prons.extend(pron)
+            prons.extend([" "])
+        return prons[:-1] if prons else []
+
+    def _query_word(self, o_word: str) -> List[str]:
+        word = o_word.lower()
+        if word in self.cmu:
+            if o_word == "A": return ["AH0"]
+            return self.cmu[word][0]
+        if o_word.istitle() and word in self.namedict:
+            return self.namedict[word][0]
+        if word.endswith("'s") and len(word) > 2:
+            base_pron = self._query_word(word[:-2])
+            if base_pron:
+                last_ph = base_pron[-1]
+                if last_ph in {"S", "Z", "SH", "ZH", "CH", "JH"}: return base_pron + ["AH0", "Z"]
+                if last_ph in {"P", "T", "K", "F", "TH"}: return base_pron + ["S"]
+                return base_pron + ["Z"]
+        if "-" in word and len(word) > 1:
+            parts = [p for p in word.split("-") if p]
+            if len(parts) > 1:
+                result = [ph for part in parts for ph in self._query_word(part)]
+                if result: return result
+        segments = segment_text(word)
+        if len(segments) > 1 and "".join(segments) == word:
+            result = [ph for segment in segments for ph in self._query_word(segment)]
+            if result: return result
+
+        return self.predict(o_word)
+
+
+_g2p_instance: CleanG2p = CleanG2p()
+
+
+def g2p(text: str) -> List[str]:
+    if _g2p_instance is None: raise RuntimeError("G2P model is not available.")
+    raw_phonemes = _g2p_instance(text)
+    undesired = {" ", "<pad>", "UW", "</s>", "<s>"}
+    phones = ["UNK" if ph == "<unk>" else ph for ph in raw_phonemes if ph not in undesired]
+    return replace_phs(phones)
+
+
+def english_to_phones(text: str) -> List[int]:
+    phones = g2p(text)
+    phones = [symbol_to_id_v2[ph] for ph in phones]
+    return phones
diff --git a/genie_tts/G2P/English/Normalization.py b/genie_tts/G2P/English/Normalization.py
index ce5046d73caeaf8750e67a6088012d13d777737a..513a1ec770ad18ffb38479e9ee564acc32c5fed8 100644
--- a/genie_tts/G2P/English/Normalization.py
+++ b/genie_tts/G2P/English/Normalization.py
@@ -1,286 +1,286 @@
-import re
-import unicodedata
-from calendar import month_name
-
-
-# ------------------- 核心：自实现数字转单词 (替代 inflect) -------------------
-
-def _number_to_words_custom(num_str):
-    """一个不依赖inflect的、简化的数字到单词转换器。"""
-    num_str = str(num_str).strip()
-    if not num_str.isdigit(): return num_str
-
-    num = int(num_str)
-    if num == 0: return 'zero'
-
-    units = ["", "one", "two", "three", "four", "five", "six", "seven", "eight", "nine"]
-    teens = ["ten", "eleven", "twelve", "thirteen", "fourteen", "fifteen", "sixteen", "seventeen", "eighteen",
-             "nineteen"]
-    tens = ["", "", "twenty", "thirty", "forty", "fifty", "sixty", "seventy", "eighty", "ninety"]
-    thousands = ["", "thousand", "million", "billion", "trillion"]
-
-    def convert_less_than_thousand(n):
-        if n == 0: return ""
-        if n < 10: return units[n]
-        if n < 20: return teens[n - 10]
-        if n < 100: return tens[n // 10] + (" " + units[n % 10] if n % 10 != 0 else "")
-        return units[n // 100] + " hundred" + (" " + convert_less_than_thousand(n % 100) if n % 100 != 0 else "")
-
-    words = []
-    i = 0
-    if num == 0: return "zero"
-    while num > 0:
-        if num % 1000 != 0:
-            words.insert(0, convert_less_than_thousand(num % 1000) + " " + thousands[i])
-        num //= 1000
-        i += 1
-    return " ".join(words).strip()
-
-
-def _ordinal_custom(num_str):
-    """一个不依赖inflect的、简化的序数词转换器。"""
-    num = int(num_str)
-    if 10 <= num % 100 <= 20:
-        suffix = 'th'
-    else:
-        suffix = {1: 'st', 2: 'nd', 3: 'rd'}.get(num % 10, 'th')
-    return _number_to_words_custom(str(num)) + suffix
-
-
-# ------------------- 初始化和常量定义 (无 inflect) -------------------
-
-_measurement_map = {
-    "km/h": ["kilometer per hour", "kilometers per hour"], "mph": ["mile per hour", "miles per hour"],
-    "°C": ["degree celsius", "degrees celsius"], "°F": ["degree fahrenheit", "degrees fahrenheit"],
-    "tbsp": ["tablespoon", "tablespoons"], "tsp": ["teaspoon", "teaspoons"],
-    "km": ["kilometer", "kilometers"], "kg": ["kilogram", "kilograms"], "min": ["minute", "minutes"],
-    "ft": ["foot", "feet"], "cm": ["centimeter", "centimeters"], "m": ["meter", "meters"],
-    "L": ["liter", "liters"], "h": ["hour", "hours"], "s": ["second", "seconds"],
-}
-
-_abbreviations = [
-    (re.compile(r"\bMr\.(?=[\s,.]|\Z)", re.IGNORECASE), "Mister"),
-    (re.compile(r"\bMrs\.(?=[\s,.]|\Z)", re.IGNORECASE), "Missus"),
-    (re.compile(r"\bDr\.(?=[\s,.]|\Z)", re.IGNORECASE), "Doctor"),
-    (re.compile(r"\bProf\.(?=[\s,.]|\Z)", re.IGNORECASE), "Professor"),
-    (re.compile(r"\bSt\.(?=[\s,.]|\Z)", re.IGNORECASE), "Street"),
-    (re.compile(r"\bCo\.(?=[\s,.]|\Z)", re.IGNORECASE), "Company"),
-    (re.compile(r"\bLtd\.(?=[\s,.]|\Z)", re.IGNORECASE), "Limited"),
-    (re.compile(r"\be\.g\.(?=[\s,.]|\Z)", re.IGNORECASE), "for example"),
-    (re.compile(r"\bi\.e\.(?=[\s,.]|\Z)", re.IGNORECASE), "that is"),
-]
-
-# ------------------- 正则表达式定义 (与原来保持一致) -------------------
-_currency_suffix_re = re.compile(r"([£$€])([\d,.]*\d)\s*(million|billion|thousand)\b", re.IGNORECASE)
-_phone_re = re.compile(r"(\+?\d{1,3}-)?\b(\d{3})-(?:(\d{3})-)?(\d{4})\b")
-_roman_re = re.compile(r"\b(XIX|XVIII|XVII|XVI|XV|XIV|XIII|XII|XI|X|IX|VIII|VII|VI|V|IV|III|II)\b", re.IGNORECASE)
-_decade_re = re.compile(r"\b((?:1[89]|20)\d0)s\b")
-_score_re = re.compile(r"\b(\d{1,2})-(\d{1,2})\b")
-_dimension_re = re.compile(r"\b(\d+(?:\.\d+)?)\s*x\s*(\d+(?:\.\d+)?)(?:\s*x\s*(\d+(?:\.\d+)?))?\b")
-_alphanumeric_re = re.compile(r"\b([a-zA-Z]+[0-9]+|[0-9]+[a-zA-Z]+)\b")
-_date_re = re.compile(r"\b(0?[1-9]|1[0-2])/([0-2]?\d|3[01])/(\d{2,4})\b")
-_ordinal_number_re = re.compile(r"\b(\d+)\. ")
-_comma_number_re = re.compile(r"(\d[\d,]+\d)")
-_currency_re = re.compile(r"([£$€])(\d*\.?\d+)|(\d*\.?\d+)\s*([£$€])")
-_time_re = re.compile(r"\b([01]?\d|2[0-3]):([0-5]\d)(?::([0-5]\d))?(\s*(?:a\.?m\.?|p\.?m\.?))?\b", re.IGNORECASE)
-units = "|".join(re.escape(key) for key in sorted(_measurement_map.keys(), key=len, reverse=True))
-_measurement_re = re.compile(rf"(?<!\w)(-?(?:\d+/\d+|\d+(?:\.\d+)?))\s*({units})\b")
-_fraction_re = re.compile(r"\b(\d+)/(\d+)\b")
-_decimal_number_re = re.compile(r"(\d+\.\d+)")
-_ordinal_re = re.compile(r"\b\d+(st|nd|rd|th)\b")
-_acronym_re = re.compile(r"\b[A-Z]{2,}\b")
-_number_re = re.compile(r"(?<!\w)-?\d+(?!\w)")
-
-
-# ------------------- 替换与扩展函数 (全部使用 _number_to_words_custom) -------------------
-def _expand_currency_suffix(m):
-    symbol, amount_str, suffix = m.groups()
-    major_map = {"$": "dollars", "£": "pounds", "€": "euros"}
-    amount_word = _number_to_words_custom(amount_str.replace(",", ""))
-    currency_word = major_map.get(symbol, "")
-    return f"{amount_word} {suffix} {currency_word}"
-
-
-def _expand_phone_number(m):
-    country, area, exch, line = m.groups()
-    parts = []
-    if country:
-        country_words = []
-        if country.startswith('+'): country_words.append('plus')
-        digits_only = re.sub(r'\D', '', country)
-        if digits_only: country_words.append(' '.join(_number_to_words_custom(d) for d in digits_only))
-        parts.append(' '.join(country_words))
-    parts.append(' '.join(_number_to_words_custom(c) for c in area))
-    if exch: parts.append(' '.join(_number_to_words_custom(c) for c in exch))
-    parts.append(' '.join(_number_to_words_custom(c) for c in line))
-    return ", ".join(parts)
-
-
-def _expand_roman(m):
-    roman_map = {
-        "ii": "two", "iii": "three", "iv": "four", "v": "five", "vi": "six", "vii": "seven",
-        "viii": "eight", "ix": "nine", "x": "ten", "xi": "eleven", "xii": "twelve",
-        "xiii": "thirteen", "xiv": "fourteen", "xv": "fifteen", "xvi": "sixteen",
-        "xvii": "seventeen", "xviii": "eighteen", "xix": "nineteen"
-    }
-    return roman_map.get(m.group(1).lower(), m.group(1))
-
-
-def _expand_decade(m):
-    year_str = m.group(1)
-    year_words = _expand_number_positive(year_str)
-    if year_words.endswith('ty'): return f"{year_words[:-1]}ies"
-    return f"{year_words}s"
-
-
-def _expand_dimension(m):
-    parts = [p for p in m.groups() if p is not None]
-    expanded_parts = [_number_to_words_custom(p) for p in parts]
-    return " by ".join(expanded_parts)
-
-
-def _expand_score(m):
-    return f"{_number_to_words_custom(m.group(1))} to {_number_to_words_custom(m.group(2))}"
-
-
-def _expand_alphanumeric(m):
-    text = m.group(0)
-    parts = re.findall(r'[a-zA-Z]+|[0-9]+', text)
-    expanded_parts = []
-    for part in parts:
-        if part.isalpha():
-            expanded_parts.append(' '.join(list(part)))
-        elif part.isdigit():
-            expanded_parts.append(' '.join(_number_to_words_custom(c) for c in part))
-    return ' '.join(expanded_parts)
-
-
-def _convert_ordinal(m):
-    return _ordinal_custom(m.group(1)) + ", "
-
-
-def _remove_commas(m): return m.group(1).replace(",", "")
-
-
-def _expand_time(m):
-    h_str, m_str, s_str, am_pm = m.groups()
-    h, m = int(h_str), int(m_str)
-    h_word = _number_to_words_custom(h if h <= 12 or not am_pm else h - 12)
-    if h == 0 and am_pm: h_word = "twelve"
-    m_word = ""
-    if m > 0: m_word = f" oh {_number_to_words_custom(m)}" if m < 10 else f" {_number_to_words_custom(m)}"
-    result = f"{h_word}{m_word}".lstrip()
-    if s_str: result += f" and {_number_to_words_custom(int(s_str))} seconds"
-    if am_pm: result += ' pm' if 'p' in am_pm.lower() else ' am'
-    return result
-
-
-def _expand_measurement(m):
-    num_str, unit = m.groups()
-    is_neg = num_str.startswith('-')
-    if is_neg: num_str = num_str[1:]
-    if '/' in num_str:
-        num_word = _expand_fraction(re.match(_fraction_re, num_str))
-        is_plural = True
-    else:
-        num_word = _number_to_words_custom(num_str)
-        is_plural = float(num_str) != 1
-    unit_word = _measurement_map[unit][1] if is_plural else _measurement_map[unit][0]
-    result = f"{num_word} {unit_word}"
-    return f"minus {result}" if is_neg else result
-
-
-def _expand_currency(m):
-    symbol, amount_str = (m.group(1), m.group(2)) if m.group(1) else (m.group(4), m.group(3))
-    amount_str = (amount_str or "").replace(",", "")
-    if amount_str.startswith('.'): amount_str = '0' + amount_str
-    major_map = {"$": ("dollar", "dollars"), "£": ("pound", "pounds"), "€": ("euro", "euros")}
-    minor_map = {"$": ("cent", "cents"), "£": ("penny", "pence"), "€": ("cent", "cents")}
-    major_singular, major_plural = major_map.get(symbol, ("", ""))
-    parts = amount_str.split('.')
-    major_val = int(parts[0]) if parts[0] else 0
-    minor_val = int(parts[1].ljust(2, '0')) if len(parts) > 1 and parts[1] else 0
-    result = []
-    if major_val > 0:
-        result.append(f"{_number_to_words_custom(major_val)} {major_singular if major_val == 1 else major_plural}")
-    if minor_val > 0:
-        minor_singular, minor_plural = minor_map.get(symbol, ("", ""))
-        result.append(f"{_number_to_words_custom(minor_val)} {minor_singular if minor_val == 1 else minor_plural}")
-    return " and ".join(result) or f"zero {major_plural}"
-
-
-def _expand_decimal_number(m):
-    num_str = m.group(1)
-    parts = num_str.split('.')
-    integer_part = _number_to_words_custom(parts[0])
-    fractional_part = ' '.join(_number_to_words_custom(digit) for digit in parts[1])
-    return f"{integer_part} point {fractional_part}"
-
-
-def _expand_date(m):
-    month, day, year = m.groups()
-    month_word = month_name[int(month)]
-    day_word = _ordinal_custom(day)
-    year_num = int(year)
-    if len(year) == 2: year_num += 2000 if year_num < 50 else 1900
-    return f"{month_word} {day_word}, {_expand_number_positive(str(year_num))}"
-
-
-def _expand_fraction(m):
-    n, d = int(m.group(1)), int(m.group(2))
-    if d == 0: return m.group(0)
-    common_fractions = {(1, 2): "one half", (1, 4): "one quarter", (3, 4): "three quarters"}
-    if (n, d) in common_fractions: return common_fractions[(n, d)]
-    return f"{_number_to_words_custom(n)} over {_number_to_words_custom(d)}"
-
-
-def _expand_ordinal_word(m):
-    return _ordinal_custom(m.group(0)[:-2])
-
-
-def _expand_number(m):
-    num_str = m.group(0)
-    if num_str.startswith('-'): return f"minus {_expand_number_positive(num_str[1:])}"
-    return _expand_number_positive(num_str)
-
-
-def _expand_number_positive(num_str):
-    num = int(num_str)
-    if 2000 <= num < 2010: return f"two thousand and {_number_to_words_custom(num % 100)}"
-    if 1100 <= num < 2100 and num % 100 != 0:
-        return f"{_number_to_words_custom(num // 100)} {_number_to_words_custom(num % 100)}"
-    return _number_to_words_custom(num_str)
-
-
-def _expand_acronym(m): return " ".join(m.group(0))
-
-
-def normalize(text):
-    text = "".join(char for char in unicodedata.normalize("NFD", text) if unicodedata.category(char) != "Mn")
-    text = re.sub(r"@", " at ", text)
-    for regex, replacement in _abbreviations: text = regex.sub(replacement, text)
-    text = re.sub(_currency_suffix_re, _expand_currency_suffix, text)
-    text = re.sub(_phone_re, _expand_phone_number, text)
-    text = re.sub(_dimension_re, _expand_dimension, text)
-    text = re.sub(_roman_re, _expand_roman, text)
-    text = re.sub(_decade_re, _expand_decade, text)
-    text = re.sub(_score_re, _expand_score, text)
-    text = re.sub(_date_re, _expand_date, text)
-    text = re.sub(_time_re, _expand_time, text)
-    text = re.sub(_ordinal_number_re, _convert_ordinal, text)
-    text = re.sub(_comma_number_re, _remove_commas, text)
-    text = re.sub(_currency_re, _expand_currency, text)
-    text = re.sub(_measurement_re, _expand_measurement, text)
-    text = re.sub(_fraction_re, _expand_fraction, text)
-    text = re.sub(_decimal_number_re, _expand_decimal_number, text)
-    text = re.sub(_ordinal_re, _expand_ordinal_word, text)
-    text = re.sub(_alphanumeric_re, _expand_alphanumeric, text)
-    text = re.sub(_acronym_re, _expand_acronym, text)
-    text = re.sub(_number_re, _expand_number, text)
-    text = text.lower()
-    text = re.sub(r"%", " percent", text)
-    domain_re = re.compile(r'\b([a-z0-9-]+)\.([a-z]{2,})\b')
-    while domain_re.search(text): text = domain_re.sub(r'\1 dot \2', text)
-    text = re.sub(r"[^a-z0-9'.,?!:;-]", " ", text)
-    text = re.sub(r"\s+", " ", text)
-    return text.strip()
+import re
+import unicodedata
+from calendar import month_name
+
+
+# ------------------- 核心：自实现数字转单词 (替代 inflect) -------------------
+
+def _number_to_words_custom(num_str):
+    """一个不依赖inflect的、简化的数字到单词转换器。"""
+    num_str = str(num_str).strip()
+    if not num_str.isdigit(): return num_str
+
+    num = int(num_str)
+    if num == 0: return 'zero'
+
+    units = ["", "one", "two", "three", "four", "five", "six", "seven", "eight", "nine"]
+    teens = ["ten", "eleven", "twelve", "thirteen", "fourteen", "fifteen", "sixteen", "seventeen", "eighteen",
+             "nineteen"]
+    tens = ["", "", "twenty", "thirty", "forty", "fifty", "sixty", "seventy", "eighty", "ninety"]
+    thousands = ["", "thousand", "million", "billion", "trillion"]
+
+    def convert_less_than_thousand(n):
+        if n == 0: return ""
+        if n < 10: return units[n]
+        if n < 20: return teens[n - 10]
+        if n < 100: return tens[n // 10] + (" " + units[n % 10] if n % 10 != 0 else "")
+        return units[n // 100] + " hundred" + (" " + convert_less_than_thousand(n % 100) if n % 100 != 0 else "")
+
+    words = []
+    i = 0
+    if num == 0: return "zero"
+    while num > 0:
+        if num % 1000 != 0:
+            words.insert(0, convert_less_than_thousand(num % 1000) + " " + thousands[i])
+        num //= 1000
+        i += 1
+    return " ".join(words).strip()
+
+
+def _ordinal_custom(num_str):
+    """一个不依赖inflect的、简化的序数词转换器。"""
+    num = int(num_str)
+    if 10 <= num % 100 <= 20:
+        suffix = 'th'
+    else:
+        suffix = {1: 'st', 2: 'nd', 3: 'rd'}.get(num % 10, 'th')
+    return _number_to_words_custom(str(num)) + suffix
+
+
+# ------------------- 初始化和常量定义 (无 inflect) -------------------
+
+_measurement_map = {
+    "km/h": ["kilometer per hour", "kilometers per hour"], "mph": ["mile per hour", "miles per hour"],
+    "°C": ["degree celsius", "degrees celsius"], "°F": ["degree fahrenheit", "degrees fahrenheit"],
+    "tbsp": ["tablespoon", "tablespoons"], "tsp": ["teaspoon", "teaspoons"],
+    "km": ["kilometer", "kilometers"], "kg": ["kilogram", "kilograms"], "min": ["minute", "minutes"],
+    "ft": ["foot", "feet"], "cm": ["centimeter", "centimeters"], "m": ["meter", "meters"],
+    "L": ["liter", "liters"], "h": ["hour", "hours"], "s": ["second", "seconds"],
+}
+
+_abbreviations = [
+    (re.compile(r"\bMr\.(?=[\s,.]|\Z)", re.IGNORECASE), "Mister"),
+    (re.compile(r"\bMrs\.(?=[\s,.]|\Z)", re.IGNORECASE), "Missus"),
+    (re.compile(r"\bDr\.(?=[\s,.]|\Z)", re.IGNORECASE), "Doctor"),
+    (re.compile(r"\bProf\.(?=[\s,.]|\Z)", re.IGNORECASE), "Professor"),
+    (re.compile(r"\bSt\.(?=[\s,.]|\Z)", re.IGNORECASE), "Street"),
+    (re.compile(r"\bCo\.(?=[\s,.]|\Z)", re.IGNORECASE), "Company"),
+    (re.compile(r"\bLtd\.(?=[\s,.]|\Z)", re.IGNORECASE), "Limited"),
+    (re.compile(r"\be\.g\.(?=[\s,.]|\Z)", re.IGNORECASE), "for example"),
+    (re.compile(r"\bi\.e\.(?=[\s,.]|\Z)", re.IGNORECASE), "that is"),
+]
+
+# ------------------- 正则表达式定义 (与原来保持一致) -------------------
+_currency_suffix_re = re.compile(r"([£$€])([\d,.]*\d)\s*(million|billion|thousand)\b", re.IGNORECASE)
+_phone_re = re.compile(r"(\+?\d{1,3}-)?\b(\d{3})-(?:(\d{3})-)?(\d{4})\b")
+_roman_re = re.compile(r"\b(XIX|XVIII|XVII|XVI|XV|XIV|XIII|XII|XI|X|IX|VIII|VII|VI|V|IV|III|II)\b", re.IGNORECASE)
+_decade_re = re.compile(r"\b((?:1[89]|20)\d0)s\b")
+_score_re = re.compile(r"\b(\d{1,2})-(\d{1,2})\b")
+_dimension_re = re.compile(r"\b(\d+(?:\.\d+)?)\s*x\s*(\d+(?:\.\d+)?)(?:\s*x\s*(\d+(?:\.\d+)?))?\b")
+_alphanumeric_re = re.compile(r"\b([a-zA-Z]+[0-9]+|[0-9]+[a-zA-Z]+)\b")
+_date_re = re.compile(r"\b(0?[1-9]|1[0-2])/([0-2]?\d|3[01])/(\d{2,4})\b")
+_ordinal_number_re = re.compile(r"\b(\d+)\. ")
+_comma_number_re = re.compile(r"(\d[\d,]+\d)")
+_currency_re = re.compile(r"([£$€])(\d*\.?\d+)|(\d*\.?\d+)\s*([£$€])")
+_time_re = re.compile(r"\b([01]?\d|2[0-3]):([0-5]\d)(?::([0-5]\d))?(\s*(?:a\.?m\.?|p\.?m\.?))?\b", re.IGNORECASE)
+units = "|".join(re.escape(key) for key in sorted(_measurement_map.keys(), key=len, reverse=True))
+_measurement_re = re.compile(rf"(?<!\w)(-?(?:\d+/\d+|\d+(?:\.\d+)?))\s*({units})\b")
+_fraction_re = re.compile(r"\b(\d+)/(\d+)\b")
+_decimal_number_re = re.compile(r"(\d+\.\d+)")
+_ordinal_re = re.compile(r"\b\d+(st|nd|rd|th)\b")
+_acronym_re = re.compile(r"\b[A-Z]{2,}\b")
+_number_re = re.compile(r"(?<!\w)-?\d+(?!\w)")
+
+
+# ------------------- 替换与扩展函数 (全部使用 _number_to_words_custom) -------------------
+def _expand_currency_suffix(m):
+    symbol, amount_str, suffix = m.groups()
+    major_map = {"$": "dollars", "£": "pounds", "€": "euros"}
+    amount_word = _number_to_words_custom(amount_str.replace(",", ""))
+    currency_word = major_map.get(symbol, "")
+    return f"{amount_word} {suffix} {currency_word}"
+
+
+def _expand_phone_number(m):
+    country, area, exch, line = m.groups()
+    parts = []
+    if country:
+        country_words = []
+        if country.startswith('+'): country_words.append('plus')
+        digits_only = re.sub(r'\D', '', country)
+        if digits_only: country_words.append(' '.join(_number_to_words_custom(d) for d in digits_only))
+        parts.append(' '.join(country_words))
+    parts.append(' '.join(_number_to_words_custom(c) for c in area))
+    if exch: parts.append(' '.join(_number_to_words_custom(c) for c in exch))
+    parts.append(' '.join(_number_to_words_custom(c) for c in line))
+    return ", ".join(parts)
+
+
+def _expand_roman(m):
+    roman_map = {
+        "ii": "two", "iii": "three", "iv": "four", "v": "five", "vi": "six", "vii": "seven",
+        "viii": "eight", "ix": "nine", "x": "ten", "xi": "eleven", "xii": "twelve",
+        "xiii": "thirteen", "xiv": "fourteen", "xv": "fifteen", "xvi": "sixteen",
+        "xvii": "seventeen", "xviii": "eighteen", "xix": "nineteen"
+    }
+    return roman_map.get(m.group(1).lower(), m.group(1))
+
+
+def _expand_decade(m):
+    year_str = m.group(1)
+    year_words = _expand_number_positive(year_str)
+    if year_words.endswith('ty'): return f"{year_words[:-1]}ies"
+    return f"{year_words}s"
+
+
+def _expand_dimension(m):
+    parts = [p for p in m.groups() if p is not None]
+    expanded_parts = [_number_to_words_custom(p) for p in parts]
+    return " by ".join(expanded_parts)
+
+
+def _expand_score(m):
+    return f"{_number_to_words_custom(m.group(1))} to {_number_to_words_custom(m.group(2))}"
+
+
+def _expand_alphanumeric(m):
+    text = m.group(0)
+    parts = re.findall(r'[a-zA-Z]+|[0-9]+', text)
+    expanded_parts = []
+    for part in parts:
+        if part.isalpha():
+            expanded_parts.append(' '.join(list(part)))
+        elif part.isdigit():
+            expanded_parts.append(' '.join(_number_to_words_custom(c) for c in part))
+    return ' '.join(expanded_parts)
+
+
+def _convert_ordinal(m):
+    return _ordinal_custom(m.group(1)) + ", "
+
+
+def _remove_commas(m): return m.group(1).replace(",", "")
+
+
+def _expand_time(m):
+    h_str, m_str, s_str, am_pm = m.groups()
+    h, m = int(h_str), int(m_str)
+    h_word = _number_to_words_custom(h if h <= 12 or not am_pm else h - 12)
+    if h == 0 and am_pm: h_word = "twelve"
+    m_word = ""
+    if m > 0: m_word = f" oh {_number_to_words_custom(m)}" if m < 10 else f" {_number_to_words_custom(m)}"
+    result = f"{h_word}{m_word}".lstrip()
+    if s_str: result += f" and {_number_to_words_custom(int(s_str))} seconds"
+    if am_pm: result += ' pm' if 'p' in am_pm.lower() else ' am'
+    return result
+
+
+def _expand_measurement(m):
+    num_str, unit = m.groups()
+    is_neg = num_str.startswith('-')
+    if is_neg: num_str = num_str[1:]
+    if '/' in num_str:
+        num_word = _expand_fraction(re.match(_fraction_re, num_str))
+        is_plural = True
+    else:
+        num_word = _number_to_words_custom(num_str)
+        is_plural = float(num_str) != 1
+    unit_word = _measurement_map[unit][1] if is_plural else _measurement_map[unit][0]
+    result = f"{num_word} {unit_word}"
+    return f"minus {result}" if is_neg else result
+
+
+def _expand_currency(m):
+    symbol, amount_str = (m.group(1), m.group(2)) if m.group(1) else (m.group(4), m.group(3))
+    amount_str = (amount_str or "").replace(",", "")
+    if amount_str.startswith('.'): amount_str = '0' + amount_str
+    major_map = {"$": ("dollar", "dollars"), "£": ("pound", "pounds"), "€": ("euro", "euros")}
+    minor_map = {"$": ("cent", "cents"), "£": ("penny", "pence"), "€": ("cent", "cents")}
+    major_singular, major_plural = major_map.get(symbol, ("", ""))
+    parts = amount_str.split('.')
+    major_val = int(parts[0]) if parts[0] else 0
+    minor_val = int(parts[1].ljust(2, '0')) if len(parts) > 1 and parts[1] else 0
+    result = []
+    if major_val > 0:
+        result.append(f"{_number_to_words_custom(major_val)} {major_singular if major_val == 1 else major_plural}")
+    if minor_val > 0:
+        minor_singular, minor_plural = minor_map.get(symbol, ("", ""))
+        result.append(f"{_number_to_words_custom(minor_val)} {minor_singular if minor_val == 1 else minor_plural}")
+    return " and ".join(result) or f"zero {major_plural}"
+
+
+def _expand_decimal_number(m):
+    num_str = m.group(1)
+    parts = num_str.split('.')
+    integer_part = _number_to_words_custom(parts[0])
+    fractional_part = ' '.join(_number_to_words_custom(digit) for digit in parts[1])
+    return f"{integer_part} point {fractional_part}"
+
+
+def _expand_date(m):
+    month, day, year = m.groups()
+    month_word = month_name[int(month)]
+    day_word = _ordinal_custom(day)
+    year_num = int(year)
+    if len(year) == 2: year_num += 2000 if year_num < 50 else 1900
+    return f"{month_word} {day_word}, {_expand_number_positive(str(year_num))}"
+
+
+def _expand_fraction(m):
+    n, d = int(m.group(1)), int(m.group(2))
+    if d == 0: return m.group(0)
+    common_fractions = {(1, 2): "one half", (1, 4): "one quarter", (3, 4): "three quarters"}
+    if (n, d) in common_fractions: return common_fractions[(n, d)]
+    return f"{_number_to_words_custom(n)} over {_number_to_words_custom(d)}"
+
+
+def _expand_ordinal_word(m):
+    return _ordinal_custom(m.group(0)[:-2])
+
+
+def _expand_number(m):
+    num_str = m.group(0)
+    if num_str.startswith('-'): return f"minus {_expand_number_positive(num_str[1:])}"
+    return _expand_number_positive(num_str)
+
+
+def _expand_number_positive(num_str):
+    num = int(num_str)
+    if 2000 <= num < 2010: return f"two thousand and {_number_to_words_custom(num % 100)}"
+    if 1100 <= num < 2100 and num % 100 != 0:
+        return f"{_number_to_words_custom(num // 100)} {_number_to_words_custom(num % 100)}"
+    return _number_to_words_custom(num_str)
+
+
+def _expand_acronym(m): return " ".join(m.group(0))
+
+
+def normalize(text):
+    text = "".join(char for char in unicodedata.normalize("NFD", text) if unicodedata.category(char) != "Mn")
+    text = re.sub(r"@", " at ", text)
+    for regex, replacement in _abbreviations: text = regex.sub(replacement, text)
+    text = re.sub(_currency_suffix_re, _expand_currency_suffix, text)
+    text = re.sub(_phone_re, _expand_phone_number, text)
+    text = re.sub(_dimension_re, _expand_dimension, text)
+    text = re.sub(_roman_re, _expand_roman, text)
+    text = re.sub(_decade_re, _expand_decade, text)
+    text = re.sub(_score_re, _expand_score, text)
+    text = re.sub(_date_re, _expand_date, text)
+    text = re.sub(_time_re, _expand_time, text)
+    text = re.sub(_ordinal_number_re, _convert_ordinal, text)
+    text = re.sub(_comma_number_re, _remove_commas, text)
+    text = re.sub(_currency_re, _expand_currency, text)
+    text = re.sub(_measurement_re, _expand_measurement, text)
+    text = re.sub(_fraction_re, _expand_fraction, text)
+    text = re.sub(_decimal_number_re, _expand_decimal_number, text)
+    text = re.sub(_ordinal_re, _expand_ordinal_word, text)
+    text = re.sub(_alphanumeric_re, _expand_alphanumeric, text)
+    text = re.sub(_acronym_re, _expand_acronym, text)
+    text = re.sub(_number_re, _expand_number, text)
+    text = text.lower()
+    text = re.sub(r"%", " percent", text)
+    domain_re = re.compile(r'\b([a-z0-9-]+)\.([a-z]{2,})\b')
+    while domain_re.search(text): text = domain_re.sub(r'\1 dot \2', text)
+    text = re.sub(r"[^a-z0-9'.,?!:;-]", " ", text)
+    text = re.sub(r"\s+", " ", text)
+    return text.strip()
diff --git a/genie_tts/G2P/English/WordSegment.py b/genie_tts/G2P/English/WordSegment.py
index 0359d9ca60ad76fbecc4ad2ba92d635942c6b959..26539258da8ddcae171dc0993cd13afa53136684 100644
--- a/genie_tts/G2P/English/WordSegment.py
+++ b/genie_tts/G2P/English/WordSegment.py
@@ -1,143 +1,143 @@
-import io
-import math
-import os
-from typing import List, Iterator, Tuple, Dict
-
-from ...Core.Resources import English_G2P_DIR
-
-
-class WordSegmenter:
-    """
-    Contains the core logic for word segmentation, adapted from the original library.
-    """
-    ALPHABET = set('abcdefghijklmnopqrstuvwxyz0123456789')
-    TOTAL = 1024908267229.0
-    LIMIT = 24
-
-    def __init__(self):
-        self.unigrams: Dict[str, float] = {}
-        self.bigrams: Dict[str, float] = {}
-        self.words: List[str] = []
-        self.total: float = 0.0
-
-    def load(self, data_directory: str):
-        """
-        Load unigram, bigram, and word counts from the specified data directory.
-        This is the main modification from the original library.
-        """
-        unigrams_path = os.path.join(data_directory, 'unigrams.txt')
-        bigrams_path = os.path.join(data_directory, 'bigrams.txt')
-        words_path = os.path.join(data_directory, 'words.txt')
-
-        for file_path in [unigrams_path, bigrams_path, words_path]:
-            if not os.path.exists(file_path):
-                raise FileNotFoundError(
-                    f"Word segmentation data file not found: {file_path}. "
-                    "Please ensure the data directory is correct."
-                )
-
-        self.unigrams.update(self._parse(unigrams_path))
-        self.bigrams.update(self._parse(bigrams_path))
-        with io.open(words_path, encoding='utf-8') as reader:
-            self.words.extend(reader.read().splitlines())
-
-        self.total = self.TOTAL
-
-    @staticmethod
-    def _parse(filename: str) -> Dict[str, float]:
-        """Read `filename` and parse tab-separated file of word and count pairs."""
-        with io.open(filename, encoding='utf-8') as reader:
-            # CORRECTED LINE: The generator now defines `line` before the comprehension uses it.
-            lines = (line.split('\t') for line in reader)
-            return {word: float(number) for word, number in lines if len(word) > 0 and len(number) > 0}
-
-    def score(self, word: str, previous: str = None) -> float:
-        """Score `word` in the context of `previous` word."""
-        if previous is None:
-            if word in self.unigrams:
-                return self.unigrams[word] / self.total
-            return 10.0 / (self.total * 10 ** len(word))
-
-        bigram = f'{previous} {word}'
-        if bigram in self.bigrams and previous in self.unigrams:
-            return self.bigrams[bigram] / self.total / self.score(previous)
-
-        return self.score(word)
-
-    def isegment(self, text: str) -> Iterator[str]:
-        """Return iterator of words that is the best segmenation of `text`."""
-        memo = {}
-
-        def search(text: str, previous: str = '<s>') -> Tuple[float, List[str]]:
-            if text == '':
-                return 0.0, []
-
-            def candidates() -> Iterator[Tuple[float, List[str]]]:
-                for prefix, suffix in self._divide(text):
-                    prefix_score = math.log10(self.score(prefix, previous))
-
-                    pair = (suffix, prefix)
-                    if pair not in memo:
-                        memo[pair] = search(suffix, prefix)
-                    suffix_score, suffix_words = memo[pair]
-
-                    yield prefix_score + suffix_score, [prefix] + suffix_words
-
-            return max(candidates())
-
-        clean_text = self._clean(text)
-
-        # Original logic to avoid recursion limits by chunking
-        size = 250
-        prefix = ''
-        if len(clean_text) > size:
-            for offset in range(0, len(clean_text), size):
-                chunk = clean_text[offset:(offset + size)]
-                _, chunk_words = search(prefix + chunk)
-
-                if len(chunk_words) > 5:
-                    prefix = ''.join(chunk_words[-5:])
-                    del chunk_words[-5:]
-                else:  # handle case where chunk is small
-                    prefix = ''.join(chunk_words)
-                    chunk_words = []
-
-                for word in chunk_words:
-                    yield word
-
-            _, prefix_words = search(prefix)
-            for word in prefix_words:
-                yield word
-        else:
-            _, words = search(clean_text)
-            for word in words:
-                yield word
-
-    def segment(self, text: str) -> List[str]:
-        """Return list of words that is the best segmenation of `text`."""
-        return list(self.isegment(text))
-
-    def _divide(self, text: str) -> Iterator[Tuple[str, str]]:
-        """Yield `(prefix, suffix)` pairs from `text`."""
-        for pos in range(1, min(len(text), self.LIMIT) + 1):
-            yield text[:pos], text[pos:]
-
-    @classmethod
-    def _clean(cls, text: str) -> str:
-        """Return `text` lower-cased with non-alphanumeric characters removed."""
-        text_lower = text.lower()
-        return ''.join(letter for letter in text_lower if letter in cls.ALPHABET)
-
-
-# --- Public Interface ---
-# Create a single instance to be used by the importing module.
-
-_segmenter = WordSegmenter()
-_segmenter.load(os.path.join(English_G2P_DIR, 'wordsegment'))
-
-
-def segment_text(text: str) -> List[str]:
-    """
-    Public function to segment a text string into a list of words.
-    """
-    return _segmenter.segment(text)
+import io
+import math
+import os
+from typing import List, Iterator, Tuple, Dict
+
+from ...Core.Resources import English_G2P_DIR
+
+
+class WordSegmenter:
+    """
+    Contains the core logic for word segmentation, adapted from the original library.
+    """
+    ALPHABET = set('abcdefghijklmnopqrstuvwxyz0123456789')
+    TOTAL = 1024908267229.0
+    LIMIT = 24
+
+    def __init__(self):
+        self.unigrams: Dict[str, float] = {}
+        self.bigrams: Dict[str, float] = {}
+        self.words: List[str] = []
+        self.total: float = 0.0
+
+    def load(self, data_directory: str):
+        """
+        Load unigram, bigram, and word counts from the specified data directory.
+        This is the main modification from the original library.
+        """
+        unigrams_path = os.path.join(data_directory, 'unigrams.txt')
+        bigrams_path = os.path.join(data_directory, 'bigrams.txt')
+        words_path = os.path.join(data_directory, 'words.txt')
+
+        for file_path in [unigrams_path, bigrams_path, words_path]:
+            if not os.path.exists(file_path):
+                raise FileNotFoundError(
+                    f"Word segmentation data file not found: {file_path}. "
+                    "Please ensure the data directory is correct."
+                )
+
+        self.unigrams.update(self._parse(unigrams_path))
+        self.bigrams.update(self._parse(bigrams_path))
+        with io.open(words_path, encoding='utf-8') as reader:
+            self.words.extend(reader.read().splitlines())
+
+        self.total = self.TOTAL
+
+    @staticmethod
+    def _parse(filename: str) -> Dict[str, float]:
+        """Read `filename` and parse tab-separated file of word and count pairs."""
+        with io.open(filename, encoding='utf-8') as reader:
+            # CORRECTED LINE: The generator now defines `line` before the comprehension uses it.
+            lines = (line.split('\t') for line in reader)
+            return {word: float(number) for word, number in lines if len(word) > 0 and len(number) > 0}
+
+    def score(self, word: str, previous: str = None) -> float:
+        """Score `word` in the context of `previous` word."""
+        if previous is None:
+            if word in self.unigrams:
+                return self.unigrams[word] / self.total
+            return 10.0 / (self.total * 10 ** len(word))
+
+        bigram = f'{previous} {word}'
+        if bigram in self.bigrams and previous in self.unigrams:
+            return self.bigrams[bigram] / self.total / self.score(previous)
+
+        return self.score(word)
+
+    def isegment(self, text: str) -> Iterator[str]:
+        """Return iterator of words that is the best segmenation of `text`."""
+        memo = {}
+
+        def search(text: str, previous: str = '<s>') -> Tuple[float, List[str]]:
+            if text == '':
+                return 0.0, []
+
+            def candidates() -> Iterator[Tuple[float, List[str]]]:
+                for prefix, suffix in self._divide(text):
+                    prefix_score = math.log10(self.score(prefix, previous))
+
+                    pair = (suffix, prefix)
+                    if pair not in memo:
+                        memo[pair] = search(suffix, prefix)
+                    suffix_score, suffix_words = memo[pair]
+
+                    yield prefix_score + suffix_score, [prefix] + suffix_words
+
+            return max(candidates())
+
+        clean_text = self._clean(text)
+
+        # Original logic to avoid recursion limits by chunking
+        size = 250
+        prefix = ''
+        if len(clean_text) > size:
+            for offset in range(0, len(clean_text), size):
+                chunk = clean_text[offset:(offset + size)]
+                _, chunk_words = search(prefix + chunk)
+
+                if len(chunk_words) > 5:
+                    prefix = ''.join(chunk_words[-5:])
+                    del chunk_words[-5:]
+                else:  # handle case where chunk is small
+                    prefix = ''.join(chunk_words)
+                    chunk_words = []
+
+                for word in chunk_words:
+                    yield word
+
+            _, prefix_words = search(prefix)
+            for word in prefix_words:
+                yield word
+        else:
+            _, words = search(clean_text)
+            for word in words:
+                yield word
+
+    def segment(self, text: str) -> List[str]:
+        """Return list of words that is the best segmenation of `text`."""
+        return list(self.isegment(text))
+
+    def _divide(self, text: str) -> Iterator[Tuple[str, str]]:
+        """Yield `(prefix, suffix)` pairs from `text`."""
+        for pos in range(1, min(len(text), self.LIMIT) + 1):
+            yield text[:pos], text[pos:]
+
+    @classmethod
+    def _clean(cls, text: str) -> str:
+        """Return `text` lower-cased with non-alphanumeric characters removed."""
+        text_lower = text.lower()
+        return ''.join(letter for letter in text_lower if letter in cls.ALPHABET)
+
+
+# --- Public Interface ---
+# Create a single instance to be used by the importing module.
+
+_segmenter = WordSegmenter()
+_segmenter.load(os.path.join(English_G2P_DIR, 'wordsegment'))
+
+
+def segment_text(text: str) -> List[str]:
+    """
+    Public function to segment a text string into a list of words.
+    """
+    return _segmenter.segment(text)
diff --git a/genie_tts/G2P/Japanese/JapaneseG2P.py b/genie_tts/G2P/Japanese/JapaneseG2P.py
index 94362fc4b62601530de7b08699e43e91eaa4e591..194d9b459666595b2b609668a5cf50f0fe3f7719 100644
--- a/genie_tts/G2P/Japanese/JapaneseG2P.py
+++ b/genie_tts/G2P/Japanese/JapaneseG2P.py
@@ -1,150 +1,150 @@
-# -*- coding: utf-8 -*-
-"""
-用于纯日语的 G2P。
-"""
-import re
-import pyopenjtalk
-from typing import List
-from ..SymbolsV2 import symbols_v2, symbol_to_id_v2
-
-# 匹配连续的标点符号
-_CONSECUTIVE_PUNCTUATION_RE = re.compile(r"([,./?!~…・])\1+")
-
-# 匹配需要转换为日语读法的特殊符号
-_SYMBOLS_TO_JAPANESE = [
-    (re.compile("%"), "パーセント"),
-    (re.compile("％"), "パーセント"),
-]
-
-# 匹配日语字符（汉字、假名、全角字母数字等）
-_JAPANESE_CHARACTERS_RE = re.compile(
-    r"[A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]"
-)
-
-# 匹配非日语字符（标点、空格等）
-_JAPANESE_MARKS_RE = re.compile(
-    r"[^A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]"
-)
-
-
-class JapaneseG2P:
-    """
-    一个简化的、封装好的日语Grapheme-to-Phoneme（字素到音素）转换器。
-
-    本版本假设 pyopenjtalk 库已安装，并且不使用任何用户自定义词典。
-    它专注于提供一个纯粹、高效的文本到音素转换接口。
-    """
-
-    @staticmethod
-    def _text_normalize(text: str) -> str:
-        """对输入文本进行基础的规范化处理。"""
-        for regex, replacement in _SYMBOLS_TO_JAPANESE:
-            text = re.sub(regex, replacement, text)
-        text = _CONSECUTIVE_PUNCTUATION_RE.sub(r"\1", text)
-        text = text.lower()
-        return text
-
-    @staticmethod
-    def _post_replace_phoneme(ph: str) -> str:
-        """对单个音素或标点进行后处理替换。"""
-        rep_map = {
-            "：": ",", "；": ",", "，": ",", "。": ".",
-            "！": "!", "？": "?", "\n": ".", "·": ",",
-            "、": ",", "...": "…",
-        }
-        return rep_map.get(ph, ph)
-
-    @staticmethod
-    def _numeric_feature_by_regex(regex: str, s: str) -> int:
-        """从OpenJTalk标签中提取数值特征。"""
-        match = re.search(regex, s)
-        return int(match.group(1)) if match else -50
-
-    @staticmethod
-    def _pyopenjtalk_g2p_prosody(text: str) -> List[str]:
-        """使用pyopenjtalk提取音素及韵律符号。"""
-        labels = pyopenjtalk.make_label(pyopenjtalk.run_frontend(text))
-        phones = []
-        for n, lab_curr in enumerate(labels):
-            p3 = re.search(r"-(.*?)\+", lab_curr).group(1)
-            if p3 in "AEIOU":
-                p3 = p3.lower()
-
-            if p3 == "sil":
-                if n == 0:
-                    phones.append("^")
-                elif n == len(labels) - 1:
-                    e3 = JapaneseG2P._numeric_feature_by_regex(r"!(\d+)_", lab_curr)
-                    phones.append("?" if e3 == 1 else "$")
-                continue
-            elif p3 == "pau":
-                phones.append("_")
-                continue
-            else:
-                phones.append(p3)
-
-            a1 = JapaneseG2P._numeric_feature_by_regex(r"/A:([0-9\-]+)\+", lab_curr)
-            a2 = JapaneseG2P._numeric_feature_by_regex(r"\+(\d+)\+", lab_curr)
-            a3 = JapaneseG2P._numeric_feature_by_regex(r"\+(\d+)/", lab_curr)
-            f1 = JapaneseG2P._numeric_feature_by_regex(r"/F:(\d+)_", lab_curr)
-            lab_next = labels[n + 1] if n + 1 < len(labels) else ""
-            a2_next = JapaneseG2P._numeric_feature_by_regex(r"\+(\d+)\+", lab_next)
-
-            if a3 == 1 and a2_next == 1 and p3 in "aeiouAEIOUNcl":
-                phones.append("#")
-            elif a1 == 0 and a2_next == a2 + 1 and a2 != f1:
-                phones.append("]")
-            elif a2 == 1 and a2_next == 2:
-                phones.append("[")
-
-        return phones
-
-    @staticmethod
-    def g2p(text: str, with_prosody: bool = True) -> List[str]:
-        """
-        将日语文本转换为音素序列。
-
-        Args:
-            text (str): 待转换的日语文本。
-            with_prosody (bool): 是否在输出中包含韵律符号。默认为 True。
-
-        Returns:
-            List[str]: 音素和符号的列表。
-        """
-        if not text.strip():
-            return []
-
-        # 1. 文本规范化
-        norm_text = JapaneseG2P._text_normalize(text)
-
-        # 2. 使用标点符号分割字符串，得到日语文本片段
-        japanese_segments = _JAPANESE_MARKS_RE.split(norm_text)
-        punctuation_marks = _JAPANESE_MARKS_RE.findall(norm_text)
-
-        phonemes = []
-        for i, segment in enumerate(japanese_segments):
-            if segment:
-                if with_prosody:  # 移除分析结果中句首(^)/句尾($)的符号，因为我们按片段处理
-                    phones = JapaneseG2P._pyopenjtalk_g2p_prosody(segment)[1:-1]
-                else:
-                    phones = pyopenjtalk.g2p(segment).split(" ")
-                phonemes.extend(phones)
-
-            # 将对应的标点符号添加回来
-            if i < len(punctuation_marks):
-                mark = punctuation_marks[i].strip()
-                if mark:
-                    phonemes.append(mark)
-
-        # 3. 对最终列表中的每个元素进行后处理（主要转换全角标点）
-        processed_phonemes = [JapaneseG2P._post_replace_phoneme(p) for p in phonemes]
-
-        return processed_phonemes
-
-
-def japanese_to_phones(text: str) -> List[int]:
-    phones = JapaneseG2P.g2p(text)
-    phones = [ph for ph in phones if ph in symbols_v2]
-    # print(phones)
-    phones = [symbol_to_id_v2[ph] for ph in phones]
-    return phones
+# -*- coding: utf-8 -*-
+"""
+用于纯日语的 G2P。
+"""
+import re
+import pyopenjtalk
+from typing import List
+from ..SymbolsV2 import symbols_v2, symbol_to_id_v2
+
+# 匹配连续的标点符号
+_CONSECUTIVE_PUNCTUATION_RE = re.compile(r"([,./?!~…・])\1+")
+
+# 匹配需要转换为日语读法的特殊符号
+_SYMBOLS_TO_JAPANESE = [
+    (re.compile("%"), "パーセント"),
+    (re.compile("％"), "パーセント"),
+]
+
+# 匹配日语字符（汉字、假名、全角字母数字等）
+_JAPANESE_CHARACTERS_RE = re.compile(
+    r"[A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]"
+)
+
+# 匹配非日语字符（标点、空格等）
+_JAPANESE_MARKS_RE = re.compile(
+    r"[^A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]"
+)
+
+
+class JapaneseG2P:
+    """
+    一个简化的、封装好的日语Grapheme-to-Phoneme（字素到音素）转换器。
+
+    本版本假设 pyopenjtalk 库已安装，并且不使用任何用户自定义词典。
+    它专注于提供一个纯粹、高效的文本到音素转换接口。
+    """
+
+    @staticmethod
+    def _text_normalize(text: str) -> str:
+        """对输入文本进行基础的规范化处理。"""
+        for regex, replacement in _SYMBOLS_TO_JAPANESE:
+            text = re.sub(regex, replacement, text)
+        text = _CONSECUTIVE_PUNCTUATION_RE.sub(r"\1", text)
+        text = text.lower()
+        return text
+
+    @staticmethod
+    def _post_replace_phoneme(ph: str) -> str:
+        """对单个音素或标点进行后处理替换。"""
+        rep_map = {
+            "：": ",", "；": ",", "，": ",", "。": ".",
+            "！": "!", "？": "?", "\n": ".", "·": ",",
+            "、": ",", "...": "…",
+        }
+        return rep_map.get(ph, ph)
+
+    @staticmethod
+    def _numeric_feature_by_regex(regex: str, s: str) -> int:
+        """从OpenJTalk标签中提取数值特征。"""
+        match = re.search(regex, s)
+        return int(match.group(1)) if match else -50
+
+    @staticmethod
+    def _pyopenjtalk_g2p_prosody(text: str) -> List[str]:
+        """使用pyopenjtalk提取音素及韵律符号。"""
+        labels = pyopenjtalk.make_label(pyopenjtalk.run_frontend(text))
+        phones = []
+        for n, lab_curr in enumerate(labels):
+            p3 = re.search(r"-(.*?)\+", lab_curr).group(1)
+            if p3 in "AEIOU":
+                p3 = p3.lower()
+
+            if p3 == "sil":
+                if n == 0:
+                    phones.append("^")
+                elif n == len(labels) - 1:
+                    e3 = JapaneseG2P._numeric_feature_by_regex(r"!(\d+)_", lab_curr)
+                    phones.append("?" if e3 == 1 else "$")
+                continue
+            elif p3 == "pau":
+                phones.append("_")
+                continue
+            else:
+                phones.append(p3)
+
+            a1 = JapaneseG2P._numeric_feature_by_regex(r"/A:([0-9\-]+)\+", lab_curr)
+            a2 = JapaneseG2P._numeric_feature_by_regex(r"\+(\d+)\+", lab_curr)
+            a3 = JapaneseG2P._numeric_feature_by_regex(r"\+(\d+)/", lab_curr)
+            f1 = JapaneseG2P._numeric_feature_by_regex(r"/F:(\d+)_", lab_curr)
+            lab_next = labels[n + 1] if n + 1 < len(labels) else ""
+            a2_next = JapaneseG2P._numeric_feature_by_regex(r"\+(\d+)\+", lab_next)
+
+            if a3 == 1 and a2_next == 1 and p3 in "aeiouAEIOUNcl":
+                phones.append("#")
+            elif a1 == 0 and a2_next == a2 + 1 and a2 != f1:
+                phones.append("]")
+            elif a2 == 1 and a2_next == 2:
+                phones.append("[")
+
+        return phones
+
+    @staticmethod
+    def g2p(text: str, with_prosody: bool = True) -> List[str]:
+        """
+        将日语文本转换为音素序列。
+
+        Args:
+            text (str): 待转换的日语文本。
+            with_prosody (bool): 是否在输出中包含韵律符号。默认为 True。
+
+        Returns:
+            List[str]: 音素和符号的列表。
+        """
+        if not text.strip():
+            return []
+
+        # 1. 文本规范化
+        norm_text = JapaneseG2P._text_normalize(text)
+
+        # 2. 使用标点符号分割字符串，得到日语文本片段
+        japanese_segments = _JAPANESE_MARKS_RE.split(norm_text)
+        punctuation_marks = _JAPANESE_MARKS_RE.findall(norm_text)
+
+        phonemes = []
+        for i, segment in enumerate(japanese_segments):
+            if segment:
+                if with_prosody:  # 移除分析结果中句首(^)/句尾($)的符号，因为我们按片段处理
+                    phones = JapaneseG2P._pyopenjtalk_g2p_prosody(segment)[1:-1]
+                else:
+                    phones = pyopenjtalk.g2p(segment).split(" ")
+                phonemes.extend(phones)
+
+            # 将对应的标点符号添加回来
+            if i < len(punctuation_marks):
+                mark = punctuation_marks[i].strip()
+                if mark:
+                    phonemes.append(mark)
+
+        # 3. 对最终列表中的每个元素进行后处理（主要转换全角标点）
+        processed_phonemes = [JapaneseG2P._post_replace_phoneme(p) for p in phonemes]
+
+        return processed_phonemes
+
+
+def japanese_to_phones(text: str) -> List[int]:
+    phones = JapaneseG2P.g2p(text)
+    phones = [ph for ph in phones if ph in symbols_v2]
+    # print(phones)
+    phones = [symbol_to_id_v2[ph] for ph in phones]
+    return phones
diff --git a/genie_tts/G2P/__pycache__/SymbolsV2.cpython-311.pyc b/genie_tts/G2P/__pycache__/SymbolsV2.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3cfd9e35f1ef5c9228bf2c6ffebc88024e3f3baf
Binary files /dev/null and b/genie_tts/G2P/__pycache__/SymbolsV2.cpython-311.pyc differ
diff --git a/genie_tts/G2P/__pycache__/__init__.cpython-311.pyc b/genie_tts/G2P/__pycache__/__init__.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c0d8703d9de2a82812e870c42c43fc81129b7dcd
Binary files /dev/null and b/genie_tts/G2P/__pycache__/__init__.cpython-311.pyc differ
diff --git a/genie_tts/GUI/AudioPlayer.py b/genie_tts/GUI/AudioPlayer.py
index 76e615ce9985896bd2efca8a199fd2d54a649dc7..bc972bdd30360ea4ea2a13ad7292eee7fd151e29 100644
--- a/genie_tts/GUI/AudioPlayer.py
+++ b/genie_tts/GUI/AudioPlayer.py
@@ -1,94 +1,94 @@
-import sounddevice as sd
-import threading
-import queue
-from typing import Union, Optional, Callable
-import numpy as np
-import os
-import soundfile as sf
-
-
-def run_in_sub_thread(func) -> Callable[..., threading.Thread]:
-    def wrapper(*args, **kwargs) -> threading.Thread:
-        thread = threading.Thread(target=func, args=args, kwargs=kwargs)
-        thread.daemon = True
-        thread.start()
-        return thread
-
-    return wrapper
-
-
-class AudioPlayer:
-    CHUNK_SIZE: int = 1024
-
-    def __init__(self):
-        self._task_queue: queue.Queue[bytes | str] = queue.Queue()
-        self._worker_thread: Optional[threading.Thread] = None
-        self._stop_event: threading.Event = threading.Event()
-        self._start_worker()
-
-    def _start_worker(self):
-        """启动工作线程（如果它尚未运行或已关闭）。"""
-        if self._worker_thread and self._worker_thread.is_alive():
-            return
-        self._stop_event.clear()
-        self._worker_thread = self._playback_worker()
-
-    @run_in_sub_thread
-    def _playback_worker(self) -> None:
-        while not self._stop_event.is_set():
-            try:
-                task: str = self._task_queue.get(timeout=0.1)
-            except queue.Empty:
-                continue
-
-            stream = None
-            try:
-                if isinstance(task, str) and os.path.isfile(task):
-                    with sf.SoundFile(task, 'r') as f:
-                        if sd is not None:
-                            stream = sd.OutputStream(
-                                samplerate=f.samplerate,
-                                channels=f.channels,
-                                dtype='float32',
-                            )
-                            stream.start()
-                        while not self._stop_event.is_set():
-                            chunk = f.read(self.CHUNK_SIZE, dtype='float32')
-                            if not chunk.any():
-                                break
-                            stream.write(chunk)
-            except Exception as e:
-                if isinstance(e, sf.SoundFileError):
-                    print(f"无法读取或解析音频文件: {task}, 错误: {e}")
-                else:
-                    print(f"播放时发生错误: {e}")
-            finally:
-                if stream:
-                    stream.stop()
-                    stream.close()
-                self._task_queue.task_done()
-
-    def play(self, source: Union[str, np.ndarray]):
-        """将音频源加入播放队列。"""
-        self._start_worker()
-        self._task_queue.put(source)
-
-    def stop(self):
-        """停止播放并清空播放队列。"""
-        self._stop_event.set()
-        if self._worker_thread and self._worker_thread.is_alive():
-            self._worker_thread.join()
-        self._stop_event.clear()
-
-        with self._task_queue.mutex:
-            self._task_queue.queue.clear()
-            while self._task_queue.unfinished_tasks > 0:
-                self._task_queue.task_done()
-
-    def wait(self):
-        """阻塞，直到队列中所有任务都播放完成。"""
-        self._task_queue.join()
-
-    def close(self):
-        """永久关闭播放器并释放资源。"""
-        self.stop()
+import sounddevice as sd
+import threading
+import queue
+from typing import Union, Optional, Callable
+import numpy as np
+import os
+import soundfile as sf
+
+
+def run_in_sub_thread(func) -> Callable[..., threading.Thread]:
+    def wrapper(*args, **kwargs) -> threading.Thread:
+        thread = threading.Thread(target=func, args=args, kwargs=kwargs)
+        thread.daemon = True
+        thread.start()
+        return thread
+
+    return wrapper
+
+
+class AudioPlayer:
+    CHUNK_SIZE: int = 1024
+
+    def __init__(self):
+        self._task_queue: queue.Queue[bytes | str] = queue.Queue()
+        self._worker_thread: Optional[threading.Thread] = None
+        self._stop_event: threading.Event = threading.Event()
+        self._start_worker()
+
+    def _start_worker(self):
+        """启动工作线程（如果它尚未运行或已关闭）。"""
+        if self._worker_thread and self._worker_thread.is_alive():
+            return
+        self._stop_event.clear()
+        self._worker_thread = self._playback_worker()
+
+    @run_in_sub_thread
+    def _playback_worker(self) -> None:
+        while not self._stop_event.is_set():
+            try:
+                task: str = self._task_queue.get(timeout=0.1)
+            except queue.Empty:
+                continue
+
+            stream = None
+            try:
+                if isinstance(task, str) and os.path.isfile(task):
+                    with sf.SoundFile(task, 'r') as f:
+                        if sd is not None:
+                            stream = sd.OutputStream(
+                                samplerate=f.samplerate,
+                                channels=f.channels,
+                                dtype='float32',
+                            )
+                            stream.start()
+                        while not self._stop_event.is_set():
+                            chunk = f.read(self.CHUNK_SIZE, dtype='float32')
+                            if not chunk.any():
+                                break
+                            stream.write(chunk)
+            except Exception as e:
+                if isinstance(e, sf.SoundFileError):
+                    print(f"无法读取或解析音频文件: {task}, 错误: {e}")
+                else:
+                    print(f"播放时发生错误: {e}")
+            finally:
+                if stream:
+                    stream.stop()
+                    stream.close()
+                self._task_queue.task_done()
+
+    def play(self, source: Union[str, np.ndarray]):
+        """将音频源加入播放队列。"""
+        self._start_worker()
+        self._task_queue.put(source)
+
+    def stop(self):
+        """停止播放并清空播放队列。"""
+        self._stop_event.set()
+        if self._worker_thread and self._worker_thread.is_alive():
+            self._worker_thread.join()
+        self._stop_event.clear()
+
+        with self._task_queue.mutex:
+            self._task_queue.queue.clear()
+            while self._task_queue.unfinished_tasks > 0:
+                self._task_queue.task_done()
+
+    def wait(self):
+        """阻塞，直到队列中所有任务都播放完成。"""
+        self._task_queue.join()
+
+    def close(self):
+        """永久关闭播放器并释放资源。"""
+        self.stop()
diff --git a/genie_tts/GUI/ConverterWidget.py b/genie_tts/GUI/ConverterWidget.py
index a1d6bbf8596e1c6aebe67dfd328992dada328614..57e53222797f0cf310366b8e6a6a387198806e0a 100644
--- a/genie_tts/GUI/ConverterWidget.py
+++ b/genie_tts/GUI/ConverterWidget.py
@@ -1,204 +1,204 @@
-import sys
-import os
-import datetime
-
-from PySide6.QtWidgets import (
-    QApplication, QWidget, QVBoxLayout, QPushButton, QTextEdit, QFileDialog,
-    QListView, QTreeView, QAbstractItemView
-)
-from PySide6.QtCore import Signal, QObject, QSettings, QThread
-
-from ..Converter.Converter import convert
-from ..Converter.v2.Converter import find_ckpt_and_pth
-
-
-def get_timestamp_msg(message: str, level: str = "INFO") -> str:
-    """辅助函数：生成类似 Logging 格式的带时间戳字符串"""
-    now = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
-    return f"{now} - {level} - {message}"
-
-
-class Worker(QObject):
-    finished = Signal()
-    log_signal = Signal(str)
-
-    def __init__(self, folders):
-        super().__init__()
-        self.folders = folders
-
-    def log(self, message: str, level: str = "INFO"):
-        """内部辅助方法，用于格式化并发送日志"""
-        formatted_msg = get_timestamp_msg(message, level)
-        self.log_signal.emit(formatted_msg)
-
-    def run(self):
-        """执行转换任务"""
-        try:
-            root_output_dir = os.path.abspath("./Output")
-            for folder in self.folders:
-                character_name: str = os.path.basename(folder)
-                output_dir: str = os.path.join(root_output_dir, character_name)
-                if os.path.exists(output_dir):
-                    self.log(f'输出文件夹 {output_dir} 已存在，将覆盖内容。', "WARNING")
-                torch_ckpt_path, torch_pth_path = find_ckpt_and_pth(folder)
-                if not torch_ckpt_path or not torch_pth_path:
-                    self.log(f'无法处理文件夹 {folder} 。请保证文件夹内有 GPT—SOVITS V2 导出的 .pth 和 .ckpt 模型。',
-                             "ERROR")
-                    continue
-                self.log(f'正在处理 {folder} 。')
-                # 调用转换逻辑
-                convert(torch_ckpt_path, torch_pth_path, output_dir)
-                self.log(f'{folder} 处理完成。')  # 可选：提示完成
-            os.startfile(root_output_dir)
-        except Exception as e:
-            self.log(f"任务执行过程中发生未捕获异常: {str(e)}", "ERROR")
-        finally:
-            self.finished.emit()
-
-
-class ConverterWidget(QWidget):
-    def __init__(self):
-        super().__init__()
-        self.setWindowTitle('GENIE Converter (PySide6 Version)')
-        self.resize(1280, 720)
-
-        self.settings = QSettings("MyCompany", "GENIE Converter")
-
-        main_layout = QVBoxLayout(self)
-        main_layout.setContentsMargins(20, 20, 20, 20)
-        main_layout.setSpacing(15)
-
-        self.folder_button = QPushButton('📂 选择一个或多个文件夹')
-        self.folder_button.setFixedHeight(40)
-        self.folder_button.clicked.connect(self.open_folder_dialog)
-
-        self.log_display = QTextEdit()
-        self.log_display.setReadOnly(True)
-
-        main_layout.addWidget(self.folder_button)
-        main_layout.addWidget(self.log_display)
-
-        self.apply_stylesheet()
-
-        self.thread = None
-        self.worker = None
-
-        self.append_formatted_log("欢迎使用 GENIE Converter!")
-        self.append_formatted_log("支持将 GPT—SOVITS V2/V2ProPlus 模型导出为 GENIE 引擎所需的格式。")
-        self.append_formatted_log("请选择一个或多个文件夹，每个文件夹中包含一对 .pth 和 .ckpt 文件。")
-        self.append_formatted_log("您可以使用 Ctrl 或 Shift 键来进行多选。\n")
-
-    def apply_stylesheet(self):
-        self.setStyleSheet("""
-            QWidget {
-                background-color: #2b2b2b;
-                color: #f0f0f0;
-                font-family: 'Segoe UI', 'Microsoft YaHei', 'Arial';
-                font-size: 14px;
-            }
-            QPushButton {
-                background-color: #007bff;
-                color: white;
-                border: none;
-                padding: 10px;
-                border-radius: 5px;
-                font-weight: bold;
-            }
-            QPushButton:hover { background-color: #0056b3; }
-            QPushButton:pressed { background-color: #004494; }
-            QPushButton:disabled {
-                background-color: #555;
-                color: #aaa;
-            }
-            QTextEdit {
-                background-color: #1e1e1e;
-                border: 1px solid #444;
-                border-radius: 5px;
-                padding: 8px;
-                font-family: 'Consolas', 'Courier New', monospace;
-            }
-            QScrollBar:vertical {
-                border: none; background: #2b2b2b; width: 12px; margin: 0;
-            }
-            QScrollBar::handle:vertical {
-                background: #555; min-height: 20px; border-radius: 6px;
-            }
-            QScrollBar::handle:vertical:hover { background: #007bff; }
-            QScrollBar::add-line:vertical, QScrollBar::sub-line:vertical { height: 0px; }
-        """)
-
-    def append_log(self, text: str):
-        """直接给主线程调用的日志打印方法"""
-        self.log_display.append(text)
-        self.scroll_to_bottom()
-
-    def append_formatted_log(self, text: str, level="INFO"):
-        """给主线程调用的带格式日志方法"""
-        msg = get_timestamp_msg(text, level)
-        self.append_log(msg)
-
-    def scroll_to_bottom(self):
-        self.log_display.verticalScrollBar().setValue(
-            self.log_display.verticalScrollBar().maximum()
-        )
-
-    def open_folder_dialog(self):
-        last_dir = self.settings.value("last_dir", "")
-
-        dialog = QFileDialog(self, '请选择文件夹', str(last_dir))
-        dialog.setFileMode(QFileDialog.FileMode.Directory)
-        dialog.setOption(QFileDialog.Option.DontUseNativeDialog, True)
-
-        list_view = dialog.findChild(QListView, 'listView')
-        if list_view:
-            list_view.setSelectionMode(QAbstractItemView.SelectionMode.ExtendedSelection)
-
-        tree_view = dialog.findChild(QTreeView)
-        if tree_view:
-            tree_view.setSelectionMode(QAbstractItemView.SelectionMode.ExtendedSelection)
-
-        if dialog.exec():
-            selected_folders = dialog.selectedFiles()
-            if selected_folders:
-                self.run_conversion_task(selected_folders)
-                self.settings.setValue("last_dir", os.path.dirname(selected_folders[0]))
-
-    def run_conversion_task(self, folders):
-        self.folder_button.setEnabled(False)
-        self.folder_button.setText("🔄 转换中，请稍候...")
-
-        self.thread = QThread()
-        self.worker = Worker(folders)
-        self.worker.moveToThread(self.thread)
-
-        # 连接线程控制信号
-        self.thread.started.connect(self.worker.run)
-        self.worker.finished.connect(self.on_conversion_finished)
-        self.worker.log_signal.connect(self.append_log)
-
-        self.thread.start()
-
-    def on_conversion_finished(self):
-        self.thread.quit()
-        self.thread.wait()
-        self.worker.deleteLater()
-        self.thread.deleteLater()
-        self.thread = None
-        self.worker = None
-
-        self.folder_button.setEnabled(True)
-        self.folder_button.setText("📂 选择一个或多个文件夹")
-        self.append_formatted_log("所有任务已完成。", "INFO")
-
-    def closeEvent(self, event):
-        if self.thread is not None and self.thread.isRunning():
-            self.thread.quit()
-            self.thread.wait()
-        super().closeEvent(event)
-
-
-def start_gui() -> None:
-    app = QApplication(sys.argv)
-    window = ConverterWidget()
-    window.show()
-    sys.exit(app.exec())
+import sys
+import os
+import datetime
+
+from PySide6.QtWidgets import (
+    QApplication, QWidget, QVBoxLayout, QPushButton, QTextEdit, QFileDialog,
+    QListView, QTreeView, QAbstractItemView
+)
+from PySide6.QtCore import Signal, QObject, QSettings, QThread
+
+from ..Converter.Converter import convert
+from ..Converter.v2.Converter import find_ckpt_and_pth
+
+
+def get_timestamp_msg(message: str, level: str = "INFO") -> str:
+    """辅助函数：生成类似 Logging 格式的带时间戳字符串"""
+    now = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
+    return f"{now} - {level} - {message}"
+
+
+class Worker(QObject):
+    finished = Signal()
+    log_signal = Signal(str)
+
+    def __init__(self, folders):
+        super().__init__()
+        self.folders = folders
+
+    def log(self, message: str, level: str = "INFO"):
+        """内部辅助方法，用于格式化并发送日志"""
+        formatted_msg = get_timestamp_msg(message, level)
+        self.log_signal.emit(formatted_msg)
+
+    def run(self):
+        """执行转换任务"""
+        try:
+            root_output_dir = os.path.abspath("./Output")
+            for folder in self.folders:
+                character_name: str = os.path.basename(folder)
+                output_dir: str = os.path.join(root_output_dir, character_name)
+                if os.path.exists(output_dir):
+                    self.log(f'输出文件夹 {output_dir} 已存在，将覆盖内容。', "WARNING")
+                torch_ckpt_path, torch_pth_path = find_ckpt_and_pth(folder)
+                if not torch_ckpt_path or not torch_pth_path:
+                    self.log(f'无法处理文件夹 {folder} 。请保证文件夹内有 GPT—SOVITS V2 导出的 .pth 和 .ckpt 模型。',
+                             "ERROR")
+                    continue
+                self.log(f'正在处理 {folder} 。')
+                # 调用转换逻辑
+                convert(torch_ckpt_path, torch_pth_path, output_dir)
+                self.log(f'{folder} 处理完成。')  # 可选：提示完成
+            os.startfile(root_output_dir)
+        except Exception as e:
+            self.log(f"任务执行过程中发生未捕获异常: {str(e)}", "ERROR")
+        finally:
+            self.finished.emit()
+
+
+class ConverterWidget(QWidget):
+    def __init__(self):
+        super().__init__()
+        self.setWindowTitle('GENIE Converter (PySide6 Version)')
+        self.resize(1280, 720)
+
+        self.settings = QSettings("MyCompany", "GENIE Converter")
+
+        main_layout = QVBoxLayout(self)
+        main_layout.setContentsMargins(20, 20, 20, 20)
+        main_layout.setSpacing(15)
+
+        self.folder_button = QPushButton('📂 选择一个或多个文件夹')
+        self.folder_button.setFixedHeight(40)
+        self.folder_button.clicked.connect(self.open_folder_dialog)
+
+        self.log_display = QTextEdit()
+        self.log_display.setReadOnly(True)
+
+        main_layout.addWidget(self.folder_button)
+        main_layout.addWidget(self.log_display)
+
+        self.apply_stylesheet()
+
+        self.thread = None
+        self.worker = None
+
+        self.append_formatted_log("欢迎使用 GENIE Converter!")
+        self.append_formatted_log("支持将 GPT—SOVITS V2/V2ProPlus 模型导出为 GENIE 引擎所需的格式。")
+        self.append_formatted_log("请选择一个或多个文件夹，每个文件夹中包含一对 .pth 和 .ckpt 文件。")
+        self.append_formatted_log("您可以使用 Ctrl 或 Shift 键来进行多选。\n")
+
+    def apply_stylesheet(self):
+        self.setStyleSheet("""
+            QWidget {
+                background-color: #2b2b2b;
+                color: #f0f0f0;
+                font-family: 'Segoe UI', 'Microsoft YaHei', 'Arial';
+                font-size: 14px;
+            }
+            QPushButton {
+                background-color: #007bff;
+                color: white;
+                border: none;
+                padding: 10px;
+                border-radius: 5px;
+                font-weight: bold;
+            }
+            QPushButton:hover { background-color: #0056b3; }
+            QPushButton:pressed { background-color: #004494; }
+            QPushButton:disabled {
+                background-color: #555;
+                color: #aaa;
+            }
+            QTextEdit {
+                background-color: #1e1e1e;
+                border: 1px solid #444;
+                border-radius: 5px;
+                padding: 8px;
+                font-family: 'Consolas', 'Courier New', monospace;
+            }
+            QScrollBar:vertical {
+                border: none; background: #2b2b2b; width: 12px; margin: 0;
+            }
+            QScrollBar::handle:vertical {
+                background: #555; min-height: 20px; border-radius: 6px;
+            }
+            QScrollBar::handle:vertical:hover { background: #007bff; }
+            QScrollBar::add-line:vertical, QScrollBar::sub-line:vertical { height: 0px; }
+        """)
+
+    def append_log(self, text: str):
+        """直接给主线程调用的日志打印方法"""
+        self.log_display.append(text)
+        self.scroll_to_bottom()
+
+    def append_formatted_log(self, text: str, level="INFO"):
+        """给主线程调用的带格式日志方法"""
+        msg = get_timestamp_msg(text, level)
+        self.append_log(msg)
+
+    def scroll_to_bottom(self):
+        self.log_display.verticalScrollBar().setValue(
+            self.log_display.verticalScrollBar().maximum()
+        )
+
+    def open_folder_dialog(self):
+        last_dir = self.settings.value("last_dir", "")
+
+        dialog = QFileDialog(self, '请选择文件夹', str(last_dir))
+        dialog.setFileMode(QFileDialog.FileMode.Directory)
+        dialog.setOption(QFileDialog.Option.DontUseNativeDialog, True)
+
+        list_view = dialog.findChild(QListView, 'listView')
+        if list_view:
+            list_view.setSelectionMode(QAbstractItemView.SelectionMode.ExtendedSelection)
+
+        tree_view = dialog.findChild(QTreeView)
+        if tree_view:
+            tree_view.setSelectionMode(QAbstractItemView.SelectionMode.ExtendedSelection)
+
+        if dialog.exec():
+            selected_folders = dialog.selectedFiles()
+            if selected_folders:
+                self.run_conversion_task(selected_folders)
+                self.settings.setValue("last_dir", os.path.dirname(selected_folders[0]))
+
+    def run_conversion_task(self, folders):
+        self.folder_button.setEnabled(False)
+        self.folder_button.setText("🔄 转换中，请稍候...")
+
+        self.thread = QThread()
+        self.worker = Worker(folders)
+        self.worker.moveToThread(self.thread)
+
+        # 连接线程控制信号
+        self.thread.started.connect(self.worker.run)
+        self.worker.finished.connect(self.on_conversion_finished)
+        self.worker.log_signal.connect(self.append_log)
+
+        self.thread.start()
+
+    def on_conversion_finished(self):
+        self.thread.quit()
+        self.thread.wait()
+        self.worker.deleteLater()
+        self.thread.deleteLater()
+        self.thread = None
+        self.worker = None
+
+        self.folder_button.setEnabled(True)
+        self.folder_button.setText("📂 选择一个或多个文件夹")
+        self.append_formatted_log("所有任务已完成。", "INFO")
+
+    def closeEvent(self, event):
+        if self.thread is not None and self.thread.isRunning():
+            self.thread.quit()
+            self.thread.wait()
+        super().closeEvent(event)
+
+
+def start_gui() -> None:
+    app = QApplication(sys.argv)
+    window = ConverterWidget()
+    window.show()
+    sys.exit(app.exec())
diff --git a/genie_tts/GUI/GUI.py b/genie_tts/GUI/GUI.py
index c2b4e4408dc8d8567ce30e4bb3ada1f494d62185..8e5716e8e87ab348551a4b9c3e42e6971c1d2af6 100644
--- a/genie_tts/GUI/GUI.py
+++ b/genie_tts/GUI/GUI.py
@@ -1,649 +1,649 @@
-import sys
-import os
-import shutil
-from typing import List, Optional, TextIO, Any
-import uuid
-
-import soundfile as sf
-import numpy as np
-
-from PySide6.QtWidgets import (
-    QMainWindow, QWidget, QVBoxLayout, QHBoxLayout, QPushButton, QLabel, QLineEdit, QTextEdit,
-    QFileDialog, QGroupBox, QScrollArea, QMessageBox, QTabWidget, QFormLayout, QFrame,
-)
-from PySide6.QtCore import (
-    Qt, Signal, Slot, QObject,
-)
-from PySide6.QtGui import QTextCursor, QCloseEvent
-
-from ..Utils.TextSplitter import TextSplitter
-from .Utils import (
-    generate_output_filenames, FileSelectorWidget, FileSelectionMode, MyComboBox, sanitize_filename,
-    MyTextEdit
-)
-from .AudioPlayer import AudioPlayer
-from .PresetManager import PresetManager
-from .ServerManager import InferenceWorker
-from .ConverterWidget import ConverterWidget
-
-"""
-抄自 Genie CUDA Runtime
-"""
-
-CACHE_DIR = './UserData/Cache/GenieGUI'
-os.makedirs(CACHE_DIR, exist_ok=True)
-
-
-# ==================== 后台工作线程 ====================
-
-class LogRedirector(QObject):
-    """重定向 stdout 到 Signal"""
-    textWritten = Signal(str)
-
-    def __init__(self):
-        super().__init__()
-        self._old_stdout: TextIO = sys.stdout
-
-    def write(self, text: Any):
-        text = str(text)
-        self.textWritten.emit(text)
-        if self._old_stdout is not None:
-            self._old_stdout.write(text)
-
-    def flush(self):
-        pass
-
-
-# ==================== UI 组件实现 ====================
-
-class PreviewItemWidget(QFrame):
-    """单条音频预览组件"""
-
-    def __init__(
-            self,
-            index: int,
-            text: str,
-            file_path: str,
-            player: AudioPlayer,
-            parent: QWidget = None
-    ):
-        super().__init__(parent)
-        self.text: str = text
-        self.file_path: str = file_path
-        self.player: AudioPlayer = player
-        self.setFrameShape(QFrame.Shape.StyledPanel)
-        self.setFrameShadow(QFrame.Shadow.Raised)
-
-        layout = QHBoxLayout(self)
-        layout.setContentsMargins(5, 5, 5, 5)
-
-        # 编号
-        lbl_id = QLabel(f"#{index}")
-        lbl_id.setFixedWidth(40)
-        lbl_id.setStyleSheet("font-weight: bold; color: #555;")
-
-        # 文本
-        lbl_text = QLabel(text)
-        lbl_text.setFixedWidth(240)
-        lbl_text.setToolTip(text)
-
-        # 按钮 - 播放
-        btn_play = QPushButton("▶ 播放")
-        btn_play.setFixedWidth(80)
-        btn_play.clicked.connect(self._play_audio)
-
-        # 按钮 - 保存
-        btn_save = QPushButton("⬇ 保存")
-        btn_save.setFixedWidth(80)
-        btn_save.clicked.connect(self._save_file)
-
-        # 按钮 - 删除 (新增)
-        btn_del = QPushButton("🗑 删除")
-        btn_del.setFixedWidth(80)
-        btn_del.setStyleSheet("color: #ff4d4d;")  # 以此区分删除按钮
-        btn_del.clicked.connect(self._delete_item)
-
-        layout.addWidget(lbl_id)
-        layout.addWidget(lbl_text, 1)  # Stretch
-        layout.addWidget(btn_play)
-        layout.addWidget(btn_save)
-        layout.addWidget(btn_del)  # 添加到布局
-
-    def _play_audio(self):
-        # 播放前先停止其他播放
-        self.player.stop()
-        self.player.play(self.file_path)
-
-    def _save_file(self):
-        filename = sanitize_filename(self.text)
-        save_path, _ = QFileDialog.getSaveFileName(
-            self, "保存音频", f"{filename}.wav", "WAV Audio (*.wav)"
-        )
-        if save_path:
-            try:
-                shutil.copy(self.file_path, save_path)
-                QMessageBox.information(self, "成功", "文件保存成功！")
-            except Exception as e:
-                QMessageBox.critical(self, "错误", f"保存失败: {e}")
-
-    def _delete_item(self):
-        """删除当前条目及对应的文件"""
-        # 1. 停止播放
-        self.player.stop()
-
-        # 2. 尝试删除物理文件 (避免垃圾堆积)
-        try:
-            if os.path.exists(self.file_path):
-                os.remove(self.file_path)
-                print(f"[INFO] 已删除文件: {self.file_path}")
-        except Exception as e:
-            print(f"[WARN] 删除文件失败: {e}")
-
-        # 3. 从界面移除自身
-        self.deleteLater()
-
-
-class LogWidget(QWidget):
-    """日志显示Tab"""
-
-    def __init__(self, parent: QWidget = None):
-        super().__init__(parent)
-        layout = QVBoxLayout(self)
-        self.text_edit: QTextEdit = QTextEdit()
-        self.text_edit.setReadOnly(True)
-        self.text_edit.setStyleSheet(
-            "background-color: #1e1e1e;"
-            "color: #ecf0f1;"
-            "font-family: Consolas;"
-            "font-size: 12pt;"
-        )
-        layout.addWidget(self.text_edit)
-
-    @Slot(str)
-    def append_log(self, text: str):
-        self.text_edit.moveCursor(QTextCursor.MoveOperation.End)
-        self.text_edit.insertPlainText(text)
-        self.text_edit.moveCursor(QTextCursor.MoveOperation.End)
-
-
-class TTSWidget(QWidget):
-    """TTS 主交互界面"""
-
-    def __init__(self, player: AudioPlayer, parent: QWidget = None):
-        super().__init__(parent)
-        self.player: AudioPlayer = player
-        self.splitter: TextSplitter = TextSplitter()
-        self.current_gen_id: int = 0
-        self.current_worker: Optional[InferenceWorker] = None
-
-        main_layout = QVBoxLayout(self)
-
-        # ---------------- 顶部：预设管理器 ----------------
-        self.preset_manager = PresetManager(
-            presets_file='./UserData/GenieGuiConfig.json',
-            state_getter=self.get_ui_state,
-        )
-        self.preset_manager.sig_load_state.connect(self.apply_ui_state)
-        main_layout.addWidget(self.preset_manager)
-
-        # ---------------- 中间：滚动设置区 ----------------
-        scroll = QScrollArea()
-        scroll.setWidgetResizable(True)
-        content_widget = QWidget()
-
-        content_layout = QHBoxLayout(content_widget)
-        left_column_layout = QVBoxLayout()
-        right_column_layout = QVBoxLayout()
-
-        # ==================== 左侧列内容 ====================
-
-        # 模型设置组
-        group_model = QGroupBox("模型设置")
-        self.layout_model = QFormLayout()
-        self.combo_model_type = MyComboBox()
-        self.combo_model_type.addItems(["Genie-TTS"])
-        self.combo_model_type.currentTextChanged.connect(self._update_model_ui_visibility)
-        self.combo_model_type.setEnabled(False)
-        self.file_gpt = FileSelectorWidget("gpt_path", FileSelectionMode.FILE, "Checkpoints (*.ckpt)")
-        self.file_vits = FileSelectorWidget("vits_path", FileSelectionMode.FILE, "Models (*.pth)")
-        self.file_genie = FileSelectorWidget("genie_dir", FileSelectionMode.DIRECTORY)
-        self.file_gpt.pathChanged.connect(self._on_gpt_path_changed)
-        self.file_vits.pathChanged.connect(self._on_vits_path_changed)
-        self.layout_model.addRow("模型类型:", self.combo_model_type)
-        self.layout_model.addRow("GPT模型 (.ckpt):", self.file_gpt)
-        self.layout_model.addRow("VITS模型 (.pth):", self.file_vits)
-        self.layout_model.addRow("Genie模型目录:", self.file_genie)
-        group_model.setLayout(self.layout_model)
-        # 参考音频组
-        group_ref = QGroupBox("参考音频")
-        layout_ref = QFormLayout()
-        self.file_ref_audio = FileSelectorWidget("ref_audio", FileSelectionMode.FILE, "Audio (*.wav *.mp3)")
-        self.input_ref_text = QLineEdit()
-        self.input_ref_text.setPlaceholderText("请输入参考音频对应的文本...")
-        btn_play_ref = QPushButton("▶️")
-        btn_play_ref.setFixedWidth(30)
-        btn_play_ref.clicked.connect(self._play_ref_audio)
-        hbox_ref_text = QHBoxLayout()
-        hbox_ref_text.addWidget(self.input_ref_text)
-        hbox_ref_text.addWidget(btn_play_ref)
-        layout_ref.addRow("音频文件:", self.file_ref_audio)
-        layout_ref.addRow("音频文本:", hbox_ref_text)
-        group_ref.setLayout(layout_ref)
-
-        left_column_layout.addWidget(group_model)
-        left_column_layout.addWidget(group_ref)
-        left_column_layout.addStretch()
-
-        # ==================== 右侧列内容 ====================
-
-        # === 推理设置组 ===
-        group_infer = QGroupBox("推理参数")
-        layout_infer = QFormLayout()
-        self.combo_device = MyComboBox()
-        self.combo_device.addItems(["CPU"])
-        self.combo_device.setEnabled(False)
-        self.combo_quality = MyComboBox()
-        self.combo_quality.addItems(["质量优先"])
-        self.combo_quality.setEnabled(False)
-        self.combo_split = MyComboBox()
-        self.combo_split.addItems(["不切分", "智能切分", "按行切分"])
-        self.combo_mode = MyComboBox()
-        self.combo_mode.addItems(["串行推理"])
-        self.combo_mode.setEnabled(False)
-        self.combo_lang = MyComboBox()
-        self.combo_lang.addItems(["Chinese", "English", "Japanese"])
-        layout_infer.addRow("推理设备:\n(重启生效)", self.combo_device)
-        layout_infer.addRow("推理需求:", self.combo_quality)
-        layout_infer.addRow("分句方式:", self.combo_split)
-        layout_infer.addRow("推理模式:", self.combo_mode)
-        layout_infer.addRow("目标语言:", self.combo_lang)
-        group_infer.setLayout(layout_infer)
-
-        # === 自动保存组 ===
-        group_save = QGroupBox("自动保存设置")
-        self.layout_save = QFormLayout()
-        self.combo_save_mode = MyComboBox()
-        self.combo_save_mode.addItems(["禁用自动保存", "保存为单个文件", "保存为多个文件"])
-        self.combo_save_mode.currentIndexChanged.connect(self._update_save_ui_state)
-        default_out_path = os.path.join(os.path.expanduser("~"), "Desktop", "Genie 输出语音")
-        self.file_out_dir = FileSelectorWidget("out_dir", FileSelectionMode.DIRECTORY)
-        self.file_out_dir.set_path(default_out_path)
-        self.layout_save.addRow("保存方式:", self.combo_save_mode)
-        self.layout_save.addRow("输出文件夹:", self.file_out_dir)
-        group_save.setLayout(self.layout_save)
-
-        right_column_layout.addWidget(group_infer)
-        right_column_layout.addWidget(group_save)
-        right_column_layout.addStretch()
-
-        content_layout.addLayout(left_column_layout, 1)
-        content_layout.addLayout(right_column_layout, 1)
-        scroll.setWidget(content_widget)
-        main_layout.addWidget(scroll, 5)
-
-        # ==================== 底部：输入控制 + 输出预览 ====================
-
-        # 创建底部容器 widget
-        bottom_widget = QWidget()
-        bottom_layout = QHBoxLayout(bottom_widget)
-        bottom_layout.setContentsMargins(0, 0, 0, 0)  # 去除边距让它贴合
-
-        # --- 输入控制组 ---
-        group_input = QGroupBox("目标文本")
-        layout_input = QVBoxLayout()
-        self.text_input = MyTextEdit()
-        self.text_input.setPlaceholderText("请输入要合成的目标文本...")
-        self.text_input.setFixedHeight(300)
-        self.btn_start = QPushButton("开始推理")
-        self.btn_start.setFixedHeight(40)
-        self.btn_start.setStyleSheet("""
-            QPushButton {
-                background-color: #4CAF50; 
-                color: white; 
-                font-weight: bold; 
-                border-radius: 5px;
-            }
-            QPushButton:hover { background-color: #45a049; }
-            QPushButton:disabled { background-color: #cccccc; }
-        """)
-        self.btn_start.clicked.connect(self._start_inference)
-        layout_input.addWidget(self.text_input)
-        layout_input.addWidget(self.btn_start)
-        group_input.setLayout(layout_input)
-
-        # --- 输出预览组 ---
-        group_preview = QGroupBox("输出音频预览")
-        preview_layout = QVBoxLayout()
-        self.preview_scroll = QScrollArea()
-        self.preview_scroll.setWidgetResizable(True)
-        self.preview_container = QWidget()
-        self.preview_list_layout = QVBoxLayout(self.preview_container)
-        self.preview_list_layout.setAlignment(Qt.AlignmentFlag.AlignTop)
-        self.preview_scroll.setWidget(self.preview_container)
-        preview_layout.addWidget(self.preview_scroll)
-        group_preview.setLayout(preview_layout)
-
-        bottom_layout.addWidget(group_input, 1)
-        bottom_layout.addWidget(group_preview, 1)
-        main_layout.addWidget(bottom_widget, 3)
-
-        self.apply_ui_state(self.preset_manager.current_preset_data)
-
-    # ==================== 状态管理接口 (供 PresetManager 调用) ====================
-
-    @property
-    def current_preset_name(self) -> str:
-        return self.preset_manager.current_preset_name
-
-    @property
-    def current_preset_data(self) -> dict:
-        return self.preset_manager.current_preset_data
-
-    def get_ui_state(self) -> dict:
-        """收集当前UI状态为字典"""
-        return {
-            "model_type": self.combo_model_type.currentText(),
-            "gpt_path": self.file_gpt.get_path(),
-            "vits_path": self.file_vits.get_path(),
-            "genie_dir": self.file_genie.get_path(),
-            "ref_audio": self.file_ref_audio.get_path(),
-            "ref_text": self.input_ref_text.text(),
-            "device": self.combo_device.currentText().lower(),
-            "quality": self.combo_quality.currentText(),
-            "split": self.combo_split.currentText(),
-            "mode": self.combo_mode.currentText(),
-            "lang": self.combo_lang.currentText(),
-            "save_mode": self.combo_save_mode.currentText(),
-            "out_dir": self.file_out_dir.get_path()
-        }
-
-    @Slot(dict)
-    def apply_ui_state(self, data: dict) -> None:
-        """将字典数据应用到UI"""
-
-        def set_combo_text(combo: MyComboBox, text: str) -> None:
-            index = combo.findText(text)
-            if index >= 0:
-                combo.setCurrentIndex(index)
-
-        set_combo_text(self.combo_model_type, data.get("model_type", ""))
-        self.file_gpt.set_path(data.get("gpt_path", ""), block_signals=True)
-        self.file_vits.set_path(data.get("vits_path", ""), block_signals=True)
-        self.file_genie.set_path(data.get("genie_dir", ""))
-        self.file_ref_audio.set_path(data.get("ref_audio", ""))
-        self.input_ref_text.setText(data.get("ref_text", ""))
-
-        set_combo_text(self.combo_device, data.get("device", ""))
-        set_combo_text(self.combo_quality, data.get("quality", ""))
-        set_combo_text(self.combo_split, data.get("split", ""))
-        set_combo_text(self.combo_mode, data.get("mode", ""))
-        set_combo_text(self.combo_lang, data.get("lang", ""))
-        set_combo_text(self.combo_save_mode, data.get("save_mode", ""))
-
-        self.file_out_dir.set_path(data.get("out_dir", ""))
-
-        # 确保UI显隐状态正确
-        self._update_model_ui_visibility()
-        self._update_save_ui_state()
-
-    # ==================== UI 逻辑处理 ====================
-
-    def _update_model_ui_visibility(self, *args) -> None:
-        """根据模型类型控制文件选择器的显隐"""
-        is_gpt = self.combo_model_type.currentText() == "GPT-SoVITS"
-        self.layout_model.setRowVisible(self.file_gpt, is_gpt)
-        self.layout_model.setRowVisible(self.file_vits, is_gpt)
-        self.layout_model.setRowVisible(self.file_genie, not is_gpt)
-
-    @Slot(str)
-    def _on_gpt_path_changed(self, path: str):
-        if path and os.path.exists(path) and not self.file_vits.get_path():
-            self._try_auto_fill_sibling(path, ".pth", self.file_vits)
-
-    @Slot(str)
-    def _on_vits_path_changed(self, path: str):
-        if path and os.path.exists(path) and not self.file_gpt.get_path():
-            self._try_auto_fill_sibling(path, ".ckpt", self.file_gpt)
-
-    @staticmethod
-    def _try_auto_fill_sibling(current_path: str, target_ext: str, target_widget: FileSelectorWidget):
-        try:
-            directory = os.path.dirname(current_path)
-            if not os.path.exists(directory):
-                return
-            for f in os.listdir(directory):
-                if f.lower().endswith(target_ext.lower()):
-                    full_path = os.path.join(directory, f)
-                    target_widget.set_path(full_path)
-                    print(f"[INFO] 自动关联模型文件: {full_path}")
-                    break
-        except Exception as e:
-            print(f"[WARN] 自动关联文件失败: {e}")
-
-    def _update_save_ui_state(self) -> None:
-        enabled = self.combo_save_mode.currentText() != "禁用自动保存"
-        self.layout_save.setRowVisible(self.file_out_dir, enabled)
-
-    def _play_ref_audio(self) -> None:
-        path = self.file_ref_audio.get_path()
-        if os.path.exists(path):
-            self.player.stop()
-            self.player.play(path)
-        else:
-            QMessageBox.warning(self, "错误", "参考音频文件不存在")
-
-    def _get_split_texts(self, text: str) -> List[str]:
-        method = self.combo_split.currentText()
-        if method == "不切分":
-            return [text]
-        elif method == "按行切分":
-            return [line.strip() for line in text.split('\n') if line.strip()]
-        elif method == "智能切分":
-            return self.splitter.split(text)
-        return [text]
-
-    def _start_inference(self) -> None:
-        text = self.text_input.toPlainText().strip()
-        if not text:
-            QMessageBox.warning(self, "提示", "请输入目标文本")
-            return
-
-        ref_path = self.file_ref_audio.get_path()
-        ref_text = self.input_ref_text.text().strip()
-        if not ref_path or not ref_text:
-            QMessageBox.warning(self, "提示", "请设置参考音频")
-            return
-
-        if not self.file_genie.get_path():
-            QMessageBox.warning(self, "提示", "请选择Genie模型目录")
-            return
-
-        out_dir = self.file_out_dir.get_path()
-        save_mode = self.combo_save_mode.currentText()
-        if not out_dir and save_mode != "禁用自动保存":
-            desktop = os.path.join(os.path.expanduser("~"), "Desktop", "Genie Output")
-            self.file_out_dir.set_path(desktop)
-            print(f"[INFO] 未设置输出文件夹, 将在桌面创建!")
-
-        self.btn_start.setEnabled(False)
-        self.btn_start.setText("推理中...")
-        self._chain_import_model()
-
-    # ==================== 推理链式调用 ====================
-
-    def _chain_import_model(self) -> None:
-        req = {
-            "character_name": self.current_preset_name,
-            "onnx_model_dir": self.file_genie.get_path(),
-            "language": self.combo_lang.currentText(),
-        }
-        worker = InferenceWorker(req, mode="load_character")
-        worker.finished.connect(lambda s, m, d: self._on_import_finished(s, m))
-        worker.start()
-        self.current_worker = worker
-
-    @Slot(bool, str)
-    def _on_import_finished(self, success: bool, msg: str) -> None:
-        if not success:
-            self._reset_ui_state()
-            QMessageBox.critical(self, "模型加载失败", msg)
-            return
-        print(f"[INFO] {msg}")
-        self._chain_set_ref()
-
-    def _chain_set_ref(self) -> None:
-        req = {
-            "character_name": self.current_preset_name,
-            "audio_path": self.file_ref_audio.get_path(),
-            "audio_text": self.input_ref_text.text().strip(),
-            "language": self.combo_lang.currentText(),
-        }
-        worker = InferenceWorker(req, mode="set_reference_audio")
-        worker.finished.connect(lambda s, m, d: self._on_set_ref_finished(s, m))
-        worker.start()
-        self.current_worker = worker
-
-    @Slot(bool, str)
-    def _on_set_ref_finished(self, success: bool, msg: str) -> None:
-        if not success:
-            self._reset_ui_state()
-            QMessageBox.critical(self, "设置参考音频失败", msg)
-            return
-        print(f"[INFO] {msg}")
-        self._chain_tts()
-
-    def _chain_tts(self) -> None:
-        text_full = self.text_input.toPlainText().strip()
-        text_list = self._get_split_texts(text_full)
-
-        print(f"[INFO] 开始串行推理, 分句结果: {text_list}")
-        self._process_serial_step(0, text_list, [], 32000)
-
-    def _process_serial_step(
-            self,
-            index: int,
-            text_list: List[str],
-            audio_accumulator: List[np.ndarray],
-            sample_rate: int
-    ) -> None:
-        # 1. 终止条件：所有句子处理完毕
-        if index >= len(text_list):
-            save_mode = self.combo_save_mode.currentText()
-            out_dir = self.file_out_dir.get_path()
-
-            if out_dir:
-                os.makedirs(out_dir, exist_ok=True)
-
-            if audio_accumulator and save_mode != "保存为多个文件":
-                full_text = ''.join(text_list)
-                full_audio = np.concatenate(audio_accumulator, axis=0)
-                if save_mode == "保存为单个文件":
-                    target_names = generate_output_filenames(folder=out_dir, original_texts=[full_text])
-                    save_path = os.path.join(out_dir, target_names[0])
-                else:  # "禁用自动保存"
-                    save_path = os.path.join(CACHE_DIR, f"{uuid.uuid4().hex}.wav")
-                sf.write(save_path, data=full_audio, samplerate=sample_rate, subtype='PCM_16')
-                self._add_to_preview(full_text, save_path)
-
-            print(f"\n[INFO] 串行推理全部完成，共 {len(text_list)} 句。")
-            self._reset_ui_state()
-            return
-
-        # 2. 递归进行：发起当前句子的请求
-        req = {
-            "character_name": self.current_preset_name,
-            "text": text_list[index],
-        }
-        worker = InferenceWorker(req, mode="tts")
-        worker.finished.connect(
-            lambda s, m, d: self._on_serial_step_finished(s, m, d, index, text_list, audio_accumulator)
-        )
-        worker.start()
-        self.current_worker = worker
-
-    @Slot(bool, str, object, int, object, object, object)
-    def _on_serial_step_finished(
-            self,
-            success: bool,
-            msg: str,
-            return_data: dict,
-            index: int,
-            text_list: List[str],
-            audio_accumulator: List[np.ndarray]
-    ) -> None:
-        if not success:
-            self._reset_ui_state()
-            QMessageBox.critical(self, "推理失败", f"第 {index + 1} 句出错: {msg}")
-            return
-
-        sr = return_data.get("sample_rate", 32000)
-        audio_list = return_data.get("audio_list", [])
-        save_mode = self.combo_save_mode.currentText()
-        out_dir = self.file_out_dir.get_path()
-        if out_dir:
-            os.makedirs(out_dir, exist_ok=True)
-
-        if audio_list:
-            audio_accumulator.append(audio_list[0])
-            if save_mode == "保存为多个文件":
-                target_names = generate_output_filenames(folder=out_dir, original_texts=[text_list[index]])
-                save_path = os.path.join(out_dir, target_names[0])
-                sf.write(save_path, data=audio_list[0], samplerate=sr, subtype='FLOAT')
-                self._add_to_preview(text_list[index], save_path)
-        else:
-            print(f"[WARN] 第 {index + 1} 句返回空音频")
-
-        # 继续处理下一句
-        self._process_serial_step(index + 1, text_list, audio_accumulator, sr)
-
-    def _add_to_preview(self, text: str, path: str) -> None:
-        item = PreviewItemWidget(self.current_gen_id, text, path, self.player)
-        self.preview_list_layout.insertWidget(0, item)
-        self.current_gen_id += 1
-
-    def _reset_ui_state(self) -> None:
-        self.btn_start.setEnabled(True)
-        self.btn_start.setText("开始推理")
-
-    def closeEvent(self, event: QCloseEvent) -> None:
-        # 委托 PresetManager 处理保存逻辑
-        self.preset_manager.shutdown()
-        super().closeEvent(event)
-
-
-class MainWindow(QMainWindow):
-    def __init__(self):
-        super().__init__()
-        self.setWindowTitle("Genie TTS Inference GUI")
-        self.resize(1300, 900)
-
-        # 初始化音频播放器
-        self.player: AudioPlayer = AudioPlayer()
-
-        # 初始化日志重定向
-        self.log_widget: LogWidget = LogWidget()
-        sys.stdout = LogRedirector()
-        sys.stdout.textWritten.connect(self.log_widget.append_log)
-
-        # 初始化主界面
-        self.tabs: QTabWidget = QTabWidget()
-        self.tts_widget = TTSWidget(self.player)
-        self.conv_widget = ConverterWidget()
-
-        self.tabs.addTab(self.log_widget, "GUI Log")
-        self.tabs.addTab(self.tts_widget, "TTS Inference")
-        self.tabs.addTab(self.conv_widget, "Converter")
-        self.tabs.setCurrentIndex(1)  # 默认显示TTS页
-
-        self.setCentralWidget(self.tabs)
-
-    def closeEvent(self, event: QCloseEvent) -> None:
-        if os.path.exists(CACHE_DIR):
-            shutil.rmtree(CACHE_DIR)
-        if hasattr(self, 'player'):
-            self.player.stop()
-        # 线程安全退出后，再恢复 stdout
-        sys.stdout = sys.__stdout__
-        if hasattr(self, 'tts_widget'):
-            self.tts_widget.closeEvent(event)
-        event.accept()
+import sys
+import os
+import shutil
+from typing import List, Optional, TextIO, Any
+import uuid
+
+import soundfile as sf
+import numpy as np
+
+from PySide6.QtWidgets import (
+    QMainWindow, QWidget, QVBoxLayout, QHBoxLayout, QPushButton, QLabel, QLineEdit, QTextEdit,
+    QFileDialog, QGroupBox, QScrollArea, QMessageBox, QTabWidget, QFormLayout, QFrame,
+)
+from PySide6.QtCore import (
+    Qt, Signal, Slot, QObject,
+)
+from PySide6.QtGui import QTextCursor, QCloseEvent
+
+from ..Utils.TextSplitter import TextSplitter
+from .Utils import (
+    generate_output_filenames, FileSelectorWidget, FileSelectionMode, MyComboBox, sanitize_filename,
+    MyTextEdit
+)
+from .AudioPlayer import AudioPlayer
+from .PresetManager import PresetManager
+from .ServerManager import InferenceWorker
+from .ConverterWidget import ConverterWidget
+
+"""
+抄自 Genie CUDA Runtime
+"""
+
+CACHE_DIR = './UserData/Cache/GenieGUI'
+os.makedirs(CACHE_DIR, exist_ok=True)
+
+
+# ==================== 后台工作线程 ====================
+
+class LogRedirector(QObject):
+    """重定向 stdout 到 Signal"""
+    textWritten = Signal(str)
+
+    def __init__(self):
+        super().__init__()
+        self._old_stdout: TextIO = sys.stdout
+
+    def write(self, text: Any):
+        text = str(text)
+        self.textWritten.emit(text)
+        if self._old_stdout is not None:
+            self._old_stdout.write(text)
+
+    def flush(self):
+        pass
+
+
+# ==================== UI 组件实现 ====================
+
+class PreviewItemWidget(QFrame):
+    """单条音频预览组件"""
+
+    def __init__(
+            self,
+            index: int,
+            text: str,
+            file_path: str,
+            player: AudioPlayer,
+            parent: QWidget = None
+    ):
+        super().__init__(parent)
+        self.text: str = text
+        self.file_path: str = file_path
+        self.player: AudioPlayer = player
+        self.setFrameShape(QFrame.Shape.StyledPanel)
+        self.setFrameShadow(QFrame.Shadow.Raised)
+
+        layout = QHBoxLayout(self)
+        layout.setContentsMargins(5, 5, 5, 5)
+
+        # 编号
+        lbl_id = QLabel(f"#{index}")
+        lbl_id.setFixedWidth(40)
+        lbl_id.setStyleSheet("font-weight: bold; color: #555;")
+
+        # 文本
+        lbl_text = QLabel(text)
+        lbl_text.setFixedWidth(240)
+        lbl_text.setToolTip(text)
+
+        # 按钮 - 播放
+        btn_play = QPushButton("▶ 播放")
+        btn_play.setFixedWidth(80)
+        btn_play.clicked.connect(self._play_audio)
+
+        # 按钮 - 保存
+        btn_save = QPushButton("⬇ 保存")
+        btn_save.setFixedWidth(80)
+        btn_save.clicked.connect(self._save_file)
+
+        # 按钮 - 删除 (新增)
+        btn_del = QPushButton("🗑 删除")
+        btn_del.setFixedWidth(80)
+        btn_del.setStyleSheet("color: #ff4d4d;")  # 以此区分删除按钮
+        btn_del.clicked.connect(self._delete_item)
+
+        layout.addWidget(lbl_id)
+        layout.addWidget(lbl_text, 1)  # Stretch
+        layout.addWidget(btn_play)
+        layout.addWidget(btn_save)
+        layout.addWidget(btn_del)  # 添加到布局
+
+    def _play_audio(self):
+        # 播放前先停止其他播放
+        self.player.stop()
+        self.player.play(self.file_path)
+
+    def _save_file(self):
+        filename = sanitize_filename(self.text)
+        save_path, _ = QFileDialog.getSaveFileName(
+            self, "保存音频", f"{filename}.wav", "WAV Audio (*.wav)"
+        )
+        if save_path:
+            try:
+                shutil.copy(self.file_path, save_path)
+                QMessageBox.information(self, "成功", "文件保存成功！")
+            except Exception as e:
+                QMessageBox.critical(self, "错误", f"保存失败: {e}")
+
+    def _delete_item(self):
+        """删除当前条目及对应的文件"""
+        # 1. 停止播放
+        self.player.stop()
+
+        # 2. 尝试删除物理文件 (避免垃圾堆积)
+        try:
+            if os.path.exists(self.file_path):
+                os.remove(self.file_path)
+                print(f"[INFO] 已删除文件: {self.file_path}")
+        except Exception as e:
+            print(f"[WARN] 删除文件失败: {e}")
+
+        # 3. 从界面移除自身
+        self.deleteLater()
+
+
+class LogWidget(QWidget):
+    """日志显示Tab"""
+
+    def __init__(self, parent: QWidget = None):
+        super().__init__(parent)
+        layout = QVBoxLayout(self)
+        self.text_edit: QTextEdit = QTextEdit()
+        self.text_edit.setReadOnly(True)
+        self.text_edit.setStyleSheet(
+            "background-color: #1e1e1e;"
+            "color: #ecf0f1;"
+            "font-family: Consolas;"
+            "font-size: 12pt;"
+        )
+        layout.addWidget(self.text_edit)
+
+    @Slot(str)
+    def append_log(self, text: str):
+        self.text_edit.moveCursor(QTextCursor.MoveOperation.End)
+        self.text_edit.insertPlainText(text)
+        self.text_edit.moveCursor(QTextCursor.MoveOperation.End)
+
+
+class TTSWidget(QWidget):
+    """TTS 主交互界面"""
+
+    def __init__(self, player: AudioPlayer, parent: QWidget = None):
+        super().__init__(parent)
+        self.player: AudioPlayer = player
+        self.splitter: TextSplitter = TextSplitter()
+        self.current_gen_id: int = 0
+        self.current_worker: Optional[InferenceWorker] = None
+
+        main_layout = QVBoxLayout(self)
+
+        # ---------------- 顶部：预设管理器 ----------------
+        self.preset_manager = PresetManager(
+            presets_file='./UserData/GenieGuiConfig.json',
+            state_getter=self.get_ui_state,
+        )
+        self.preset_manager.sig_load_state.connect(self.apply_ui_state)
+        main_layout.addWidget(self.preset_manager)
+
+        # ---------------- 中间：滚动设置区 ----------------
+        scroll = QScrollArea()
+        scroll.setWidgetResizable(True)
+        content_widget = QWidget()
+
+        content_layout = QHBoxLayout(content_widget)
+        left_column_layout = QVBoxLayout()
+        right_column_layout = QVBoxLayout()
+
+        # ==================== 左侧列内容 ====================
+
+        # 模型设置组
+        group_model = QGroupBox("模型设置")
+        self.layout_model = QFormLayout()
+        self.combo_model_type = MyComboBox()
+        self.combo_model_type.addItems(["Genie-TTS"])
+        self.combo_model_type.currentTextChanged.connect(self._update_model_ui_visibility)
+        self.combo_model_type.setEnabled(False)
+        self.file_gpt = FileSelectorWidget("gpt_path", FileSelectionMode.FILE, "Checkpoints (*.ckpt)")
+        self.file_vits = FileSelectorWidget("vits_path", FileSelectionMode.FILE, "Models (*.pth)")
+        self.file_genie = FileSelectorWidget("genie_dir", FileSelectionMode.DIRECTORY)
+        self.file_gpt.pathChanged.connect(self._on_gpt_path_changed)
+        self.file_vits.pathChanged.connect(self._on_vits_path_changed)
+        self.layout_model.addRow("模型类型:", self.combo_model_type)
+        self.layout_model.addRow("GPT模型 (.ckpt):", self.file_gpt)
+        self.layout_model.addRow("VITS模型 (.pth):", self.file_vits)
+        self.layout_model.addRow("Genie模型目录:", self.file_genie)
+        group_model.setLayout(self.layout_model)
+        # 参考音频组
+        group_ref = QGroupBox("参考音频")
+        layout_ref = QFormLayout()
+        self.file_ref_audio = FileSelectorWidget("ref_audio", FileSelectionMode.FILE, "Audio (*.wav *.mp3)")
+        self.input_ref_text = QLineEdit()
+        self.input_ref_text.setPlaceholderText("请输入参考音频对应的文本...")
+        btn_play_ref = QPushButton("▶️")
+        btn_play_ref.setFixedWidth(30)
+        btn_play_ref.clicked.connect(self._play_ref_audio)
+        hbox_ref_text = QHBoxLayout()
+        hbox_ref_text.addWidget(self.input_ref_text)
+        hbox_ref_text.addWidget(btn_play_ref)
+        layout_ref.addRow("音频文件:", self.file_ref_audio)
+        layout_ref.addRow("音频文本:", hbox_ref_text)
+        group_ref.setLayout(layout_ref)
+
+        left_column_layout.addWidget(group_model)
+        left_column_layout.addWidget(group_ref)
+        left_column_layout.addStretch()
+
+        # ==================== 右侧列内容 ====================
+
+        # === 推理设置组 ===
+        group_infer = QGroupBox("推理参数")
+        layout_infer = QFormLayout()
+        self.combo_device = MyComboBox()
+        self.combo_device.addItems(["CPU"])
+        self.combo_device.setEnabled(False)
+        self.combo_quality = MyComboBox()
+        self.combo_quality.addItems(["质量优先"])
+        self.combo_quality.setEnabled(False)
+        self.combo_split = MyComboBox()
+        self.combo_split.addItems(["不切分", "智能切分", "按行切分"])
+        self.combo_mode = MyComboBox()
+        self.combo_mode.addItems(["串行推理"])
+        self.combo_mode.setEnabled(False)
+        self.combo_lang = MyComboBox()
+        self.combo_lang.addItems(["Chinese", "English", "Japanese"])
+        layout_infer.addRow("推理设备:\n(重启生效)", self.combo_device)
+        layout_infer.addRow("推理需求:", self.combo_quality)
+        layout_infer.addRow("分句方式:", self.combo_split)
+        layout_infer.addRow("推理模式:", self.combo_mode)
+        layout_infer.addRow("目标语言:", self.combo_lang)
+        group_infer.setLayout(layout_infer)
+
+        # === 自动保存组 ===
+        group_save = QGroupBox("自动保存设置")
+        self.layout_save = QFormLayout()
+        self.combo_save_mode = MyComboBox()
+        self.combo_save_mode.addItems(["禁用自动保存", "保存为单个文件", "保存为多个文件"])
+        self.combo_save_mode.currentIndexChanged.connect(self._update_save_ui_state)
+        default_out_path = os.path.join(os.path.expanduser("~"), "Desktop", "Genie 输出语音")
+        self.file_out_dir = FileSelectorWidget("out_dir", FileSelectionMode.DIRECTORY)
+        self.file_out_dir.set_path(default_out_path)
+        self.layout_save.addRow("保存方式:", self.combo_save_mode)
+        self.layout_save.addRow("输出文件夹:", self.file_out_dir)
+        group_save.setLayout(self.layout_save)
+
+        right_column_layout.addWidget(group_infer)
+        right_column_layout.addWidget(group_save)
+        right_column_layout.addStretch()
+
+        content_layout.addLayout(left_column_layout, 1)
+        content_layout.addLayout(right_column_layout, 1)
+        scroll.setWidget(content_widget)
+        main_layout.addWidget(scroll, 5)
+
+        # ==================== 底部：输入控制 + 输出预览 ====================
+
+        # 创建底部容器 widget
+        bottom_widget = QWidget()
+        bottom_layout = QHBoxLayout(bottom_widget)
+        bottom_layout.setContentsMargins(0, 0, 0, 0)  # 去除边距让它贴合
+
+        # --- 输入控制组 ---
+        group_input = QGroupBox("目标文本")
+        layout_input = QVBoxLayout()
+        self.text_input = MyTextEdit()
+        self.text_input.setPlaceholderText("请输入要合成的目标文本...")
+        self.text_input.setFixedHeight(300)
+        self.btn_start = QPushButton("开始推理")
+        self.btn_start.setFixedHeight(40)
+        self.btn_start.setStyleSheet("""
+            QPushButton {
+                background-color: #4CAF50; 
+                color: white; 
+                font-weight: bold; 
+                border-radius: 5px;
+            }
+            QPushButton:hover { background-color: #45a049; }
+            QPushButton:disabled { background-color: #cccccc; }
+        """)
+        self.btn_start.clicked.connect(self._start_inference)
+        layout_input.addWidget(self.text_input)
+        layout_input.addWidget(self.btn_start)
+        group_input.setLayout(layout_input)
+
+        # --- 输出预览组 ---
+        group_preview = QGroupBox("输出音频预览")
+        preview_layout = QVBoxLayout()
+        self.preview_scroll = QScrollArea()
+        self.preview_scroll.setWidgetResizable(True)
+        self.preview_container = QWidget()
+        self.preview_list_layout = QVBoxLayout(self.preview_container)
+        self.preview_list_layout.setAlignment(Qt.AlignmentFlag.AlignTop)
+        self.preview_scroll.setWidget(self.preview_container)
+        preview_layout.addWidget(self.preview_scroll)
+        group_preview.setLayout(preview_layout)
+
+        bottom_layout.addWidget(group_input, 1)
+        bottom_layout.addWidget(group_preview, 1)
+        main_layout.addWidget(bottom_widget, 3)
+
+        self.apply_ui_state(self.preset_manager.current_preset_data)
+
+    # ==================== 状态管理接口 (供 PresetManager 调用) ====================
+
+    @property
+    def current_preset_name(self) -> str:
+        return self.preset_manager.current_preset_name
+
+    @property
+    def current_preset_data(self) -> dict:
+        return self.preset_manager.current_preset_data
+
+    def get_ui_state(self) -> dict:
+        """收集当前UI状态为字典"""
+        return {
+            "model_type": self.combo_model_type.currentText(),
+            "gpt_path": self.file_gpt.get_path(),
+            "vits_path": self.file_vits.get_path(),
+            "genie_dir": self.file_genie.get_path(),
+            "ref_audio": self.file_ref_audio.get_path(),
+            "ref_text": self.input_ref_text.text(),
+            "device": self.combo_device.currentText().lower(),
+            "quality": self.combo_quality.currentText(),
+            "split": self.combo_split.currentText(),
+            "mode": self.combo_mode.currentText(),
+            "lang": self.combo_lang.currentText(),
+            "save_mode": self.combo_save_mode.currentText(),
+            "out_dir": self.file_out_dir.get_path()
+        }
+
+    @Slot(dict)
+    def apply_ui_state(self, data: dict) -> None:
+        """将字典数据应用到UI"""
+
+        def set_combo_text(combo: MyComboBox, text: str) -> None:
+            index = combo.findText(text)
+            if index >= 0:
+                combo.setCurrentIndex(index)
+
+        set_combo_text(self.combo_model_type, data.get("model_type", ""))
+        self.file_gpt.set_path(data.get("gpt_path", ""), block_signals=True)
+        self.file_vits.set_path(data.get("vits_path", ""), block_signals=True)
+        self.file_genie.set_path(data.get("genie_dir", ""))
+        self.file_ref_audio.set_path(data.get("ref_audio", ""))
+        self.input_ref_text.setText(data.get("ref_text", ""))
+
+        set_combo_text(self.combo_device, data.get("device", ""))
+        set_combo_text(self.combo_quality, data.get("quality", ""))
+        set_combo_text(self.combo_split, data.get("split", ""))
+        set_combo_text(self.combo_mode, data.get("mode", ""))
+        set_combo_text(self.combo_lang, data.get("lang", ""))
+        set_combo_text(self.combo_save_mode, data.get("save_mode", ""))
+
+        self.file_out_dir.set_path(data.get("out_dir", ""))
+
+        # 确保UI显隐状态正确
+        self._update_model_ui_visibility()
+        self._update_save_ui_state()
+
+    # ==================== UI 逻辑处理 ====================
+
+    def _update_model_ui_visibility(self, *args) -> None:
+        """根据模型类型控制文件选择器的显隐"""
+        is_gpt = self.combo_model_type.currentText() == "GPT-SoVITS"
+        self.layout_model.setRowVisible(self.file_gpt, is_gpt)
+        self.layout_model.setRowVisible(self.file_vits, is_gpt)
+        self.layout_model.setRowVisible(self.file_genie, not is_gpt)
+
+    @Slot(str)
+    def _on_gpt_path_changed(self, path: str):
+        if path and os.path.exists(path) and not self.file_vits.get_path():
+            self._try_auto_fill_sibling(path, ".pth", self.file_vits)
+
+    @Slot(str)
+    def _on_vits_path_changed(self, path: str):
+        if path and os.path.exists(path) and not self.file_gpt.get_path():
+            self._try_auto_fill_sibling(path, ".ckpt", self.file_gpt)
+
+    @staticmethod
+    def _try_auto_fill_sibling(current_path: str, target_ext: str, target_widget: FileSelectorWidget):
+        try:
+            directory = os.path.dirname(current_path)
+            if not os.path.exists(directory):
+                return
+            for f in os.listdir(directory):
+                if f.lower().endswith(target_ext.lower()):
+                    full_path = os.path.join(directory, f)
+                    target_widget.set_path(full_path)
+                    print(f"[INFO] 自动关联模型文件: {full_path}")
+                    break
+        except Exception as e:
+            print(f"[WARN] 自动关联文件失败: {e}")
+
+    def _update_save_ui_state(self) -> None:
+        enabled = self.combo_save_mode.currentText() != "禁用自动保存"
+        self.layout_save.setRowVisible(self.file_out_dir, enabled)
+
+    def _play_ref_audio(self) -> None:
+        path = self.file_ref_audio.get_path()
+        if os.path.exists(path):
+            self.player.stop()
+            self.player.play(path)
+        else:
+            QMessageBox.warning(self, "错误", "参考音频文件不存在")
+
+    def _get_split_texts(self, text: str) -> List[str]:
+        method = self.combo_split.currentText()
+        if method == "不切分":
+            return [text]
+        elif method == "按行切分":
+            return [line.strip() for line in text.split('\n') if line.strip()]
+        elif method == "智能切分":
+            return self.splitter.split(text)
+        return [text]
+
+    def _start_inference(self) -> None:
+        text = self.text_input.toPlainText().strip()
+        if not text:
+            QMessageBox.warning(self, "提示", "请输入目标文本")
+            return
+
+        ref_path = self.file_ref_audio.get_path()
+        ref_text = self.input_ref_text.text().strip()
+        if not ref_path or not ref_text:
+            QMessageBox.warning(self, "提示", "请设置参考音频")
+            return
+
+        if not self.file_genie.get_path():
+            QMessageBox.warning(self, "提示", "请选择Genie模型目录")
+            return
+
+        out_dir = self.file_out_dir.get_path()
+        save_mode = self.combo_save_mode.currentText()
+        if not out_dir and save_mode != "禁用自动保存":
+            desktop = os.path.join(os.path.expanduser("~"), "Desktop", "Genie Output")
+            self.file_out_dir.set_path(desktop)
+            print(f"[INFO] 未设置输出文件夹, 将在桌面创建!")
+
+        self.btn_start.setEnabled(False)
+        self.btn_start.setText("推理中...")
+        self._chain_import_model()
+
+    # ==================== 推理链式调用 ====================
+
+    def _chain_import_model(self) -> None:
+        req = {
+            "character_name": self.current_preset_name,
+            "onnx_model_dir": self.file_genie.get_path(),
+            "language": self.combo_lang.currentText(),
+        }
+        worker = InferenceWorker(req, mode="load_character")
+        worker.finished.connect(lambda s, m, d: self._on_import_finished(s, m))
+        worker.start()
+        self.current_worker = worker
+
+    @Slot(bool, str)
+    def _on_import_finished(self, success: bool, msg: str) -> None:
+        if not success:
+            self._reset_ui_state()
+            QMessageBox.critical(self, "模型加载失败", msg)
+            return
+        print(f"[INFO] {msg}")
+        self._chain_set_ref()
+
+    def _chain_set_ref(self) -> None:
+        req = {
+            "character_name": self.current_preset_name,
+            "audio_path": self.file_ref_audio.get_path(),
+            "audio_text": self.input_ref_text.text().strip(),
+            "language": self.combo_lang.currentText(),
+        }
+        worker = InferenceWorker(req, mode="set_reference_audio")
+        worker.finished.connect(lambda s, m, d: self._on_set_ref_finished(s, m))
+        worker.start()
+        self.current_worker = worker
+
+    @Slot(bool, str)
+    def _on_set_ref_finished(self, success: bool, msg: str) -> None:
+        if not success:
+            self._reset_ui_state()
+            QMessageBox.critical(self, "设置参考音频失败", msg)
+            return
+        print(f"[INFO] {msg}")
+        self._chain_tts()
+
+    def _chain_tts(self) -> None:
+        text_full = self.text_input.toPlainText().strip()
+        text_list = self._get_split_texts(text_full)
+
+        print(f"[INFO] 开始串行推理, 分句结果: {text_list}")
+        self._process_serial_step(0, text_list, [], 32000)
+
+    def _process_serial_step(
+            self,
+            index: int,
+            text_list: List[str],
+            audio_accumulator: List[np.ndarray],
+            sample_rate: int
+    ) -> None:
+        # 1. 终止条件：所有句子处理完毕
+        if index >= len(text_list):
+            save_mode = self.combo_save_mode.currentText()
+            out_dir = self.file_out_dir.get_path()
+
+            if out_dir:
+                os.makedirs(out_dir, exist_ok=True)
+
+            if audio_accumulator and save_mode != "保存为多个文件":
+                full_text = ''.join(text_list)
+                full_audio = np.concatenate(audio_accumulator, axis=0)
+                if save_mode == "保存为单个文件":
+                    target_names = generate_output_filenames(folder=out_dir, original_texts=[full_text])
+                    save_path = os.path.join(out_dir, target_names[0])
+                else:  # "禁用自动保存"
+                    save_path = os.path.join(CACHE_DIR, f"{uuid.uuid4().hex}.wav")
+                sf.write(save_path, data=full_audio, samplerate=sample_rate, subtype='PCM_16')
+                self._add_to_preview(full_text, save_path)
+
+            print(f"\n[INFO] 串行推理全部完成，共 {len(text_list)} 句。")
+            self._reset_ui_state()
+            return
+
+        # 2. 递归进行：发起当前句子的请求
+        req = {
+            "character_name": self.current_preset_name,
+            "text": text_list[index],
+        }
+        worker = InferenceWorker(req, mode="tts")
+        worker.finished.connect(
+            lambda s, m, d: self._on_serial_step_finished(s, m, d, index, text_list, audio_accumulator)
+        )
+        worker.start()
+        self.current_worker = worker
+
+    @Slot(bool, str, object, int, object, object, object)
+    def _on_serial_step_finished(
+            self,
+            success: bool,
+            msg: str,
+            return_data: dict,
+            index: int,
+            text_list: List[str],
+            audio_accumulator: List[np.ndarray]
+    ) -> None:
+        if not success:
+            self._reset_ui_state()
+            QMessageBox.critical(self, "推理失败", f"第 {index + 1} 句出错: {msg}")
+            return
+
+        sr = return_data.get("sample_rate", 32000)
+        audio_list = return_data.get("audio_list", [])
+        save_mode = self.combo_save_mode.currentText()
+        out_dir = self.file_out_dir.get_path()
+        if out_dir:
+            os.makedirs(out_dir, exist_ok=True)
+
+        if audio_list:
+            audio_accumulator.append(audio_list[0])
+            if save_mode == "保存为多个文件":
+                target_names = generate_output_filenames(folder=out_dir, original_texts=[text_list[index]])
+                save_path = os.path.join(out_dir, target_names[0])
+                sf.write(save_path, data=audio_list[0], samplerate=sr, subtype='FLOAT')
+                self._add_to_preview(text_list[index], save_path)
+        else:
+            print(f"[WARN] 第 {index + 1} 句返回空音频")
+
+        # 继续处理下一句
+        self._process_serial_step(index + 1, text_list, audio_accumulator, sr)
+
+    def _add_to_preview(self, text: str, path: str) -> None:
+        item = PreviewItemWidget(self.current_gen_id, text, path, self.player)
+        self.preview_list_layout.insertWidget(0, item)
+        self.current_gen_id += 1
+
+    def _reset_ui_state(self) -> None:
+        self.btn_start.setEnabled(True)
+        self.btn_start.setText("开始推理")
+
+    def closeEvent(self, event: QCloseEvent) -> None:
+        # 委托 PresetManager 处理保存逻辑
+        self.preset_manager.shutdown()
+        super().closeEvent(event)
+
+
+class MainWindow(QMainWindow):
+    def __init__(self):
+        super().__init__()
+        self.setWindowTitle("Genie TTS Inference GUI")
+        self.resize(1300, 900)
+
+        # 初始化音频播放器
+        self.player: AudioPlayer = AudioPlayer()
+
+        # 初始化日志重定向
+        self.log_widget: LogWidget = LogWidget()
+        sys.stdout = LogRedirector()
+        sys.stdout.textWritten.connect(self.log_widget.append_log)
+
+        # 初始化主界面
+        self.tabs: QTabWidget = QTabWidget()
+        self.tts_widget = TTSWidget(self.player)
+        self.conv_widget = ConverterWidget()
+
+        self.tabs.addTab(self.log_widget, "GUI Log")
+        self.tabs.addTab(self.tts_widget, "TTS Inference")
+        self.tabs.addTab(self.conv_widget, "Converter")
+        self.tabs.setCurrentIndex(1)  # 默认显示TTS页
+
+        self.setCentralWidget(self.tabs)
+
+    def closeEvent(self, event: QCloseEvent) -> None:
+        if os.path.exists(CACHE_DIR):
+            shutil.rmtree(CACHE_DIR)
+        if hasattr(self, 'player'):
+            self.player.stop()
+        # 线程安全退出后，再恢复 stdout
+        sys.stdout = sys.__stdout__
+        if hasattr(self, 'tts_widget'):
+            self.tts_widget.closeEvent(event)
+        event.accept()
diff --git a/genie_tts/GUI/PresetManager.py b/genie_tts/GUI/PresetManager.py
index 15f25131bfcfb19f10cea7267651241c8f2be497..9f62a8bf841b127b2546d0972c569732c5b87d25 100644
--- a/genie_tts/GUI/PresetManager.py
+++ b/genie_tts/GUI/PresetManager.py
@@ -1,188 +1,188 @@
-import os
-import json
-from typing import Callable, Optional, Dict
-
-from PySide6.QtWidgets import (
-    QWidget, QVBoxLayout, QHBoxLayout, QPushButton, QLabel,
-    QComboBox, QGroupBox, QMessageBox, QInputDialog
-)
-from PySide6.QtCore import Signal, QSettings, Slot
-
-
-class PresetManager(QWidget):
-    # 信号：通知主界面加载数据
-    sig_load_state = Signal(dict)
-
-    def __init__(self,
-                 presets_file: str,
-                 state_getter: Callable[[], dict],
-                 parent: QWidget = None):
-        super().__init__(parent)
-        self.presets_file: str = presets_file
-        self.state_getter: Callable[[], dict] = state_getter
-        self.presets: Dict[str, dict] = {}
-        self.current_preset_name: Optional[str] = None
-        self._init_ui()
-        self._load_from_disk()
-        # 初始化完成后，应用一次初始状态
-        self._apply_initial_preset()
-
-    def _init_ui(self):
-        layout = QVBoxLayout(self)
-        layout.setContentsMargins(0, 0, 0, 0)
-
-        group = QGroupBox("预设管理 (自动保存)")
-        h_layout = QHBoxLayout()
-
-        self.lbl_info = QLabel("当前:")
-
-        self.combo_presets = QComboBox()
-        self.combo_presets.textActivated.connect(self._on_preset_switch_triggered)
-
-        btn_new = QPushButton("新建")
-        btn_new.clicked.connect(self.create_preset)
-
-        # === 新增：重命名按钮 ===
-        btn_rename = QPushButton("重命名")
-        btn_rename.clicked.connect(self.rename_preset)
-        # =====================
-
-        btn_del = QPushButton("删除")
-        btn_del.clicked.connect(self.delete_preset)
-
-        h_layout.addWidget(self.lbl_info)
-        h_layout.addWidget(self.combo_presets, 1)
-        h_layout.addWidget(btn_new)
-        h_layout.addWidget(btn_rename)  # 添加到布局
-        h_layout.addWidget(btn_del)
-
-        group.setLayout(h_layout)
-        layout.addWidget(group)
-
-    @property
-    def current_preset_data(self) -> dict:
-        return self.presets.get(self.current_preset_name, {})
-
-    def _load_from_disk(self):
-        """从磁盘加载 JSON"""
-        if os.path.exists(self.presets_file):
-            try:
-                with open(self.presets_file, 'r', encoding='utf-8') as f:
-                    self.presets = json.load(f)
-            except Exception as e:
-                print(f"[ERROR] 预设文件损坏: {e}")
-                self.presets = {}
-
-        if not self.presets:
-            self.presets = {"Default": {}}
-
-        self.combo_presets.clear()
-        self.combo_presets.addItems(list(self.presets.keys()))
-
-    def _save_to_disk(self):
-        """写入磁盘"""
-        try:
-            os.makedirs(os.path.dirname(self.presets_file), exist_ok=True)
-            with open(self.presets_file, 'w', encoding='utf-8') as f:
-                json.dump(self.presets, f, indent=4, ensure_ascii=False)
-        except Exception as e:
-            print(f"[ERROR] 保存预设失败: {e}")
-
-    def _apply_initial_preset(self):
-        """初始化时恢复上次的选择"""
-        last_used = QSettings("MyTTS", "GUI").value("last_preset", "Default")
-        if last_used not in self.presets and self.combo_presets.count() > 0:
-            last_used = self.combo_presets.itemText(0)
-        self.current_preset_name = last_used
-        self.combo_presets.setCurrentText(last_used)
-        self._load_preset_data(last_used)
-
-    @Slot(str)
-    def _on_preset_switch_triggered(self, new_preset_name: str):
-        if new_preset_name == self.current_preset_name:
-            return
-        if self.current_preset_name:
-            self._save_current_state_to_memory(self.current_preset_name)
-        self._load_preset_data(new_preset_name)
-        self.current_preset_name = new_preset_name
-        QSettings("MyTTS", "GUI").setValue("last_preset", new_preset_name)
-        self._save_to_disk()
-
-    def _save_current_state_to_memory(self, preset_name: str):
-        """调用回调获取主界面状态，并更新到内存字典"""
-        if self.state_getter and preset_name in self.presets:
-            current_data = self.state_getter()
-            self.presets[preset_name] = current_data
-
-    def _load_preset_data(self, preset_name: str):
-        """发送信号给主界面加载数据"""
-        data = self.presets.get(preset_name, {})
-        self.sig_load_state.emit(data)
-        print(f"[INFO] 已加载预设: {preset_name}")
-
-    # ================= 公共接口 =================
-
-    def create_preset(self):
-        """新建预设"""
-        if self.current_preset_name:
-            self._save_current_state_to_memory(self.current_preset_name)
-
-        name, ok = QInputDialog.getText(self, "新建预设", "名称:")
-        if ok and name:
-            if name in self.presets:
-                QMessageBox.warning(self, "警告", "预设名已存在")
-                return
-            self.presets[name] = {}
-            self.combo_presets.addItem(name)
-            self.combo_presets.setCurrentText(name)
-            self.current_preset_name = name
-            self._save_to_disk()
-            self._load_preset_data(name)
-            print(f"[INFO] 已创建预设: {name}")
-
-    def rename_preset(self):
-        """重命名当前预设"""
-        current_name = self.current_preset_name
-        if not current_name:
-            return
-
-        # 先保存当前状态到内存，确保重命名时带走的是最新数据
-        self._save_current_state_to_memory(current_name)
-
-        new_name, ok = QInputDialog.getText(self, "重命名预设", "新名称:", text=current_name)
-        if ok and new_name and new_name != current_name:
-            if new_name in self.presets:
-                QMessageBox.warning(self, "警告", "预设名已存在")
-                return
-            # 迁移数据
-            self.presets[new_name] = self.presets.pop(current_name)
-            self.current_preset_name = new_name
-            # 更新下拉框显示的文本（更新当前选中的这一项）
-            current_index = self.combo_presets.currentIndex()
-            self.combo_presets.setItemText(current_index, new_name)
-            # 更新配置记录
-            QSettings("MyTTS", "GUI").setValue("last_preset", new_name)
-            self._save_to_disk()
-            print(f"[INFO] 已重命名预设: {current_name} -> {new_name}")
-
-    def delete_preset(self):
-        """删除当前预设"""
-        target = self.current_preset_name
-        if len(self.presets) <= 1:
-            QMessageBox.warning(self, "禁止", "至少保留一个预设")
-            return
-
-        if QMessageBox.StandardButton.Yes == QMessageBox.question(self, "确认", f"删除 '{target}'?"):
-            del self.presets[target]
-            self.combo_presets.removeItem(self.combo_presets.currentIndex())
-            new_name = self.combo_presets.currentText()
-            self.current_preset_name = new_name
-            self._load_preset_data(new_name)
-            self._save_to_disk()
-            print(f"[INFO] 已删除预设: {target}")
-
-    def shutdown(self):
-        """关闭时触发"""
-        if self.current_preset_name:
-            self._save_current_state_to_memory(self.current_preset_name)
-        self._save_to_disk()
+import os
+import json
+from typing import Callable, Optional, Dict
+
+from PySide6.QtWidgets import (
+    QWidget, QVBoxLayout, QHBoxLayout, QPushButton, QLabel,
+    QComboBox, QGroupBox, QMessageBox, QInputDialog
+)
+from PySide6.QtCore import Signal, QSettings, Slot
+
+
+class PresetManager(QWidget):
+    # 信号：通知主界面加载数据
+    sig_load_state = Signal(dict)
+
+    def __init__(self,
+                 presets_file: str,
+                 state_getter: Callable[[], dict],
+                 parent: QWidget = None):
+        super().__init__(parent)
+        self.presets_file: str = presets_file
+        self.state_getter: Callable[[], dict] = state_getter
+        self.presets: Dict[str, dict] = {}
+        self.current_preset_name: Optional[str] = None
+        self._init_ui()
+        self._load_from_disk()
+        # 初始化完成后，应用一次初始状态
+        self._apply_initial_preset()
+
+    def _init_ui(self):
+        layout = QVBoxLayout(self)
+        layout.setContentsMargins(0, 0, 0, 0)
+
+        group = QGroupBox("预设管理 (自动保存)")
+        h_layout = QHBoxLayout()
+
+        self.lbl_info = QLabel("当前:")
+
+        self.combo_presets = QComboBox()
+        self.combo_presets.textActivated.connect(self._on_preset_switch_triggered)
+
+        btn_new = QPushButton("新建")
+        btn_new.clicked.connect(self.create_preset)
+
+        # === 新增：重命名按钮 ===
+        btn_rename = QPushButton("重命名")
+        btn_rename.clicked.connect(self.rename_preset)
+        # =====================
+
+        btn_del = QPushButton("删除")
+        btn_del.clicked.connect(self.delete_preset)
+
+        h_layout.addWidget(self.lbl_info)
+        h_layout.addWidget(self.combo_presets, 1)
+        h_layout.addWidget(btn_new)
+        h_layout.addWidget(btn_rename)  # 添加到布局
+        h_layout.addWidget(btn_del)
+
+        group.setLayout(h_layout)
+        layout.addWidget(group)
+
+    @property
+    def current_preset_data(self) -> dict:
+        return self.presets.get(self.current_preset_name, {})
+
+    def _load_from_disk(self):
+        """从磁盘加载 JSON"""
+        if os.path.exists(self.presets_file):
+            try:
+                with open(self.presets_file, 'r', encoding='utf-8') as f:
+                    self.presets = json.load(f)
+            except Exception as e:
+                print(f"[ERROR] 预设文件损坏: {e}")
+                self.presets = {}
+
+        if not self.presets:
+            self.presets = {"Default": {}}
+
+        self.combo_presets.clear()
+        self.combo_presets.addItems(list(self.presets.keys()))
+
+    def _save_to_disk(self):
+        """写入磁盘"""
+        try:
+            os.makedirs(os.path.dirname(self.presets_file), exist_ok=True)
+            with open(self.presets_file, 'w', encoding='utf-8') as f:
+                json.dump(self.presets, f, indent=4, ensure_ascii=False)
+        except Exception as e:
+            print(f"[ERROR] 保存预设失败: {e}")
+
+    def _apply_initial_preset(self):
+        """初始化时恢复上次的选择"""
+        last_used = QSettings("MyTTS", "GUI").value("last_preset", "Default")
+        if last_used not in self.presets and self.combo_presets.count() > 0:
+            last_used = self.combo_presets.itemText(0)
+        self.current_preset_name = last_used
+        self.combo_presets.setCurrentText(last_used)
+        self._load_preset_data(last_used)
+
+    @Slot(str)
+    def _on_preset_switch_triggered(self, new_preset_name: str):
+        if new_preset_name == self.current_preset_name:
+            return
+        if self.current_preset_name:
+            self._save_current_state_to_memory(self.current_preset_name)
+        self._load_preset_data(new_preset_name)
+        self.current_preset_name = new_preset_name
+        QSettings("MyTTS", "GUI").setValue("last_preset", new_preset_name)
+        self._save_to_disk()
+
+    def _save_current_state_to_memory(self, preset_name: str):
+        """调用回调获取主界面状态，并更新到内存字典"""
+        if self.state_getter and preset_name in self.presets:
+            current_data = self.state_getter()
+            self.presets[preset_name] = current_data
+
+    def _load_preset_data(self, preset_name: str):
+        """发送信号给主界面加载数据"""
+        data = self.presets.get(preset_name, {})
+        self.sig_load_state.emit(data)
+        print(f"[INFO] 已加载预设: {preset_name}")
+
+    # ================= 公共接口 =================
+
+    def create_preset(self):
+        """新建预设"""
+        if self.current_preset_name:
+            self._save_current_state_to_memory(self.current_preset_name)
+
+        name, ok = QInputDialog.getText(self, "新建预设", "名称:")
+        if ok and name:
+            if name in self.presets:
+                QMessageBox.warning(self, "警告", "预设名已存在")
+                return
+            self.presets[name] = {}
+            self.combo_presets.addItem(name)
+            self.combo_presets.setCurrentText(name)
+            self.current_preset_name = name
+            self._save_to_disk()
+            self._load_preset_data(name)
+            print(f"[INFO] 已创建预设: {name}")
+
+    def rename_preset(self):
+        """重命名当前预设"""
+        current_name = self.current_preset_name
+        if not current_name:
+            return
+
+        # 先保存当前状态到内存，确保重命名时带走的是最新数据
+        self._save_current_state_to_memory(current_name)
+
+        new_name, ok = QInputDialog.getText(self, "重命名预设", "新名称:", text=current_name)
+        if ok and new_name and new_name != current_name:
+            if new_name in self.presets:
+                QMessageBox.warning(self, "警告", "预设名已存在")
+                return
+            # 迁移数据
+            self.presets[new_name] = self.presets.pop(current_name)
+            self.current_preset_name = new_name
+            # 更新下拉框显示的文本（更新当前选中的这一项）
+            current_index = self.combo_presets.currentIndex()
+            self.combo_presets.setItemText(current_index, new_name)
+            # 更新配置记录
+            QSettings("MyTTS", "GUI").setValue("last_preset", new_name)
+            self._save_to_disk()
+            print(f"[INFO] 已重命名预设: {current_name} -> {new_name}")
+
+    def delete_preset(self):
+        """删除当前预设"""
+        target = self.current_preset_name
+        if len(self.presets) <= 1:
+            QMessageBox.warning(self, "禁止", "至少保留一个预设")
+            return
+
+        if QMessageBox.StandardButton.Yes == QMessageBox.question(self, "确认", f"删除 '{target}'?"):
+            del self.presets[target]
+            self.combo_presets.removeItem(self.combo_presets.currentIndex())
+            new_name = self.combo_presets.currentText()
+            self.current_preset_name = new_name
+            self._load_preset_data(new_name)
+            self._save_to_disk()
+            print(f"[INFO] 已删除预设: {target}")
+
+    def shutdown(self):
+        """关闭时触发"""
+        if self.current_preset_name:
+            self._save_current_state_to_memory(self.current_preset_name)
+        self._save_to_disk()
diff --git a/genie_tts/GUI/ServerManager.py b/genie_tts/GUI/ServerManager.py
index 7928b80aba69632032eff0a25143faebe7c719c1..389216058a9075aa6b20d14f85226bbb81ee0d20 100644
--- a/genie_tts/GUI/ServerManager.py
+++ b/genie_tts/GUI/ServerManager.py
@@ -1,61 +1,61 @@
-from PySide6.QtCore import Signal, QThread
-
-from ..Utils.Shared import context
-from ..Internal import load_character, set_reference_audio
-from ..Core.Inference import tts_client
-from ..ModelManager import model_manager
-
-
-class InferenceWorker(QThread):
-    """执行推理任务的 Worker"""
-    finished = Signal(bool, str, object)  # success, message, data
-
-    def __init__(self, request_data: dict, mode: str):
-        super().__init__()
-        self.req: dict = request_data
-        self.mode: str = mode
-
-    def run(self) -> None:
-        try:
-            if self.mode == 'load_character':
-                load_character(
-                    character_name=self.req['character_name'],
-                    onnx_model_dir=self.req['onnx_model_dir'],
-                    language=self.req['language'],
-                )
-                self.finished.emit(True, "导入角色完成", None)
-
-            elif self.mode == 'set_reference_audio':
-                set_reference_audio(
-                    character_name=self.req['character_name'],
-                    audio_path=self.req['audio_path'],
-                    audio_text=self.req['audio_text'],
-                    language=self.req['language'],
-                )
-                self.finished.emit(True, "设置参考音频完成", None)
-
-            elif self.mode == 'tts':
-                gsv_model = model_manager.get(self.req['character_name'])
-                tts_client.stop_event.clear()
-                audio_chunk = tts_client.tts(
-                    text=self.req['text'],
-                    prompt_audio=context.current_prompt_audio,
-                    encoder=gsv_model.T2S_ENCODER,
-                    first_stage_decoder=gsv_model.T2S_FIRST_STAGE_DECODER,
-                    stage_decoder=gsv_model.T2S_STAGE_DECODER,
-                    vocoder=gsv_model.VITS,
-                    prompt_encoder=gsv_model.PROMPT_ENCODER,
-                    language=gsv_model.LANGUAGE,
-                )
-                audio_chunk = audio_chunk.squeeze()
-                try:
-                    return_data = {
-                        "sample_rate": 32000,
-                        "audio_list": [audio_chunk],
-                    }
-                    self.finished.emit(True, "推理完成", return_data)
-                except Exception as e:
-                    self.finished.emit(False, f"数据解析失败: {e}", None)
-
-        except Exception as e:
-            self.finished.emit(False, f"请求异常: {str(e)}", None)
+from PySide6.QtCore import Signal, QThread
+
+from ..Utils.Shared import context
+from ..Internal import load_character, set_reference_audio
+from ..Core.Inference import tts_client
+from ..ModelManager import model_manager
+
+
+class InferenceWorker(QThread):
+    """执行推理任务的 Worker"""
+    finished = Signal(bool, str, object)  # success, message, data
+
+    def __init__(self, request_data: dict, mode: str):
+        super().__init__()
+        self.req: dict = request_data
+        self.mode: str = mode
+
+    def run(self) -> None:
+        try:
+            if self.mode == 'load_character':
+                load_character(
+                    character_name=self.req['character_name'],
+                    onnx_model_dir=self.req['onnx_model_dir'],
+                    language=self.req['language'],
+                )
+                self.finished.emit(True, "导入角色完成", None)
+
+            elif self.mode == 'set_reference_audio':
+                set_reference_audio(
+                    character_name=self.req['character_name'],
+                    audio_path=self.req['audio_path'],
+                    audio_text=self.req['audio_text'],
+                    language=self.req['language'],
+                )
+                self.finished.emit(True, "设置参考音频完成", None)
+
+            elif self.mode == 'tts':
+                gsv_model = model_manager.get(self.req['character_name'])
+                tts_client.stop_event.clear()
+                audio_chunk = tts_client.tts(
+                    text=self.req['text'],
+                    prompt_audio=context.current_prompt_audio,
+                    encoder=gsv_model.T2S_ENCODER,
+                    first_stage_decoder=gsv_model.T2S_FIRST_STAGE_DECODER,
+                    stage_decoder=gsv_model.T2S_STAGE_DECODER,
+                    vocoder=gsv_model.VITS,
+                    prompt_encoder=gsv_model.PROMPT_ENCODER,
+                    language=gsv_model.LANGUAGE,
+                )
+                audio_chunk = audio_chunk.squeeze()
+                try:
+                    return_data = {
+                        "sample_rate": 32000,
+                        "audio_list": [audio_chunk],
+                    }
+                    self.finished.emit(True, "推理完成", return_data)
+                except Exception as e:
+                    self.finished.emit(False, f"数据解析失败: {e}", None)
+
+        except Exception as e:
+            self.finished.emit(False, f"请求异常: {str(e)}", None)
diff --git a/genie_tts/GUI/Utils.py b/genie_tts/GUI/Utils.py
index c801e2b4635bcba08c40aa64926b9ad408252727..adca79cfa3f40bfdc87841779288c4ac2e725ad7 100644
--- a/genie_tts/GUI/Utils.py
+++ b/genie_tts/GUI/Utils.py
@@ -1,257 +1,257 @@
-import os
-import re
-from enum import Enum
-from datetime import datetime
-import socket
-from typing import List
-
-from PySide6.QtWidgets import (
-    QWidget, QHBoxLayout, QPushButton, QLineEdit, QFileDialog, QMessageBox, QComboBox, QTextEdit
-)
-from PySide6.QtCore import (Qt, Signal, QSettings, Property, QObject, QEvent, QMimeData)
-from PySide6.QtGui import QFont
-
-
-def sanitize_filename(filename: str, replacement: str = '') -> str:
-    """
-    将文本清理为合法的 Windows 文件名。
-
-    Args:
-        filename (str): 原始文件名。
-        replacement (str): 非法字符的替换字符，默认为空，建议使用 "_"。
-
-    Returns:
-        str: 清理后的文件名。
-    """
-    # 1. 去除非法字符
-    # \/:*?"<>| 是标准非法字符
-    # \x00-\x1f 是控制字符 (如换行符、制表符等)，Windows 也不允许
-    cleaned = re.sub(r'[\\/:*?"<>|\x00-\x1f]', replacement, filename)
-
-    # 2. 去除首尾的空格和点
-    # Windows 文件名不能以空格或点结尾，也不能以空格开头(虽然允许但通常不推荐)
-    cleaned = cleaned.strip().rstrip('.')
-
-    # 3. 处理 Windows 保留文件名 (CON, PRN, AUX, NUL, COM1-9, LPT1-9)
-    # 这些名字不论加什么扩展名都是非法的 (例如 con.txt 也是非法的)
-    reserved_names = {
-        "CON", "PRN", "AUX", "NUL",
-        "COM1", "COM2", "COM3", "COM4", "COM5", "COM6", "COM7", "COM8", "COM9",
-        "LPT1", "LPT2", "LPT3", "LPT4", "LPT5", "LPT6", "LPT7", "LPT8", "LPT9"
-    }
-
-    # 如果文件名(全大写)是保留字，或者文件名是保留字+扩展名(如 con.txt)，则加下划线前缀
-    filename_upper = cleaned.upper()
-    file_stem = filename_upper.split('.')[0]  # 获取不带后缀的主文件名
-
-    if filename_upper in reserved_names or file_stem in reserved_names:
-        cleaned = "_" + cleaned
-
-    # 4. 处理空文件名 (如果输入全是乱码被删光了)
-    if not cleaned:
-        cleaned = "unnamed_file"
-
-    # 5. 限制长度 (Windows API 通常限制 255 字符，但在某些路径下更短)
-    cleaned = truncate_text(cleaned)
-
-    return cleaned
-
-
-def is_port_free(port: int) -> bool:
-    """返回端口是否可用"""
-    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
-        try:
-            s.bind(('', port))
-        except OSError:
-            return False
-    return True
-
-
-def find_free_port(preferred: int = 8000) -> int:
-    # 先尝试 preferred 端口
-    if is_port_free(preferred):
-        return preferred
-
-    # 否则用系统自动分配
-    s = socket.socket()
-    s.bind(('', 0))
-    port = s.getsockname()[1]
-    s.close()
-    return port
-
-
-def truncate_text(text: str, max_len: int = 30) -> str:
-    """
-    按指定宽度截断文本
-    """
-    result = ""
-    cur_len = 0
-    for ch in text:
-        char_len = 2 if ('\u4e00' <= ch <= '\u9fff') else 1
-        if cur_len + char_len > max_len:
-            break
-        result += ch
-        cur_len += char_len
-    return result
-
-
-def generate_output_filenames(folder: str, original_texts: List[str]) -> List[str]:
-    """
-    批量生成文件名：
-    输入 original_texts 列表
-    输出等长 filenames 列表
-    """
-    today = datetime.now().strftime("%Y%m%d")
-    pattern = re.compile(rf'^\[{today}]\[(\d{{3}})]')
-
-    # ① 查找当天现存最大编号
-    max_n = 0
-    if os.path.isdir(folder):
-        for name in os.listdir(folder):
-            m = pattern.match(name)
-            if m:
-                n = int(m.group(1))
-                max_n = max(max_n, n)
-
-    filenames = []
-    cur_n = max_n
-
-    # ② 依次生成新文件名
-    for text in original_texts:
-        cur_n += 1
-        n_str = f"{cur_n:03d}"
-
-        cleaned = sanitize_filename(text)
-
-        filename = f"[{today}][{n_str}]{cleaned}.wav"
-        filenames.append(filename)
-
-    return filenames
-
-
-# ==================== 通用组件 ====================
-
-class FileSelectionMode(Enum):
-    FILE = 0
-    DIRECTORY = 1
-
-
-class FileSelectorWidget(QWidget):
-    """一个包含行编辑和浏览按钮的复合控件，支持文件和文件夹选择。（原样引用并稍作适配）"""
-    pathChanged = Signal(str)
-
-    def __init__(
-            self,
-            setting_key: str,
-            selection_mode: FileSelectionMode = FileSelectionMode.DIRECTORY,
-            file_filter: str = "All Files (*)",
-            parent: QWidget = None,
-    ):
-        super().__init__(parent)
-        self.setting_key: str = setting_key
-        self.selection_mode: FileSelectionMode = selection_mode
-        self.file_filter: str = file_filter
-
-        # 使用 QSettings 模拟 STUDIO_SETTINGS
-        self.settings = QSettings("MyTTS", "GUI")
-
-        layout = QHBoxLayout(self)
-        layout.setContentsMargins(0, 0, 0, 0)
-        layout.setSpacing(5)
-
-        self.path_edit: QLineEdit = QLineEdit()
-        self.path_edit.setReadOnly(True)
-        self.path_edit.setPlaceholderText("未选择路径")
-
-        self.browse_button: QPushButton = QPushButton("📁")
-        self.browse_button.setCursor(Qt.CursorShape.PointingHandCursor)
-        default_font = QFont()
-        default_font.setPointSize(10)
-        self.browse_button.setFont(default_font)
-        self.browse_button.setFixedSize(30, 30)
-
-        self.clear_button: QPushButton = QPushButton("❌")
-        self.clear_button.setCursor(Qt.CursorShape.PointingHandCursor)
-        self.clear_button.setFont(default_font)
-        self.clear_button.setFixedSize(30, 30)
-
-        layout.addWidget(self.path_edit)
-        layout.addWidget(self.browse_button)
-        layout.addWidget(self.clear_button)
-
-        self.browse_button.clicked.connect(self._open_dialog)
-        self.clear_button.clicked.connect(self._clear_path)
-        self.path_edit.textChanged.connect(self.pathChanged)
-
-    def _open_dialog(self):
-        path = self.path_edit.text()
-        if path and os.path.exists(path):
-            start_path = path
-        else:
-            # 【修改点】默认路径改为 Desktop
-            desktop_path = os.path.join(os.path.expanduser("~"), "Desktop")
-            start_path = str(self.settings.value(
-                f"last_path_{self.setting_key}", defaultValue=desktop_path
-            ))
-
-        if self.selection_mode == FileSelectionMode.DIRECTORY:
-            selected_path = QFileDialog.getExistingDirectory(
-                self, "选择文件夹", start_path
-            )
-        else:
-            selected_path, _ = QFileDialog.getOpenFileName(
-                self, "选择文件", start_path, self.file_filter
-            )
-
-        if selected_path:
-            self.set_path(selected_path)
-            parent_path = os.path.dirname(selected_path)
-            if parent_path:
-                self.settings.setValue(
-                    f"last_path_{self.setting_key}", parent_path)
-
-    def _clear_path(self):
-        if not self.path_edit.text():
-            return
-        reply = QMessageBox.question(self, '确认', '您确定要清空路径吗？',
-                                     QMessageBox.StandardButton.Yes | QMessageBox.StandardButton.No,
-                                     QMessageBox.StandardButton.No)
-        if reply == QMessageBox.StandardButton.Yes:
-            self.set_path("")
-
-    def get_path(self) -> str:
-        text = self.path_edit.text()
-        # 即使路径不存在（可能是输入时），也返回文本供逻辑判断，或者严格校验
-        return text
-
-    def set_path(self, path: str, block_signals: bool = False):
-        if block_signals:
-            self.path_edit.blockSignals(True)
-        self.path_edit.setText(path)
-        if block_signals:
-            self.path_edit.blockSignals(False)
-
-    path = Property(str, fget=get_path, fset=set_path, notify=pathChanged)  # type: ignore
-
-
-class WheelEventFilter(QObject):
-    def eventFilter(self, obj, event):
-        if event.type() == QEvent.Type.Wheel and isinstance(obj, QComboBox):
-            return True  # 阻止默认滚轮行为
-        return super().eventFilter(obj, event)
-
-
-class MyComboBox(QComboBox):
-    def __init__(self, parent: QWidget = None):
-        super().__init__(parent)
-        self._wheelFilter = WheelEventFilter()
-        self.installEventFilter(self._wheelFilter)
-
-
-class MyTextEdit(QTextEdit):
-    def insertFromMimeData(self, source: QMimeData) -> None:
-        # 仅取纯文本
-        if source.hasText():
-            self.insertPlainText(source.text())
-        else:
-            super().insertFromMimeData(source)
+import os
+import re
+from enum import Enum
+from datetime import datetime
+import socket
+from typing import List
+
+from PySide6.QtWidgets import (
+    QWidget, QHBoxLayout, QPushButton, QLineEdit, QFileDialog, QMessageBox, QComboBox, QTextEdit
+)
+from PySide6.QtCore import (Qt, Signal, QSettings, Property, QObject, QEvent, QMimeData)
+from PySide6.QtGui import QFont
+
+
+def sanitize_filename(filename: str, replacement: str = '') -> str:
+    """
+    将文本清理为合法的 Windows 文件名。
+
+    Args:
+        filename (str): 原始文件名。
+        replacement (str): 非法字符的替换字符，默认为空，建议使用 "_"。
+
+    Returns:
+        str: 清理后的文件名。
+    """
+    # 1. 去除非法字符
+    # \/:*?"<>| 是标准非法字符
+    # \x00-\x1f 是控制字符 (如换行符、制表符等)，Windows 也不允许
+    cleaned = re.sub(r'[\\/:*?"<>|\x00-\x1f]', replacement, filename)
+
+    # 2. 去除首尾的空格和点
+    # Windows 文件名不能以空格或点结尾，也不能以空格开头(虽然允许但通常不推荐)
+    cleaned = cleaned.strip().rstrip('.')
+
+    # 3. 处理 Windows 保留文件名 (CON, PRN, AUX, NUL, COM1-9, LPT1-9)
+    # 这些名字不论加什么扩展名都是非法的 (例如 con.txt 也是非法的)
+    reserved_names = {
+        "CON", "PRN", "AUX", "NUL",
+        "COM1", "COM2", "COM3", "COM4", "COM5", "COM6", "COM7", "COM8", "COM9",
+        "LPT1", "LPT2", "LPT3", "LPT4", "LPT5", "LPT6", "LPT7", "LPT8", "LPT9"
+    }
+
+    # 如果文件名(全大写)是保留字，或者文件名是保留字+扩展名(如 con.txt)，则加下划线前缀
+    filename_upper = cleaned.upper()
+    file_stem = filename_upper.split('.')[0]  # 获取不带后缀的主文件名
+
+    if filename_upper in reserved_names or file_stem in reserved_names:
+        cleaned = "_" + cleaned
+
+    # 4. 处理空文件名 (如果输入全是乱码被删光了)
+    if not cleaned:
+        cleaned = "unnamed_file"
+
+    # 5. 限制长度 (Windows API 通常限制 255 字符，但在某些路径下更短)
+    cleaned = truncate_text(cleaned)
+
+    return cleaned
+
+
+def is_port_free(port: int) -> bool:
+    """返回端口是否可用"""
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+        try:
+            s.bind(('', port))
+        except OSError:
+            return False
+    return True
+
+
+def find_free_port(preferred: int = 8000) -> int:
+    # 先尝试 preferred 端口
+    if is_port_free(preferred):
+        return preferred
+
+    # 否则用系统自动分配
+    s = socket.socket()
+    s.bind(('', 0))
+    port = s.getsockname()[1]
+    s.close()
+    return port
+
+
+def truncate_text(text: str, max_len: int = 30) -> str:
+    """
+    按指定宽度截断文本
+    """
+    result = ""
+    cur_len = 0
+    for ch in text:
+        char_len = 2 if ('\u4e00' <= ch <= '\u9fff') else 1
+        if cur_len + char_len > max_len:
+            break
+        result += ch
+        cur_len += char_len
+    return result
+
+
+def generate_output_filenames(folder: str, original_texts: List[str]) -> List[str]:
+    """
+    批量生成文件名：
+    输入 original_texts 列表
+    输出等长 filenames 列表
+    """
+    today = datetime.now().strftime("%Y%m%d")
+    pattern = re.compile(rf'^\[{today}]\[(\d{{3}})]')
+
+    # ① 查找当天现存最大编号
+    max_n = 0
+    if os.path.isdir(folder):
+        for name in os.listdir(folder):
+            m = pattern.match(name)
+            if m:
+                n = int(m.group(1))
+                max_n = max(max_n, n)
+
+    filenames = []
+    cur_n = max_n
+
+    # ② 依次生成新文件名
+    for text in original_texts:
+        cur_n += 1
+        n_str = f"{cur_n:03d}"
+
+        cleaned = sanitize_filename(text)
+
+        filename = f"[{today}][{n_str}]{cleaned}.wav"
+        filenames.append(filename)
+
+    return filenames
+
+
+# ==================== 通用组件 ====================
+
+class FileSelectionMode(Enum):
+    FILE = 0
+    DIRECTORY = 1
+
+
+class FileSelectorWidget(QWidget):
+    """一个包含行编辑和浏览按钮的复合控件，支持文件和文件夹选择。（原样引用并稍作适配）"""
+    pathChanged = Signal(str)
+
+    def __init__(
+            self,
+            setting_key: str,
+            selection_mode: FileSelectionMode = FileSelectionMode.DIRECTORY,
+            file_filter: str = "All Files (*)",
+            parent: QWidget = None,
+    ):
+        super().__init__(parent)
+        self.setting_key: str = setting_key
+        self.selection_mode: FileSelectionMode = selection_mode
+        self.file_filter: str = file_filter
+
+        # 使用 QSettings 模拟 STUDIO_SETTINGS
+        self.settings = QSettings("MyTTS", "GUI")
+
+        layout = QHBoxLayout(self)
+        layout.setContentsMargins(0, 0, 0, 0)
+        layout.setSpacing(5)
+
+        self.path_edit: QLineEdit = QLineEdit()
+        self.path_edit.setReadOnly(True)
+        self.path_edit.setPlaceholderText("未选择路径")
+
+        self.browse_button: QPushButton = QPushButton("📁")
+        self.browse_button.setCursor(Qt.CursorShape.PointingHandCursor)
+        default_font = QFont()
+        default_font.setPointSize(10)
+        self.browse_button.setFont(default_font)
+        self.browse_button.setFixedSize(30, 30)
+
+        self.clear_button: QPushButton = QPushButton("❌")
+        self.clear_button.setCursor(Qt.CursorShape.PointingHandCursor)
+        self.clear_button.setFont(default_font)
+        self.clear_button.setFixedSize(30, 30)
+
+        layout.addWidget(self.path_edit)
+        layout.addWidget(self.browse_button)
+        layout.addWidget(self.clear_button)
+
+        self.browse_button.clicked.connect(self._open_dialog)
+        self.clear_button.clicked.connect(self._clear_path)
+        self.path_edit.textChanged.connect(self.pathChanged)
+
+    def _open_dialog(self):
+        path = self.path_edit.text()
+        if path and os.path.exists(path):
+            start_path = path
+        else:
+            # 【修改点】默认路径改为 Desktop
+            desktop_path = os.path.join(os.path.expanduser("~"), "Desktop")
+            start_path = str(self.settings.value(
+                f"last_path_{self.setting_key}", defaultValue=desktop_path
+            ))
+
+        if self.selection_mode == FileSelectionMode.DIRECTORY:
+            selected_path = QFileDialog.getExistingDirectory(
+                self, "选择文件夹", start_path
+            )
+        else:
+            selected_path, _ = QFileDialog.getOpenFileName(
+                self, "选择文件", start_path, self.file_filter
+            )
+
+        if selected_path:
+            self.set_path(selected_path)
+            parent_path = os.path.dirname(selected_path)
+            if parent_path:
+                self.settings.setValue(
+                    f"last_path_{self.setting_key}", parent_path)
+
+    def _clear_path(self):
+        if not self.path_edit.text():
+            return
+        reply = QMessageBox.question(self, '确认', '您确定要清空路径吗？',
+                                     QMessageBox.StandardButton.Yes | QMessageBox.StandardButton.No,
+                                     QMessageBox.StandardButton.No)
+        if reply == QMessageBox.StandardButton.Yes:
+            self.set_path("")
+
+    def get_path(self) -> str:
+        text = self.path_edit.text()
+        # 即使路径不存在（可能是输入时），也返回文本供逻辑判断，或者严格校验
+        return text
+
+    def set_path(self, path: str, block_signals: bool = False):
+        if block_signals:
+            self.path_edit.blockSignals(True)
+        self.path_edit.setText(path)
+        if block_signals:
+            self.path_edit.blockSignals(False)
+
+    path = Property(str, fget=get_path, fset=set_path, notify=pathChanged)  # type: ignore
+
+
+class WheelEventFilter(QObject):
+    def eventFilter(self, obj, event):
+        if event.type() == QEvent.Type.Wheel and isinstance(obj, QComboBox):
+            return True  # 阻止默认滚轮行为
+        return super().eventFilter(obj, event)
+
+
+class MyComboBox(QComboBox):
+    def __init__(self, parent: QWidget = None):
+        super().__init__(parent)
+        self._wheelFilter = WheelEventFilter()
+        self.installEventFilter(self._wheelFilter)
+
+
+class MyTextEdit(QTextEdit):
+    def insertFromMimeData(self, source: QMimeData) -> None:
+        # 仅取纯文本
+        if source.hasText():
+            self.insertPlainText(source.text())
+        else:
+            super().insertFromMimeData(source)
diff --git a/genie_tts/GetPhonesAndBert.py b/genie_tts/GetPhonesAndBert.py
index c181a948a162341643864c289fbdc81b6a281850..3594ab26e2fd1ad673c5486a08161a4ffdbcfbfd 100644
--- a/genie_tts/GetPhonesAndBert.py
+++ b/genie_tts/GetPhonesAndBert.py
@@ -1,34 +1,34 @@
-import numpy as np
-from typing import Tuple
-from .Utils.Constants import BERT_FEATURE_DIM
-from .ModelManager import model_manager
-
-
-def get_phones_and_bert(prompt_text: str, language: str = 'japanese') -> Tuple[np.ndarray, np.ndarray]:
-    if language.lower() == 'english':
-        from .G2P.English.EnglishG2P import english_to_phones
-        phones = english_to_phones(prompt_text)
-        text_bert = np.zeros((len(phones), BERT_FEATURE_DIM), dtype=np.float32)
-    elif language.lower() == 'chinese':
-        from .G2P.Chinese.ChineseG2P import chinese_to_phones
-        text_clean, _, phones, word2ph = chinese_to_phones(prompt_text)
-        if model_manager.load_roberta_model():
-            encoded = model_manager.roberta_tokenizer.encode(text_clean)
-            input_ids = np.array([encoded.ids], dtype=np.int64)
-            attention_mask = np.array([encoded.attention_mask], dtype=np.int64)
-            ort_inputs = {
-                'input_ids': input_ids,
-                'attention_mask': attention_mask,
-                'repeats': np.array(word2ph, dtype=np.int64),
-            }
-            outputs = model_manager.roberta_model.run(None, ort_inputs)
-            text_bert = outputs[0].astype(np.float32)
-        else:
-            text_bert = np.zeros((len(phones), BERT_FEATURE_DIM), dtype=np.float32)
-    else:
-        from .G2P.Japanese.JapaneseG2P import japanese_to_phones
-        phones = japanese_to_phones(prompt_text)
-        text_bert = np.zeros((len(phones), BERT_FEATURE_DIM), dtype=np.float32)
-
-    phones_seq = np.array([phones], dtype=np.int64)
-    return phones_seq, text_bert
+import numpy as np
+from typing import Tuple
+from .Utils.Constants import BERT_FEATURE_DIM
+from .ModelManager import model_manager
+
+
+def get_phones_and_bert(prompt_text: str, language: str = 'japanese') -> Tuple[np.ndarray, np.ndarray]:
+    if language.lower() == 'english':
+        from .G2P.English.EnglishG2P import english_to_phones
+        phones = english_to_phones(prompt_text)
+        text_bert = np.zeros((len(phones), BERT_FEATURE_DIM), dtype=np.float32)
+    elif language.lower() == 'chinese':
+        from .G2P.Chinese.ChineseG2P import chinese_to_phones
+        text_clean, _, phones, word2ph = chinese_to_phones(prompt_text)
+        if model_manager.load_roberta_model():
+            encoded = model_manager.roberta_tokenizer.encode(text_clean)
+            input_ids = np.array([encoded.ids], dtype=np.int64)
+            attention_mask = np.array([encoded.attention_mask], dtype=np.int64)
+            ort_inputs = {
+                'input_ids': input_ids,
+                'attention_mask': attention_mask,
+                'repeats': np.array(word2ph, dtype=np.int64),
+            }
+            outputs = model_manager.roberta_model.run(None, ort_inputs)
+            text_bert = outputs[0].astype(np.float32)
+        else:
+            text_bert = np.zeros((len(phones), BERT_FEATURE_DIM), dtype=np.float32)
+    else:
+        from .G2P.Japanese.JapaneseG2P import japanese_to_phones
+        phones = japanese_to_phones(prompt_text)
+        text_bert = np.zeros((len(phones), BERT_FEATURE_DIM), dtype=np.float32)
+
+    phones_seq = np.array([phones], dtype=np.int64)
+    return phones_seq, text_bert
diff --git a/genie_tts/ModelManager.py b/genie_tts/ModelManager.py
index f3e84bf65f6f378922fe88191e5ba1561a848e48..69eabbfc781e3286f35040dad42bb9f02c98cd9c 100644
--- a/genie_tts/ModelManager.py
+++ b/genie_tts/ModelManager.py
@@ -1,324 +1,324 @@
-"""
-不再新建 .bin 文件。
-修改后内存: 6448 MB
-修改前内存: 5952 MB
-"""
-
-import gc
-import logging
-import os
-from dataclasses import dataclass
-from typing import Optional, List, Dict
-
-import numpy as np
-import onnx
-import onnxruntime
-from onnxruntime import InferenceSession
-from tokenizers import Tokenizer
-
-from .Core.Resources import (HUBERT_MODEL_DIR, SV_MODEL, ROBERTA_MODEL_DIR)
-from .Utils.Utils import LRUCacheDict
-
-onnxruntime.set_default_logger_severity(3)
-logger = logging.getLogger(__name__)
-
-
-class GSVModelFile:
-    T2S_ENCODER_FP32: str = 't2s_encoder_fp32.onnx'
-
-    T2S_FIRST_STAGE_DECODER_FP32: str = 't2s_first_stage_decoder_fp32.onnx'
-    T2S_FIRST_STAGE_DECODER_FP16: str = 't2s_first_stage_decoder_fp16.onnx'
-    T2S_STAGE_DECODER_FP32: str = 't2s_stage_decoder_fp32.onnx'
-    T2S_STAGE_DECODER_FP16: str = 't2s_stage_decoder_fp16.onnx'
-    T2S_DECODER_WEIGHT_FP16: str = 't2s_shared_fp16.bin'
-
-    VITS_FP32: str = 'vits_fp32.onnx'
-    VITS_WEIGHT_FP16: str = 'vits_fp16.bin'
-
-    PROMPT_ENCODER: str = 'prompt_encoder_fp32.onnx'
-    PROMPT_ENCODER_WEIGHT_FP16: str = 'prompt_encoder_fp16.bin'
-
-    HUBERT_MODEL = os.path.join(HUBERT_MODEL_DIR, "chinese-hubert-base.onnx")
-    HUBERT_MODEL_WEIGHT_FP16 = os.path.join(HUBERT_MODEL_DIR, "chinese-hubert-base_weights_fp16.bin")
-
-    ROBERTA_MODEL = os.path.join(ROBERTA_MODEL_DIR, 'RoBERTa.onnx')
-    ROBERTA_TOKENIZER = os.path.join(ROBERTA_MODEL_DIR, 'roberta_tokenizer')
-
-
-@dataclass
-class GSVModel:
-    LANGUAGE: str
-    T2S_ENCODER: InferenceSession
-    T2S_FIRST_STAGE_DECODER: InferenceSession
-    T2S_STAGE_DECODER: InferenceSession
-    VITS: InferenceSession
-    PROMPT_ENCODER: Optional[InferenceSession] = None
-    PROMPT_ENCODER_PATH: Optional[str] = None
-
-
-def load_session_with_fp16_conversion(
-        onnx_path: str,
-        fp16_bin_path: str,
-        providers: List[str],
-        sess_options: Optional[onnxruntime.SessionOptions] = None
-) -> InferenceSession:
-    """
-    通用函数：读取 ONNX 和 FP16 权重文件，在内存中将权重转换为 FP32，
-    注入到 ONNX 模型中并加载 InferenceSession，不产生临时文件。
-    """
-    if not os.path.exists(onnx_path):
-        raise FileNotFoundError(f"ONNX Model not found: {onnx_path}")
-    if not os.path.exists(fp16_bin_path):
-        raise FileNotFoundError(f"FP16 Weight file not found: {fp16_bin_path}")
-
-    model_proto = onnx.load(onnx_path, load_external_data=False)
-    fp16_data = np.fromfile(fp16_bin_path, dtype=np.float16)
-    fp32_data = fp16_data.astype(np.float32)
-    fp32_bytes = fp32_data.tobytes()
-
-    # 遍历并修补模型中的 External Data Initializers
-    for tensor in model_proto.graph.initializer:
-        # 检查该 Tensor 是否使用外部数据
-        if tensor.data_location == onnx.TensorProto.EXTERNAL:
-            offset = 0
-            length = 0
-            # 解析外部数据信息
-            for entry in tensor.external_data:
-                if entry.key == 'offset':
-                    offset = int(entry.value)
-                elif entry.key == 'length':
-                    length = int(entry.value)
-
-            if offset + length > len(fp32_bytes):
-                logger.warning(
-                    f"Tensor {tensor.name} requested a data range that exceeds the size of the provided bin file. "
-                    f"Offset: {offset}, Length: {length}, Buffer: {len(fp32_bytes)}"
-                )
-                continue
-
-            tensor_data = fp32_bytes[offset: offset + length]
-            tensor.raw_data = tensor_data
-
-            del tensor.external_data[:]
-            tensor.data_location = onnx.TensorProto.DEFAULT
-
-    try:
-        session = InferenceSession(
-            model_proto.SerializeToString(),
-            providers=providers,
-            sess_options=sess_options
-        )
-        return session
-    except Exception as e:
-        logger.error(f"Failed to load in-memory model {os.path.basename(onnx_path)}: {e}")
-        raise e
-
-
-class ModelManager:
-    def __init__(self):
-        capacity_str = os.getenv('Max_Cached_Character_Models', '3')
-        self.character_to_model: Dict[str, Dict[str, Optional[InferenceSession]]] = LRUCacheDict(
-            capacity=int(capacity_str)
-        )
-        self.character_to_language: Dict[str, str] = {}
-        self.character_model_paths: Dict[str, str] = {}
-        self.providers = ["CPUExecutionProvider"]
-
-        self.cn_hubert: Optional[InferenceSession] = None
-        self.speaker_verification_model: Optional[InferenceSession] = None
-        self.roberta_model: Optional[InferenceSession] = None
-        self.roberta_tokenizer: Optional[Tokenizer] = None
-
-    def load_roberta_model(self, model_path: str = GSVModelFile.ROBERTA_MODEL) -> bool:
-        if self.roberta_model is not None:
-            return True
-        if not os.path.exists(model_path):
-            # logger.warning(f'RoBERTa model does not exist: {model_path}. BERT features will not be used.')
-            return False
-        try:
-            self.roberta_model = onnxruntime.InferenceSession(
-                model_path,
-                providers=self.providers,
-            )
-            self.roberta_tokenizer = Tokenizer.from_file(
-                os.path.join(GSVModelFile.ROBERTA_TOKENIZER, 'tokenizer.json')
-            )
-            logger.info(f"Successfully loaded RoBERTa model.")
-            return True
-        except Exception as e:
-            logger.error(
-                f"Error: Failed to load ONNX model '{GSVModelFile.ROBERTA_MODEL}'.\n"
-                f"Details: {e}"
-            )
-        return False
-
-    def load_sv_model(self, model_path: str = SV_MODEL) -> bool:
-        if self.speaker_verification_model is not None:
-            return True
-        try:
-            self.speaker_verification_model = onnxruntime.InferenceSession(
-                model_path,
-                providers=self.providers,
-            )
-            logger.info(f"Successfully loaded Speaker Verification model.")
-            return True
-        except Exception as e:
-            logger.error(
-                f"Error: Failed to load ONNX model '{SV_MODEL}'.\n"
-                f"Details: {e}"
-            )
-        return False
-
-    def load_cn_hubert(self, model_path: str = GSVModelFile.HUBERT_MODEL) -> bool:
-        if self.cn_hubert is not None:
-            return True
-        try:
-            # Hubert 也应用内存转换逻辑
-            if model_path == GSVModelFile.HUBERT_MODEL and os.path.exists(GSVModelFile.HUBERT_MODEL_WEIGHT_FP16):
-                self.cn_hubert = load_session_with_fp16_conversion(
-                    model_path,
-                    GSVModelFile.HUBERT_MODEL_WEIGHT_FP16,
-                    self.providers
-                )
-            else:
-                self.cn_hubert = onnxruntime.InferenceSession(
-                    model_path,
-                    providers=self.providers,
-                )
-            logger.info("Successfully loaded CN_HuBERT model.")
-            return True
-        except Exception as e:
-            logger.error(
-                f"Error: Failed to load ONNX model '{GSVModelFile.HUBERT_MODEL}'.\n"
-                f"Details: {e}"
-            )
-        return False
-
-    def get(self, character_name: str) -> Optional[GSVModel]:
-        character_name = character_name.lower()
-        language = self.character_to_language.get(character_name, 'Japanese')
-        if character_name in self.character_to_model:
-            model_map: dict = self.character_to_model[character_name]
-            # 简化获取逻辑
-            t2s_first_stage_decoder = model_map.get(GSVModelFile.T2S_FIRST_STAGE_DECODER_FP32) or \
-                                      model_map.get(GSVModelFile.T2S_FIRST_STAGE_DECODER_FP16)
-            t2s_stage_decoder = model_map.get(GSVModelFile.T2S_STAGE_DECODER_FP32) or \
-                                model_map.get(GSVModelFile.T2S_STAGE_DECODER_FP16)
-            prompt_encoder_path = os.path.join(self.character_model_paths[character_name], GSVModelFile.PROMPT_ENCODER)
-
-            return GSVModel(
-                LANGUAGE=language,
-                T2S_ENCODER=model_map[GSVModelFile.T2S_ENCODER_FP32],
-                T2S_FIRST_STAGE_DECODER=t2s_first_stage_decoder,
-                T2S_STAGE_DECODER=t2s_stage_decoder,
-                VITS=model_map[GSVModelFile.VITS_FP32],
-                PROMPT_ENCODER=model_map[GSVModelFile.PROMPT_ENCODER],
-                PROMPT_ENCODER_PATH=prompt_encoder_path,
-            )
-        if character_name in self.character_model_paths:
-            model_dir = self.character_model_paths[character_name]
-            if self.load_character(character_name, model_dir, language=language):
-                return self.get(character_name)
-            else:
-                del self.character_model_paths[character_name]
-                return None
-        return None
-
-    def has_character(self, character_name: str) -> bool:
-        character_name = character_name.lower()
-        return character_name in self.character_model_paths
-
-    def load_character(
-            self,
-            character_name: str,
-            model_dir: str,
-            language: str,
-    ) -> bool:
-        """
-        加载角色模型，如果需要，在内存中动态转换 FP16 权重。
-        """
-        character_name = character_name.lower()
-        if character_name in self.character_to_model:
-            _ = self.character_to_model[character_name]
-            return True
-
-        model_dict: Dict[str, Optional[InferenceSession]] = {}
-
-        # 定义 ONNX 文件到 FP16 Bin 文件的映射关系
-        onnx_to_fp16_map = {
-            GSVModelFile.T2S_FIRST_STAGE_DECODER_FP32: GSVModelFile.T2S_DECODER_WEIGHT_FP16,
-            GSVModelFile.T2S_STAGE_DECODER_FP32: GSVModelFile.T2S_DECODER_WEIGHT_FP16,
-            GSVModelFile.VITS_FP32: GSVModelFile.VITS_WEIGHT_FP16,
-            GSVModelFile.PROMPT_ENCODER: GSVModelFile.PROMPT_ENCODER_WEIGHT_FP16
-        }
-
-        # 确定需要加载的模型列表
-        model_files_to_load = [
-            GSVModelFile.T2S_ENCODER_FP32,
-            GSVModelFile.VITS_FP32,
-            GSVModelFile.PROMPT_ENCODER,
-        ]
-
-        fp32_decoders = [GSVModelFile.T2S_FIRST_STAGE_DECODER_FP32, GSVModelFile.T2S_STAGE_DECODER_FP32]
-        model_files_to_load.extend(fp32_decoders)
-
-        try:
-            for model_file in model_files_to_load:
-                model_path = os.path.normpath(os.path.join(model_dir, model_file))
-
-                # 设置 Session Options
-                sess_options = onnxruntime.SessionOptions()
-                sess_options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
-
-                if os.path.exists(model_path):
-                    fp16_bin_name = onnx_to_fp16_map.get(model_file)
-                    fp16_bin_path = os.path.join(model_dir, fp16_bin_name) if fp16_bin_name else None
-
-                    if fp16_bin_path and os.path.exists(fp16_bin_path):
-                        model_dict[model_file] = load_session_with_fp16_conversion(
-                            model_path, fp16_bin_path, self.providers, sess_options
-                        )
-                    else:
-                        model_dict[model_file] = onnxruntime.InferenceSession(
-                            model_path,
-                            providers=self.providers,
-                            sess_options=sess_options,
-                        )
-                elif model_file == GSVModelFile.PROMPT_ENCODER:
-                    model_dict[model_file] = None
-                else:
-                    raise FileNotFoundError(f'文件 {model_path} 不存在！')
-
-            # 日志信息
-            is_v2pp = model_dict[GSVModelFile.PROMPT_ENCODER] is not None
-            logger.info(
-                f"Character {character_name.capitalize()} loaded successfully.\n"
-                f"- Model Path: {model_dir}\n"
-                f"- Model Type: {'V2ProPlus' if is_v2pp else 'V2'}"
-            )
-
-            self.character_to_model[character_name] = model_dict
-            self.character_to_language[character_name] = language
-            self.character_model_paths[character_name] = model_dir
-            return True
-
-        except Exception as e:
-            logger.error(
-                f"Error: Failed to load ONNX model '{model_dir}'.\n"
-                f"Details: {e}"
-            )
-            return False
-
-    def remove_all_character(self) -> None:
-        self.character_to_model.clear()
-        gc.collect()
-
-    def remove_character(self, character_name: str) -> None:
-        character_name = character_name.lower()
-        if character_name in self.character_to_model:
-            del self.character_to_model[character_name]
-            gc.collect()
-            logger.info(f"Character {character_name.capitalize()} removed successfully.")
-
-
-model_manager: ModelManager = ModelManager()
+"""
+不再新建 .bin 文件。
+修改后内存: 6448 MB
+修改前内存: 5952 MB
+"""
+
+import gc
+import logging
+import os
+from dataclasses import dataclass
+from typing import Optional, List, Dict
+
+import numpy as np
+import onnx
+import onnxruntime
+from onnxruntime import InferenceSession
+from tokenizers import Tokenizer
+
+from .Core.Resources import (HUBERT_MODEL_DIR, SV_MODEL, ROBERTA_MODEL_DIR)
+from .Utils.Utils import LRUCacheDict
+
+onnxruntime.set_default_logger_severity(3)
+logger = logging.getLogger(__name__)
+
+
+class GSVModelFile:
+    T2S_ENCODER_FP32: str = 't2s_encoder_fp32.onnx'
+
+    T2S_FIRST_STAGE_DECODER_FP32: str = 't2s_first_stage_decoder_fp32.onnx'
+    T2S_FIRST_STAGE_DECODER_FP16: str = 't2s_first_stage_decoder_fp16.onnx'
+    T2S_STAGE_DECODER_FP32: str = 't2s_stage_decoder_fp32.onnx'
+    T2S_STAGE_DECODER_FP16: str = 't2s_stage_decoder_fp16.onnx'
+    T2S_DECODER_WEIGHT_FP16: str = 't2s_shared_fp16.bin'
+
+    VITS_FP32: str = 'vits_fp32.onnx'
+    VITS_WEIGHT_FP16: str = 'vits_fp16.bin'
+
+    PROMPT_ENCODER: str = 'prompt_encoder_fp32.onnx'
+    PROMPT_ENCODER_WEIGHT_FP16: str = 'prompt_encoder_fp16.bin'
+
+    HUBERT_MODEL = os.path.join(HUBERT_MODEL_DIR, "chinese-hubert-base.onnx")
+    HUBERT_MODEL_WEIGHT_FP16 = os.path.join(HUBERT_MODEL_DIR, "chinese-hubert-base_weights_fp16.bin")
+
+    ROBERTA_MODEL = os.path.join(ROBERTA_MODEL_DIR, 'RoBERTa.onnx')
+    ROBERTA_TOKENIZER = os.path.join(ROBERTA_MODEL_DIR, 'roberta_tokenizer')
+
+
+@dataclass
+class GSVModel:
+    LANGUAGE: str
+    T2S_ENCODER: InferenceSession
+    T2S_FIRST_STAGE_DECODER: InferenceSession
+    T2S_STAGE_DECODER: InferenceSession
+    VITS: InferenceSession
+    PROMPT_ENCODER: Optional[InferenceSession] = None
+    PROMPT_ENCODER_PATH: Optional[str] = None
+
+
+def load_session_with_fp16_conversion(
+        onnx_path: str,
+        fp16_bin_path: str,
+        providers: List[str],
+        sess_options: Optional[onnxruntime.SessionOptions] = None
+) -> InferenceSession:
+    """
+    通用函数：读取 ONNX 和 FP16 权重文件，在内存中将权重转换为 FP32，
+    注入到 ONNX 模型中并加载 InferenceSession，不产生临时文件。
+    """
+    if not os.path.exists(onnx_path):
+        raise FileNotFoundError(f"ONNX Model not found: {onnx_path}")
+    if not os.path.exists(fp16_bin_path):
+        raise FileNotFoundError(f"FP16 Weight file not found: {fp16_bin_path}")
+
+    model_proto = onnx.load(onnx_path, load_external_data=False)
+    fp16_data = np.fromfile(fp16_bin_path, dtype=np.float16)
+    fp32_data = fp16_data.astype(np.float32)
+    fp32_bytes = fp32_data.tobytes()
+
+    # 遍历并修补模型中的 External Data Initializers
+    for tensor in model_proto.graph.initializer:
+        # 检查该 Tensor 是否使用外部数据
+        if tensor.data_location == onnx.TensorProto.EXTERNAL:
+            offset = 0
+            length = 0
+            # 解析外部数据信息
+            for entry in tensor.external_data:
+                if entry.key == 'offset':
+                    offset = int(entry.value)
+                elif entry.key == 'length':
+                    length = int(entry.value)
+
+            if offset + length > len(fp32_bytes):
+                logger.warning(
+                    f"Tensor {tensor.name} requested a data range that exceeds the size of the provided bin file. "
+                    f"Offset: {offset}, Length: {length}, Buffer: {len(fp32_bytes)}"
+                )
+                continue
+
+            tensor_data = fp32_bytes[offset: offset + length]
+            tensor.raw_data = tensor_data
+
+            del tensor.external_data[:]
+            tensor.data_location = onnx.TensorProto.DEFAULT
+
+    try:
+        session = InferenceSession(
+            model_proto.SerializeToString(),
+            providers=providers,
+            sess_options=sess_options
+        )
+        return session
+    except Exception as e:
+        logger.error(f"Failed to load in-memory model {os.path.basename(onnx_path)}: {e}")
+        raise e
+
+
+class ModelManager:
+    def __init__(self):
+        capacity_str = os.getenv('Max_Cached_Character_Models', '3')
+        self.character_to_model: Dict[str, Dict[str, Optional[InferenceSession]]] = LRUCacheDict(
+            capacity=int(capacity_str)
+        )
+        self.character_to_language: Dict[str, str] = {}
+        self.character_model_paths: Dict[str, str] = {}
+        self.providers = ["CPUExecutionProvider"]
+
+        self.cn_hubert: Optional[InferenceSession] = None
+        self.speaker_verification_model: Optional[InferenceSession] = None
+        self.roberta_model: Optional[InferenceSession] = None
+        self.roberta_tokenizer: Optional[Tokenizer] = None
+
+    def load_roberta_model(self, model_path: str = GSVModelFile.ROBERTA_MODEL) -> bool:
+        if self.roberta_model is not None:
+            return True
+        if not os.path.exists(model_path):
+            # logger.warning(f'RoBERTa model does not exist: {model_path}. BERT features will not be used.')
+            return False
+        try:
+            self.roberta_model = onnxruntime.InferenceSession(
+                model_path,
+                providers=self.providers,
+            )
+            self.roberta_tokenizer = Tokenizer.from_file(
+                os.path.join(GSVModelFile.ROBERTA_TOKENIZER, 'tokenizer.json')
+            )
+            logger.info(f"Successfully loaded RoBERTa model.")
+            return True
+        except Exception as e:
+            logger.error(
+                f"Error: Failed to load ONNX model '{GSVModelFile.ROBERTA_MODEL}'.\n"
+                f"Details: {e}"
+            )
+        return False
+
+    def load_sv_model(self, model_path: str = SV_MODEL) -> bool:
+        if self.speaker_verification_model is not None:
+            return True
+        try:
+            self.speaker_verification_model = onnxruntime.InferenceSession(
+                model_path,
+                providers=self.providers,
+            )
+            logger.info(f"Successfully loaded Speaker Verification model.")
+            return True
+        except Exception as e:
+            logger.error(
+                f"Error: Failed to load ONNX model '{SV_MODEL}'.\n"
+                f"Details: {e}"
+            )
+        return False
+
+    def load_cn_hubert(self, model_path: str = GSVModelFile.HUBERT_MODEL) -> bool:
+        if self.cn_hubert is not None:
+            return True
+        try:
+            # Hubert 也应用内存转换逻辑
+            if model_path == GSVModelFile.HUBERT_MODEL and os.path.exists(GSVModelFile.HUBERT_MODEL_WEIGHT_FP16):
+                self.cn_hubert = load_session_with_fp16_conversion(
+                    model_path,
+                    GSVModelFile.HUBERT_MODEL_WEIGHT_FP16,
+                    self.providers
+                )
+            else:
+                self.cn_hubert = onnxruntime.InferenceSession(
+                    model_path,
+                    providers=self.providers,
+                )
+            logger.info("Successfully loaded CN_HuBERT model.")
+            return True
+        except Exception as e:
+            logger.error(
+                f"Error: Failed to load ONNX model '{GSVModelFile.HUBERT_MODEL}'.\n"
+                f"Details: {e}"
+            )
+        return False
+
+    def get(self, character_name: str) -> Optional[GSVModel]:
+        character_name = character_name.lower()
+        language = self.character_to_language.get(character_name, 'Japanese')
+        if character_name in self.character_to_model:
+            model_map: dict = self.character_to_model[character_name]
+            # 简化获取逻辑
+            t2s_first_stage_decoder = model_map.get(GSVModelFile.T2S_FIRST_STAGE_DECODER_FP32) or \
+                                      model_map.get(GSVModelFile.T2S_FIRST_STAGE_DECODER_FP16)
+            t2s_stage_decoder = model_map.get(GSVModelFile.T2S_STAGE_DECODER_FP32) or \
+                                model_map.get(GSVModelFile.T2S_STAGE_DECODER_FP16)
+            prompt_encoder_path = os.path.join(self.character_model_paths[character_name], GSVModelFile.PROMPT_ENCODER)
+
+            return GSVModel(
+                LANGUAGE=language,
+                T2S_ENCODER=model_map[GSVModelFile.T2S_ENCODER_FP32],
+                T2S_FIRST_STAGE_DECODER=t2s_first_stage_decoder,
+                T2S_STAGE_DECODER=t2s_stage_decoder,
+                VITS=model_map[GSVModelFile.VITS_FP32],
+                PROMPT_ENCODER=model_map[GSVModelFile.PROMPT_ENCODER],
+                PROMPT_ENCODER_PATH=prompt_encoder_path,
+            )
+        if character_name in self.character_model_paths:
+            model_dir = self.character_model_paths[character_name]
+            if self.load_character(character_name, model_dir, language=language):
+                return self.get(character_name)
+            else:
+                del self.character_model_paths[character_name]
+                return None
+        return None
+
+    def has_character(self, character_name: str) -> bool:
+        character_name = character_name.lower()
+        return character_name in self.character_model_paths
+
+    def load_character(
+            self,
+            character_name: str,
+            model_dir: str,
+            language: str,
+    ) -> bool:
+        """
+        加载角色模型，如果需要，在内存中动态转换 FP16 权重。
+        """
+        character_name = character_name.lower()
+        if character_name in self.character_to_model:
+            _ = self.character_to_model[character_name]
+            return True
+
+        model_dict: Dict[str, Optional[InferenceSession]] = {}
+
+        # 定义 ONNX 文件到 FP16 Bin 文件的映射关系
+        onnx_to_fp16_map = {
+            GSVModelFile.T2S_FIRST_STAGE_DECODER_FP32: GSVModelFile.T2S_DECODER_WEIGHT_FP16,
+            GSVModelFile.T2S_STAGE_DECODER_FP32: GSVModelFile.T2S_DECODER_WEIGHT_FP16,
+            GSVModelFile.VITS_FP32: GSVModelFile.VITS_WEIGHT_FP16,
+            GSVModelFile.PROMPT_ENCODER: GSVModelFile.PROMPT_ENCODER_WEIGHT_FP16
+        }
+
+        # 确定需要加载的模型列表
+        model_files_to_load = [
+            GSVModelFile.T2S_ENCODER_FP32,
+            GSVModelFile.VITS_FP32,
+            GSVModelFile.PROMPT_ENCODER,
+        ]
+
+        fp32_decoders = [GSVModelFile.T2S_FIRST_STAGE_DECODER_FP32, GSVModelFile.T2S_STAGE_DECODER_FP32]
+        model_files_to_load.extend(fp32_decoders)
+
+        try:
+            for model_file in model_files_to_load:
+                model_path = os.path.normpath(os.path.join(model_dir, model_file))
+
+                # 设置 Session Options
+                sess_options = onnxruntime.SessionOptions()
+                sess_options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
+
+                if os.path.exists(model_path):
+                    fp16_bin_name = onnx_to_fp16_map.get(model_file)
+                    fp16_bin_path = os.path.join(model_dir, fp16_bin_name) if fp16_bin_name else None
+
+                    if fp16_bin_path and os.path.exists(fp16_bin_path):
+                        model_dict[model_file] = load_session_with_fp16_conversion(
+                            model_path, fp16_bin_path, self.providers, sess_options
+                        )
+                    else:
+                        model_dict[model_file] = onnxruntime.InferenceSession(
+                            model_path,
+                            providers=self.providers,
+                            sess_options=sess_options,
+                        )
+                elif model_file == GSVModelFile.PROMPT_ENCODER:
+                    model_dict[model_file] = None
+                else:
+                    raise FileNotFoundError(f'文件 {model_path} 不存在！')
+
+            # 日志信息
+            is_v2pp = model_dict[GSVModelFile.PROMPT_ENCODER] is not None
+            logger.info(
+                f"Character {character_name.capitalize()} loaded successfully.\n"
+                f"- Model Path: {model_dir}\n"
+                f"- Model Type: {'V2ProPlus' if is_v2pp else 'V2'}"
+            )
+
+            self.character_to_model[character_name] = model_dict
+            self.character_to_language[character_name] = language
+            self.character_model_paths[character_name] = model_dir
+            return True
+
+        except Exception as e:
+            logger.error(
+                f"Error: Failed to load ONNX model '{model_dir}'.\n"
+                f"Details: {e}"
+            )
+            return False
+
+    def remove_all_character(self) -> None:
+        self.character_to_model.clear()
+        gc.collect()
+
+    def remove_character(self, character_name: str) -> None:
+        character_name = character_name.lower()
+        if character_name in self.character_to_model:
+            del self.character_to_model[character_name]
+            gc.collect()
+            logger.info(f"Character {character_name.capitalize()} removed successfully.")
+
+
+model_manager: ModelManager = ModelManager()
diff --git a/genie_tts/PredefinedCharacter.py b/genie_tts/PredefinedCharacter.py
index c8ba07dd6b4a5eb4107e8a5a61f1b6f1d4ec82c8..2f084cd7ad8ce02c305fc5ce976038a9f9d7e0a4 100644
--- a/genie_tts/PredefinedCharacter.py
+++ b/genie_tts/PredefinedCharacter.py
@@ -1,39 +1,39 @@
-from huggingface_hub import snapshot_download
-import os
-from typing import Dict
-
-CHARA_LANG: Dict[str, str] = {
-    'mika': 'Japanese',
-    'feibi': 'Chinese',
-    'thirtyseven': 'English',
-}
-CHARA_ALIAS_MAP: Dict[str, str] = {
-    "mika": "mika",
-    "misono mika": "mika",
-    "圣园未花": "mika",
-    "未花": "mika",
-    "みその みか": "mika",
-    "feibi": "feibi",
-    "菲比": "feibi",
-    "37": "thirtyseven",
-    "thirtyseven": "thirtyseven",
-}
-
-
-def download_chara(chara: str, version: str = "v2ProPlus") -> str:
-    local_dir = os.path.join("CharacterModels", version, chara)
-    if os.path.exists(local_dir):
-        print(f"✔ Model for '{chara}' already exists locally. Skipping download.")
-        return local_dir
-
-    print(f"🚀 Starting download of model for character '{chara}'. This may take a few moments... ⏳")
-    remote_path = f"CharacterModels/{version}/{chara}/*"
-    snapshot_download(
-        repo_id="High-Logic/Genie",
-        repo_type="model",
-        allow_patterns=remote_path,
-        local_dir=".",
-        local_dir_use_symlinks=True,  # 软链接
-    )
-    print(f"🎉 All model files for '{chara}' have been downloaded to '{os.path.abspath(local_dir)}' 📂")
-    return local_dir
+from huggingface_hub import snapshot_download
+import os
+from typing import Dict
+
+CHARA_LANG: Dict[str, str] = {
+    'mika': 'Japanese',
+    'feibi': 'Chinese',
+    'thirtyseven': 'English',
+}
+CHARA_ALIAS_MAP: Dict[str, str] = {
+    "mika": "mika",
+    "misono mika": "mika",
+    "圣园未花": "mika",
+    "未花": "mika",
+    "みその みか": "mika",
+    "feibi": "feibi",
+    "菲比": "feibi",
+    "37": "thirtyseven",
+    "thirtyseven": "thirtyseven",
+}
+
+
+def download_chara(chara: str, version: str = "v2ProPlus") -> str:
+    local_dir = os.path.join("CharacterModels", version, chara)
+    if os.path.exists(local_dir):
+        print(f"✔ Model for '{chara}' already exists locally. Skipping download.")
+        return local_dir
+
+    print(f"🚀 Starting download of model for character '{chara}'. This may take a few moments... ⏳")
+    remote_path = f"CharacterModels/{version}/{chara}/*"
+    snapshot_download(
+        repo_id="High-Logic/Genie",
+        repo_type="model",
+        allow_patterns=remote_path,
+        local_dir=".",
+        local_dir_use_symlinks=True,  # 软链接
+    )
+    print(f"🎉 All model files for '{chara}' have been downloaded to '{os.path.abspath(local_dir)}' 📂")
+    return local_dir
diff --git a/genie_tts/Server.py b/genie_tts/Server.py
index 458bccd4c2cce79f46ed85f99bebbb969821a587..499e9d86d21021fb20aabab7c8f84951bd01726d 100644
--- a/genie_tts/Server.py
+++ b/genie_tts/Server.py
@@ -1,169 +1,169 @@
-import asyncio
-import os
-from typing import AsyncIterator, Optional, Callable, Union, Dict
-import logging
-
-import uvicorn
-from fastapi import FastAPI, HTTPException
-from fastapi.responses import StreamingResponse
-from pydantic import BaseModel
-
-from .Audio.ReferenceAudio import ReferenceAudio
-from .Core.TTSPlayer import tts_player
-from .ModelManager import model_manager
-from .Utils.Shared import context
-from .Utils.Language import normalize_language
-
-logger = logging.getLogger(__name__)
-
-_reference_audios: Dict[str, dict] = {}
-SUPPORTED_AUDIO_EXTS = {'.wav', '.flac', '.ogg', '.aiff', '.aif'}
-
-app = FastAPI()
-
-
-class CharacterPayload(BaseModel):
-    character_name: str
-    onnx_model_dir: str
-    language: str
-
-
-class UnloadCharacterPayload(BaseModel):
-    character_name: str
-
-
-class ReferenceAudioPayload(BaseModel):
-    character_name: str
-    audio_path: str
-    audio_text: str
-    language: str
-
-
-class TTSPayload(BaseModel):
-    character_name: str
-    text: str
-    split_sentence: bool = False
-    save_path: Optional[str] = None
-
-
-@app.post("/load_character")
-def load_character_endpoint(payload: CharacterPayload):
-    try:
-        model_manager.load_character(
-            character_name=payload.character_name,
-            model_dir=payload.onnx_model_dir,
-            language=normalize_language(payload.language),
-        )
-        return {"status": "success", "message": f"Character '{payload.character_name}' loaded."}
-    except Exception as e:
-        raise HTTPException(status_code=500, detail=str(e))
-
-
-@app.post("/unload_character")
-def unload_character_endpoint(payload: UnloadCharacterPayload):
-    try:
-        model_manager.remove_character(character_name=payload.character_name)
-        return {"status": "success", "message": f"Character '{payload.character_name}' unloaded."}
-    except Exception as e:
-        raise HTTPException(status_code=500, detail=str(e))
-
-
-@app.post("/set_reference_audio")
-def set_reference_audio_endpoint(payload: ReferenceAudioPayload):
-    ext = os.path.splitext(payload.audio_path)[1].lower()
-    if ext not in SUPPORTED_AUDIO_EXTS:
-        raise HTTPException(
-            status_code=400,
-            detail=f"Audio format '{ext}' is not supported. Supported formats: {SUPPORTED_AUDIO_EXTS}",
-        )
-    _reference_audios[payload.character_name] = {
-        'audio_path': payload.audio_path,
-        'audio_text': payload.audio_text,
-        'language': normalize_language(payload.language),
-    }
-    return {"status": "success", "message": f"Reference audio for '{payload.character_name}' set."}
-
-
-def run_tts_in_background(
-        character_name: str,
-        text: str,
-        split_sentence: bool,
-        save_path: Optional[str],
-        chunk_callback: Callable[[Optional[bytes]], None]
-):
-    try:
-        context.current_speaker = character_name
-        context.current_prompt_audio = ReferenceAudio(
-            prompt_wav=_reference_audios[character_name]['audio_path'],
-            prompt_text=_reference_audios[character_name]['audio_text'],
-            language=_reference_audios[character_name]['language'],
-        )
-        tts_player.start_session(
-            play=False,
-            split=split_sentence,
-            save_path=save_path,
-            chunk_callback=chunk_callback,
-        )
-        tts_player.feed(text)
-        tts_player.end_session()
-        tts_player.wait_for_tts_completion()
-    except Exception as e:
-        logger.error(f"Error in TTS background task: {e}", exc_info=True)
-
-
-async def audio_stream_generator(queue: asyncio.Queue) -> AsyncIterator[bytes]:
-    while True:
-        chunk = await queue.get()
-        if chunk is None:
-            break
-        yield chunk
-
-
-@app.post("/tts")
-async def tts_endpoint(payload: TTSPayload):
-    if payload.character_name not in _reference_audios:
-        raise HTTPException(status_code=404, detail="Character not found or reference audio not set.")
-
-    loop = asyncio.get_running_loop()
-    stream_queue: asyncio.Queue[Union[bytes, None]] = asyncio.Queue()
-
-    def tts_chunk_callback(chunk: Optional[bytes]):
-        loop.call_soon_threadsafe(stream_queue.put_nowait, chunk)
-
-    loop.run_in_executor(
-        None,
-        run_tts_in_background,
-        payload.character_name,
-        payload.text,
-        payload.split_sentence,
-        payload.save_path,
-        tts_chunk_callback
-    )
-
-    return StreamingResponse(audio_stream_generator(stream_queue), media_type="audio/wav")
-
-
-@app.post("/stop")
-def stop_endpoint():
-    try:
-        tts_player.stop()
-        return {"status": "success", "message": "TTS stopped."}
-    except Exception as e:
-        raise HTTPException(status_code=500, detail=str(e))
-
-
-@app.post("/clear_reference_audio_cache")
-def clear_reference_audio_cache_endpoint():
-    try:
-        ReferenceAudio.clear_cache()
-        return {"status": "success", "message": "Reference audio cache cleared."}
-    except Exception as e:
-        raise HTTPException(status_code=500, detail=str(e))
-
-
-def start_server(host: str = "127.0.0.1", port: int = 8000, workers: int = 1):
-    uvicorn.run(app, host=host, port=port, workers=workers)
-
-
-if __name__ == "__main__":
-    start_server(host="0.0.0.0", port=8000, workers=1)
+import asyncio
+import os
+from typing import AsyncIterator, Optional, Callable, Union, Dict
+import logging
+
+import uvicorn
+from fastapi import FastAPI, HTTPException
+from fastapi.responses import StreamingResponse
+from pydantic import BaseModel
+
+from .Audio.ReferenceAudio import ReferenceAudio
+from .Core.TTSPlayer import tts_player
+from .ModelManager import model_manager
+from .Utils.Shared import context
+from .Utils.Language import normalize_language
+
+logger = logging.getLogger(__name__)
+
+_reference_audios: Dict[str, dict] = {}
+SUPPORTED_AUDIO_EXTS = {'.wav', '.flac', '.ogg', '.aiff', '.aif'}
+
+app = FastAPI()
+
+
+class CharacterPayload(BaseModel):
+    character_name: str
+    onnx_model_dir: str
+    language: str
+
+
+class UnloadCharacterPayload(BaseModel):
+    character_name: str
+
+
+class ReferenceAudioPayload(BaseModel):
+    character_name: str
+    audio_path: str
+    audio_text: str
+    language: str
+
+
+class TTSPayload(BaseModel):
+    character_name: str
+    text: str
+    split_sentence: bool = False
+    save_path: Optional[str] = None
+
+
+@app.post("/load_character")
+def load_character_endpoint(payload: CharacterPayload):
+    try:
+        model_manager.load_character(
+            character_name=payload.character_name,
+            model_dir=payload.onnx_model_dir,
+            language=normalize_language(payload.language),
+        )
+        return {"status": "success", "message": f"Character '{payload.character_name}' loaded."}
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))
+
+
+@app.post("/unload_character")
+def unload_character_endpoint(payload: UnloadCharacterPayload):
+    try:
+        model_manager.remove_character(character_name=payload.character_name)
+        return {"status": "success", "message": f"Character '{payload.character_name}' unloaded."}
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))
+
+
+@app.post("/set_reference_audio")
+def set_reference_audio_endpoint(payload: ReferenceAudioPayload):
+    ext = os.path.splitext(payload.audio_path)[1].lower()
+    if ext not in SUPPORTED_AUDIO_EXTS:
+        raise HTTPException(
+            status_code=400,
+            detail=f"Audio format '{ext}' is not supported. Supported formats: {SUPPORTED_AUDIO_EXTS}",
+        )
+    _reference_audios[payload.character_name] = {
+        'audio_path': payload.audio_path,
+        'audio_text': payload.audio_text,
+        'language': normalize_language(payload.language),
+    }
+    return {"status": "success", "message": f"Reference audio for '{payload.character_name}' set."}
+
+
+def run_tts_in_background(
+        character_name: str,
+        text: str,
+        split_sentence: bool,
+        save_path: Optional[str],
+        chunk_callback: Callable[[Optional[bytes]], None]
+):
+    try:
+        context.current_speaker = character_name
+        context.current_prompt_audio = ReferenceAudio(
+            prompt_wav=_reference_audios[character_name]['audio_path'],
+            prompt_text=_reference_audios[character_name]['audio_text'],
+            language=_reference_audios[character_name]['language'],
+        )
+        tts_player.start_session(
+            play=False,
+            split=split_sentence,
+            save_path=save_path,
+            chunk_callback=chunk_callback,
+        )
+        tts_player.feed(text)
+        tts_player.end_session()
+        tts_player.wait_for_tts_completion()
+    except Exception as e:
+        logger.error(f"Error in TTS background task: {e}", exc_info=True)
+
+
+async def audio_stream_generator(queue: asyncio.Queue) -> AsyncIterator[bytes]:
+    while True:
+        chunk = await queue.get()
+        if chunk is None:
+            break
+        yield chunk
+
+
+@app.post("/tts")
+async def tts_endpoint(payload: TTSPayload):
+    if payload.character_name not in _reference_audios:
+        raise HTTPException(status_code=404, detail="Character not found or reference audio not set.")
+
+    loop = asyncio.get_running_loop()
+    stream_queue: asyncio.Queue[Union[bytes, None]] = asyncio.Queue()
+
+    def tts_chunk_callback(chunk: Optional[bytes]):
+        loop.call_soon_threadsafe(stream_queue.put_nowait, chunk)
+
+    loop.run_in_executor(
+        None,
+        run_tts_in_background,
+        payload.character_name,
+        payload.text,
+        payload.split_sentence,
+        payload.save_path,
+        tts_chunk_callback
+    )
+
+    return StreamingResponse(audio_stream_generator(stream_queue), media_type="audio/wav")
+
+
+@app.post("/stop")
+def stop_endpoint():
+    try:
+        tts_player.stop()
+        return {"status": "success", "message": "TTS stopped."}
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))
+
+
+@app.post("/clear_reference_audio_cache")
+def clear_reference_audio_cache_endpoint():
+    try:
+        ReferenceAudio.clear_cache()
+        return {"status": "success", "message": "Reference audio cache cleared."}
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))
+
+
+def start_server(host: str = "127.0.0.1", port: int = 8000, workers: int = 1):
+    uvicorn.run(app, host=host, port=port, workers=workers)
+
+
+if __name__ == "__main__":
+    start_server(host="0.0.0.0", port=8000, workers=1)
diff --git a/genie_tts/Utils/Constants.py b/genie_tts/Utils/Constants.py
index 6b134e10b7612f06b6cc548938b27cf745575d1a..9dbafa80454155805ecee50eb104431a7450626c 100644
--- a/genie_tts/Utils/Constants.py
+++ b/genie_tts/Utils/Constants.py
@@ -1,2 +1,2 @@
-BERT_FEATURE_DIM = 1024
-PACKAGE_NAME = "genie_tts"
+BERT_FEATURE_DIM = 1024
+PACKAGE_NAME = "genie_tts"
diff --git a/genie_tts/Utils/Language.py b/genie_tts/Utils/Language.py
index 7d6f23aa8bdedcddd125c7400d51381ee3b5f1cc..345671a478ba6a5f3ff8a18305e95fb4941d9107 100644
--- a/genie_tts/Utils/Language.py
+++ b/genie_tts/Utils/Language.py
@@ -1,26 +1,26 @@
-language_map = {
-    # Chinese
-    "chinese": "Chinese",
-    "zh": "Chinese",
-    "zh-cn": "Chinese",
-    "zh-tw": "Chinese",
-    "zh-hans": "Chinese",
-    "zh-hant": "Chinese",
-
-    # English
-    "english": "English",
-    "en": "English",
-    "en-us": "English",
-    "en-gb": "English",
-    "eng": "English",
-
-    # Japanese
-    "japanese": "Japanese",
-    "jp": "Japanese",
-    "ja": "Japanese",
-    "nihongo": "Japanese",
-}
-
-
-def normalize_language(lang: str) -> str:
-    return language_map.get(lang.lower(), lang)
+language_map = {
+    # Chinese
+    "chinese": "Chinese",
+    "zh": "Chinese",
+    "zh-cn": "Chinese",
+    "zh-tw": "Chinese",
+    "zh-hans": "Chinese",
+    "zh-hant": "Chinese",
+
+    # English
+    "english": "English",
+    "en": "English",
+    "en-us": "English",
+    "en-gb": "English",
+    "eng": "English",
+
+    # Japanese
+    "japanese": "Japanese",
+    "jp": "Japanese",
+    "ja": "Japanese",
+    "nihongo": "Japanese",
+}
+
+
+def normalize_language(lang: str) -> str:
+    return language_map.get(lang.lower(), lang)
diff --git a/genie_tts/Utils/TextSplitter.py b/genie_tts/Utils/TextSplitter.py
index b9dca1045dbad88a5ed8a9315ac4d0c4307f76b8..775aa15bd3ac1bf7b6a6e83d1842ec556292ef4f 100644
--- a/genie_tts/Utils/TextSplitter.py
+++ b/genie_tts/Utils/TextSplitter.py
@@ -1,123 +1,123 @@
-import re
-from typing import List, Set, Pattern
-
-
-class TextSplitter:
-    def __init__(self, max_len: int = 40, min_len: int = 5):
-        """
-        初始化文本切分器。
-
-        :param max_len: 软限制最大长度 (Effective Length)。超过此长度遇到分隔符时会切分。
-        :param min_len: 硬限制最小长度 (Effective Length)。小于此长度遇到终止符也不会切分。
-        """
-        self.max_len: int = max_len
-        self.min_len: int = min_len
-
-        # 1. 定义基础字符集合
-        # 只要标点块中包含这些字符，就视为 Ending (终止符)
-        self.end_chars: Set[str] = {
-            '。', '！', '？', '…',
-            '!', '?', '.'
-        }
-
-        # 2. 定义标点符号全集 (用于正则匹配和长度计算过滤)
-        self.all_puncts_chars: Set[str] = self.end_chars | {
-            '，', '、', '；', '：', '——',
-            ',', ';', ':',
-            '“', '”', '‘', '’', '"', "'",
-        }
-
-        # 3. 编译正则表达式
-        # 使用非捕获组 (?:) 配合 + 号，实现贪婪匹配连续标点
-        # sort + escape 确保正则安全且优先匹配长标点
-        sorted_puncts: List[str] = sorted(list(self.all_puncts_chars), key=len, reverse=True)
-        escaped_puncts: List[str] = [re.escape(p) for p in sorted_puncts]
-        self.pattern: Pattern = re.compile(f"((?:{'|'.join(escaped_puncts)})+)")
-
-    @staticmethod
-    def get_char_width(char: str) -> int:
-        """计算单字符宽度：ASCII算1，其他（中日韩）算2"""
-        return 1 if ord(char) < 128 else 2
-
-    def get_effective_len(self, text: str) -> int:
-        """
-        计算字符串的有效长度。
-        逻辑：跳过标点符号，仅计算内容字符。
-        例如："你好......" -> 有效长度为 4 (你好)，而不是 10。
-        """
-        length = 0
-        for char in text:
-            # 如果是标点符号集合里的字符，不计入长度
-            if char in self.all_puncts_chars:
-                continue
-            length += self.get_char_width(char)
-        return length
-
-    def is_terminator_block(self, block: str) -> bool:
-        """
-        判断一个标点块是否起到结束句子的作用。
-        只要块中包含任意一个结束字符（如句号），则视为结束块。
-        """
-        for char in block:
-            if char in self.end_chars:
-                return True
-        return False
-
-    def split(self, text: str) -> List[str]:
-        """核心切分逻辑"""
-        if not text:
-            return []
-
-        text = text.replace('\n', '')
-
-        # 正则切分，segments 格式如: ['你好', '......', '我是谁', '？！', '']
-        segments: List[str] = self.pattern.split(text)
-
-        sentences: List[str] = []
-        current_buffer: str = ""
-
-        for segment in segments:
-            if not segment:
-                continue
-
-            # 判断当前片段是否是标点块（通过首字符判断即可，正则保证了一致性）
-            is_punct_block = segment[0] in self.all_puncts_chars
-
-            if is_punct_block:
-                current_buffer += segment
-
-                # 计算缓冲区内容的【有效长度】
-                eff_len = self.get_effective_len(current_buffer)
-
-                # 判断逻辑
-                if self.is_terminator_block(segment):
-                    # Case B: 结束符号 -> 检查 min_len
-                    if eff_len >= self.min_len:
-                        sentences.append(current_buffer.strip())
-                        current_buffer = ""
-                    # else: 有效长度太短，合并到下一句
-                else:
-                    # Case A-B: 分隔符号 -> 检查 max_len
-                    if eff_len >= self.max_len:
-                        sentences.append(current_buffer.strip())
-                        current_buffer = ""
-                    # else: 没到最大长度，继续累积
-            else:
-                # 纯文本
-                current_buffer += segment
-
-        # 处理残留缓冲区
-        if current_buffer:
-            self._flush_buffer(sentences, current_buffer)
-
-        return sentences
-
-    def _flush_buffer(self, sentences: List[str], buffer: str):
-        candidate = buffer.strip()
-        if not candidate:
-            return
-        eff_len = self.get_effective_len(candidate)
-        if eff_len > 0:
-            sentences.append(candidate)
-        elif sentences:
-            sentences[-1] += candidate
+import re
+from typing import List, Set, Pattern
+
+
+class TextSplitter:
+    def __init__(self, max_len: int = 40, min_len: int = 5):
+        """
+        初始化文本切分器。
+
+        :param max_len: 软限制最大长度 (Effective Length)。超过此长度遇到分隔符时会切分。
+        :param min_len: 硬限制最小长度 (Effective Length)。小于此长度遇到终止符也不会切分。
+        """
+        self.max_len: int = max_len
+        self.min_len: int = min_len
+
+        # 1. 定义基础字符集合
+        # 只要标点块中包含这些字符，就视为 Ending (终止符)
+        self.end_chars: Set[str] = {
+            '。', '！', '？', '…',
+            '!', '?', '.'
+        }
+
+        # 2. 定义标点符号全集 (用于正则匹配和长度计算过滤)
+        self.all_puncts_chars: Set[str] = self.end_chars | {
+            '，', '、', '；', '：', '——',
+            ',', ';', ':',
+            '“', '”', '‘', '’', '"', "'",
+        }
+
+        # 3. 编译正则表达式
+        # 使用非捕获组 (?:) 配合 + 号，实现贪婪匹配连续标点
+        # sort + escape 确保正则安全且优先匹配长标点
+        sorted_puncts: List[str] = sorted(list(self.all_puncts_chars), key=len, reverse=True)
+        escaped_puncts: List[str] = [re.escape(p) for p in sorted_puncts]
+        self.pattern: Pattern = re.compile(f"((?:{'|'.join(escaped_puncts)})+)")
+
+    @staticmethod
+    def get_char_width(char: str) -> int:
+        """计算单字符宽度：ASCII算1，其他（中日韩）算2"""
+        return 1 if ord(char) < 128 else 2
+
+    def get_effective_len(self, text: str) -> int:
+        """
+        计算字符串的有效长度。
+        逻辑：跳过标点符号，仅计算内容字符。
+        例如："你好......" -> 有效长度为 4 (你好)，而不是 10。
+        """
+        length = 0
+        for char in text:
+            # 如果是标点符号集合里的字符，不计入长度
+            if char in self.all_puncts_chars:
+                continue
+            length += self.get_char_width(char)
+        return length
+
+    def is_terminator_block(self, block: str) -> bool:
+        """
+        判断一个标点块是否起到结束句子的作用。
+        只要块中包含任意一个结束字符（如句号），则视为结束块。
+        """
+        for char in block:
+            if char in self.end_chars:
+                return True
+        return False
+
+    def split(self, text: str) -> List[str]:
+        """核心切分逻辑"""
+        if not text:
+            return []
+
+        text = text.replace('\n', '')
+
+        # 正则切分，segments 格式如: ['你好', '......', '我是谁', '？！', '']
+        segments: List[str] = self.pattern.split(text)
+
+        sentences: List[str] = []
+        current_buffer: str = ""
+
+        for segment in segments:
+            if not segment:
+                continue
+
+            # 判断当前片段是否是标点块（通过首字符判断即可，正则保证了一致性）
+            is_punct_block = segment[0] in self.all_puncts_chars
+
+            if is_punct_block:
+                current_buffer += segment
+
+                # 计算缓冲区内容的【有效长度】
+                eff_len = self.get_effective_len(current_buffer)
+
+                # 判断逻辑
+                if self.is_terminator_block(segment):
+                    # Case B: 结束符号 -> 检查 min_len
+                    if eff_len >= self.min_len:
+                        sentences.append(current_buffer.strip())
+                        current_buffer = ""
+                    # else: 有效长度太短，合并到下一句
+                else:
+                    # Case A-B: 分隔符号 -> 检查 max_len
+                    if eff_len >= self.max_len:
+                        sentences.append(current_buffer.strip())
+                        current_buffer = ""
+                    # else: 没到最大长度，继续累积
+            else:
+                # 纯文本
+                current_buffer += segment
+
+        # 处理残留缓冲区
+        if current_buffer:
+            self._flush_buffer(sentences, current_buffer)
+
+        return sentences
+
+    def _flush_buffer(self, sentences: List[str], buffer: str):
+        candidate = buffer.strip()
+        if not candidate:
+            return
+        eff_len = self.get_effective_len(candidate)
+        if eff_len > 0:
+            sentences.append(candidate)
+        elif sentences:
+            sentences[-1] += candidate
diff --git a/genie_tts/Utils/UserData.py b/genie_tts/Utils/UserData.py
index 1d1bd517e02d35f683e92c3536fbc6027ab2b438..a32820c04d8753d609130f29a4693731f530dbee 100644
--- a/genie_tts/Utils/UserData.py
+++ b/genie_tts/Utils/UserData.py
@@ -1,44 +1,44 @@
-import json
-from pathlib import Path
-from typing import Dict, Any
-import logging
-import importlib.resources
-
-from .Constants import PACKAGE_NAME
-
-logger = logging.getLogger(__name__)
-
-
-class UserDataManager:
-    def __init__(self, file_path: str = "./UserData.json"):
-        self.file_path = Path(file_path)
-        self._data: Dict[str, Any] = self._load()
-
-    def _load(self) -> Dict[str, Any]:
-        if self.file_path.exists():
-            try:
-                with self.file_path.open('r', encoding='utf-8') as f:
-                    return json.load(f)
-            except (json.JSONDecodeError, IOError) as e:
-                logger.warning(f"Failed to read user data file {self.file_path}. Using empty configuration. Error: {e}")
-
-                return {}
-        return {}
-
-    def _save(self):
-        try:
-            with self.file_path.open('w', encoding='utf-8') as f:
-                json.dump(self._data, f, indent=4, ensure_ascii=False)  # type: ignore
-        except IOError as e:
-            logger.warning(f"Failed to write user data file {self.file_path}. Error: {e}")
-
-    def get(self, key: str, default: Any = None) -> Any:
-        return self._data.get(key, default)
-
-    def set(self, key: str, value: Any):
-        self._data[key] = value
-        self._save()
-
-
-userdata_file: str = str(importlib.resources.files(PACKAGE_NAME) / 'UserData.json')
-userdata_manager = UserDataManager(file_path=userdata_file)
+import json
+from pathlib import Path
+from typing import Dict, Any
+import logging
+import importlib.resources
+
+from .Constants import PACKAGE_NAME
+
+logger = logging.getLogger(__name__)
+
+
+class UserDataManager:
+    def __init__(self, file_path: str = "./UserData.json"):
+        self.file_path = Path(file_path)
+        self._data: Dict[str, Any] = self._load()
+
+    def _load(self) -> Dict[str, Any]:
+        if self.file_path.exists():
+            try:
+                with self.file_path.open('r', encoding='utf-8') as f:
+                    return json.load(f)
+            except (json.JSONDecodeError, IOError) as e:
+                logger.warning(f"Failed to read user data file {self.file_path}. Using empty configuration. Error: {e}")
+
+                return {}
+        return {}
+
+    def _save(self):
+        try:
+            with self.file_path.open('w', encoding='utf-8') as f:
+                json.dump(self._data, f, indent=4, ensure_ascii=False)  # type: ignore
+        except IOError as e:
+            logger.warning(f"Failed to write user data file {self.file_path}. Error: {e}")
+
+    def get(self, key: str, default: Any = None) -> Any:
+        return self._data.get(key, default)
+
+    def set(self, key: str, value: Any):
+        self._data[key] = value
+        self._save()
+
+
+userdata_file: str = str(importlib.resources.files(PACKAGE_NAME) / 'UserData.json')
+userdata_manager = UserDataManager(file_path=userdata_file)
diff --git a/genie_tts/Utils/Utils.py b/genie_tts/Utils/Utils.py
index 1453655a72e22716f4e966ef418480497f0a314c..7a5dc3aee0a8438d7c60ea1c9170f24708dd401b 100644
--- a/genie_tts/Utils/Utils.py
+++ b/genie_tts/Utils/Utils.py
@@ -1,28 +1,28 @@
-from collections import OrderedDict
-import queue
-
-
-class LRUCacheDict(OrderedDict):
-    def __init__(self, capacity):
-        super().__init__()
-        self.capacity = capacity
-
-    def __getitem__(self, key):
-        value = super().__getitem__(key)
-        self.move_to_end(key)  # 访问后移到末尾
-        return value
-
-    def __setitem__(self, key, value):
-        if key in self:
-            self.move_to_end(key)
-        super().__setitem__(key, value)
-        if len(self) > self.capacity:
-            self.popitem(last=False)  # 删除最旧的（第一个）
-
-
-def clear_queue(q: queue.Queue) -> None:
-    while not q.empty():
-        try:
-            q.get_nowait()
-        except queue.Empty:
-            break
+from collections import OrderedDict
+import queue
+
+
+class LRUCacheDict(OrderedDict):
+    def __init__(self, capacity):
+        super().__init__()
+        self.capacity = capacity
+
+    def __getitem__(self, key):
+        value = super().__getitem__(key)
+        self.move_to_end(key)  # 访问后移到末尾
+        return value
+
+    def __setitem__(self, key, value):
+        if key in self:
+            self.move_to_end(key)
+        super().__setitem__(key, value)
+        if len(self) > self.capacity:
+            self.popitem(last=False)  # 删除最旧的（第一个）
+
+
+def clear_queue(q: queue.Queue) -> None:
+    while not q.empty():
+        try:
+            q.get_nowait()
+        except queue.Empty:
+            break
diff --git a/genie_tts/Utils/__pycache__/Constants.cpython-311.pyc b/genie_tts/Utils/__pycache__/Constants.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..545d5045f6341fee86d4c69b4672e5fd65bfce5d
Binary files /dev/null and b/genie_tts/Utils/__pycache__/Constants.cpython-311.pyc differ
diff --git a/genie_tts/Utils/__pycache__/Language.cpython-311.pyc b/genie_tts/Utils/__pycache__/Language.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5f5889d376989ff760e7ba39c580c32d5126c377
Binary files /dev/null and b/genie_tts/Utils/__pycache__/Language.cpython-311.pyc differ
diff --git a/genie_tts/Utils/__pycache__/Shared.cpython-311.pyc b/genie_tts/Utils/__pycache__/Shared.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..e806748737923a29612ec2dececb7cdf64c33ff2
Binary files /dev/null and b/genie_tts/Utils/__pycache__/Shared.cpython-311.pyc differ
diff --git a/genie_tts/Utils/__pycache__/TextSplitter.cpython-311.pyc b/genie_tts/Utils/__pycache__/TextSplitter.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..80885a6eb7908aa4a1fb49c83c4c44af056c3de8
Binary files /dev/null and b/genie_tts/Utils/__pycache__/TextSplitter.cpython-311.pyc differ
diff --git a/genie_tts/Utils/__pycache__/Utils.cpython-311.pyc b/genie_tts/Utils/__pycache__/Utils.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5d8a9a67807edd0849ee5468d9e9b328c4a9b08f
Binary files /dev/null and b/genie_tts/Utils/__pycache__/Utils.cpython-311.pyc differ
diff --git a/genie_tts/Utils/__pycache__/__init__.cpython-311.pyc b/genie_tts/Utils/__pycache__/__init__.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..02a8876a02344b41c35e92b723c01343a3e99035
Binary files /dev/null and b/genie_tts/Utils/__pycache__/__init__.cpython-311.pyc differ
diff --git a/genie_tts/__init__.py b/genie_tts/__init__.py
index d910acdb31af65801d20206bbebfb55544af3570..b0d4fdcbcd124bdac127cc4037b95de73b4092d3 100644
--- a/genie_tts/__init__.py
+++ b/genie_tts/__init__.py
@@ -1,29 +1,31 @@
-from .Internal import (
-    load_character,
-    unload_character,
-    set_reference_audio,
-    tts_async,
-    tts,
-    stop,
-    convert_to_onnx,
-    clear_reference_audio_cache,
-    load_predefined_character,
-    wait_for_playback_done,
-)
-from .Server import start_server
-from .Core.Resources import download_genie_data
-
-__all__ = [
-    "load_character",
-    "unload_character",
-    "set_reference_audio",
-    "tts_async",
-    "tts",
-    "stop",
-    "convert_to_onnx",
-    "clear_reference_audio_cache",
-    "start_server",
-    "load_predefined_character",
-    "wait_for_playback_done",
-    'download_genie_data',
-]
+from .Internal import (
+    load_character,
+    unload_character,
+    set_reference_audio,
+    tts_async,
+    tts,
+    stop,
+    convert_to_onnx,
+    clear_reference_audio_cache,
+    load_predefined_character,
+    wait_for_playback_done,
+)
+from .Server import start_server
+from .Core.Resources import download_genie_data
+from .ModelManager import model_manager
+
+__all__ = [
+    "load_character",
+    "unload_character",
+    "set_reference_audio",
+    "tts_async",
+    "tts",
+    "stop",
+    "convert_to_onnx",
+    "clear_reference_audio_cache",
+    "start_server",
+    "load_predefined_character",
+    "wait_for_playback_done",
+    'download_genie_data',
+    "model_manager",
+]
diff --git a/genie_tts/__pycache__/GetPhonesAndBert.cpython-311.pyc b/genie_tts/__pycache__/GetPhonesAndBert.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5854509ff9027dc8542f66f475b0817d88574255
Binary files /dev/null and b/genie_tts/__pycache__/GetPhonesAndBert.cpython-311.pyc differ
diff --git a/genie_tts/__pycache__/Internal.cpython-311.pyc b/genie_tts/__pycache__/Internal.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..7d12a3b2d88c15beacc3ca945587fb579e7a0ffd
Binary files /dev/null and b/genie_tts/__pycache__/Internal.cpython-311.pyc differ
diff --git a/genie_tts/__pycache__/ModelManager.cpython-311.pyc b/genie_tts/__pycache__/ModelManager.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5e0c176223b24fde61940caadc612b98765f381f
Binary files /dev/null and b/genie_tts/__pycache__/ModelManager.cpython-311.pyc differ
diff --git a/genie_tts/__pycache__/PredefinedCharacter.cpython-311.pyc b/genie_tts/__pycache__/PredefinedCharacter.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..387529a1944a268d693d20e0b632964dbc8f8d17
Binary files /dev/null and b/genie_tts/__pycache__/PredefinedCharacter.cpython-311.pyc differ
diff --git a/genie_tts/__pycache__/Server.cpython-311.pyc b/genie_tts/__pycache__/Server.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6847332b0bce32e9469d20f991a34d0c2fb8ee06
Binary files /dev/null and b/genie_tts/__pycache__/Server.cpython-311.pyc differ
diff --git a/genie_tts/__pycache__/__init__.cpython-311.pyc b/genie_tts/__pycache__/__init__.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..63d8a1e31e8d271e5f3e0865a4abe8b2ebf2e877
Binary files /dev/null and b/genie_tts/__pycache__/__init__.cpython-311.pyc differ
diff --git a/models/base/ref.wav b/models/base/ref.wav
index 60b4426d3fd8427b875fb482f42abfd9739b4510..9a7847612f1411ae1c804b93e3b99812758d6645 100644
Binary files a/models/base/ref.wav and b/models/base/ref.wav differ
diff --git a/models/god/ref.wav b/models/god/ref.wav
index 9c9b6bf056f646d4f329537f718c62e3c111fa1a..b718b57dc5f9a25e7cd6be1479420468a4bce59c 100644
Binary files a/models/god/ref.wav and b/models/god/ref.wav differ