Bundle Qwen3-ASR weights in packaged app; slim PyInstaller assets

- qwen.py resolves model from env var > bundled assets dir > HF download
- PyInstaller hook: include assets/models/asr/qwen3-asr-1.7b, exclude
legacy funasr/whisper models and superseded TTS .bin weights
(weights are not committed; copy to assets/models/asr/qwen3-asr-1.7b
before packaging)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

build/pyinstaller/hooks/hook-voice_dialogue.py

@@ -24,8 +24,29 @@ ASSETS_ROOT = PROJECT_ROOT / "assets"
 # 收集主模块的所有子模块
 hiddenimports = collect_submodules('voice_dialogue')
 datas = collect_data_files('moyoyo_tts', include_py_files=True)
+
+# 不打包的资源：
+# - 旧版 FunASR/Whisper 模型（默认引擎为内置的 Qwen3-ASR）
+# - TTS 预训练权重的 .bin（已内置等价的 model.safetensors）
+EXCLUDED_ASSET_PATTERNS = [
+    "assets/models/asr/funasr/",
+    "assets/models/asr/whisper/",
+    "chinese-roberta-wwm-ext-large/pytorch_model.bin",
+    "chinese-hubert-base/pytorch_model.bin",
+]
+
+
+def _is_excluded(source_path: str) -> bool:
+    normalized = source_path.replace("\\", "/")
+    return any(pattern in normalized for pattern in EXCLUDED_ASSET_PATTERNS)
+
+
 # 收集系统资源文件
-datas += collect_system_data_files(ASSETS_ROOT.as_posix(), "assets")
+datas += [
+    (source, dest)
+    for source, dest in collect_system_data_files(ASSETS_ROOT.as_posix(), "assets")
+    if not _is_excluded(source)
+]
 
 # ============================================================================
 # 第三方依赖配置
@@ -39,6 +60,7 @@ ML_DEPENDENCIES = [
     "pytorch_lightning",
     "huggingface_hub",
     "einops",
+    "qwen_asr",
 ]
 
 # 语音处理相关依赖
@@ -117,6 +139,7 @@ DATA_PACKAGES = [
     ("spacy", {"include_py_files": True}),
     ("misaki", {}),
     ("silero_vad", {}),
+    ("qwen_asr", {}),
 ]
 
 # 收集数据文件

src/voice_dialogue/asr/models/qwen.py

@@ -8,13 +8,25 @@ from qwen_asr import Qwen3ASRModel
 from voice_dialogue.asr.manager import asr_tables
 from voice_dialogue.asr.models.base import ASRInterface
 from voice_dialogue.asr.utils import ensure_minimum_audio_duration
+from voice_dialogue.config import paths
 from voice_dialogue.utils.logger import logger
 
-DEFAULT_MODEL = os.environ.get('QWEN_ASR_MODEL', 'Qwen/Qwen3-ASR-1.7B')
+# 内置模型目录（打包分发时随应用携带，存在则离线加载）
+BUILTIN_QWEN_ASR_MODEL_PATH = paths.ASR_MODELS_PATH / 'qwen3-asr-1.7b'
 
 TARGET_SAMPLE_RATE = 16000
 
 
+def resolve_model_path() -> str:
+    """模型来源优先级：环境变量 > 内置目录 > HuggingFace 自动下载。"""
+    env_model = os.environ.get('QWEN_ASR_MODEL')
+    if env_model:
+        return env_model
+    if (BUILTIN_QWEN_ASR_MODEL_PATH / 'config.json').exists():
+        return BUILTIN_QWEN_ASR_MODEL_PATH.as_posix()
+    return 'Qwen/Qwen3-ASR-1.7B'
+
+
 @asr_tables.register('asr_classes', 'qwen')
 class QwenASRClient(ASRInterface):
     """Qwen3-ASR 客户端（transformers 后端，macOS 上使用 MPS 加速）"""
@@ -25,7 +37,7 @@ class QwenASRClient(ASRInterface):
         self.model: typing.Optional[Qwen3ASRModel] = None
 
     def setup(self, **kwargs) -> None:
-        model_name = kwargs.get('model', DEFAULT_MODEL)
+        model_name = kwargs.get('model') or resolve_model_path()
 
         if torch.backends.mps.is_available():
             device_map, dtype = 'mps', torch.bfloat16