feat: Docker 构建预置 ASR 模型，支持离线使用

- Dockerfile 构建时预下载 faster-whisper 模型到镜像
- docker-compose 增加 ASR_MODEL_CACHE_DIR 环境变量
- asr_service 支持从缓存目录加载本地模型，无需运行时联网下载

Co-authored-by: Cursor <cursoragent@cursor.com>

api/services/asr_service.py

@@ -16,6 +16,8 @@ logger = logging.getLogger(__name__)
 ASR_MODEL_SIZE = os.getenv("ASR_MODEL_SIZE", "small")
 ASR_DEVICE = os.getenv("ASR_DEVICE", "auto")  # auto, cpu, cuda
 ASR_COMPUTE_TYPE = os.getenv("ASR_COMPUTE_TYPE", "auto")  # auto, int8, float16, float32
+# 镜像内预置模型目录，设置后直接使用本地模型不联网下载（与 Dockerfile 中 download_root 一致）
+ASR_MODEL_CACHE_DIR = os.getenv("ASR_MODEL_CACHE_DIR")
 
 
 class ASRService:
@@ -58,10 +60,14 @@ class ASRService:
                 else:
                     compute_type = "int8"  # CPU 使用 int8 量化，速度更快
             
+            download_root = ASR_MODEL_CACHE_DIR if ASR_MODEL_CACHE_DIR else None
+            local_files_only = bool(ASR_MODEL_CACHE_DIR)
             self.model = WhisperModel(
                 ASR_MODEL_SIZE,
                 device=device,
-                compute_type=compute_type
+                compute_type=compute_type,
+                download_root=download_root,
+                local_files_only=local_files_only,
             )
             
             self._model_loaded = True