ASR: add Qwen3-ASR engine, default zh/en to it on this branch

- New QwenASRClient (Qwen/Qwen3-ASR-1.7B via qwen-asr, transformers backend, MPS bf16)
- Language mapping zh/en -> qwen; VOICE_DIALOGUE_ASR=legacy restores funasr/whisper
- ASRService now passes session language to transcribe()

Requires: uv pip install qwen-asr (upgrades transformers to 4.57)

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

src/voice_dialogue/asr/manager.py

@@ -1,5 +1,6 @@
 import importlib.util
 import inspect
+import os
 import re
 from dataclasses import dataclass
 from typing import Dict, Type, List, Literal, Optional
@@ -92,11 +93,17 @@ class ASRManager:
 
     def __init__(self):
         self._asr_instances: Dict[str, ASRInterface] = {}
-        self._language_to_asr_mapping = {
-            'zh': 'funasr',  # 中文优先使用FunASR
-            'en': 'whisper',  # 英文优先使用Whisper
-            # 'auto': 'whisper',  # 自动检测默认使用Whisper
-        }
+        # 本分支默认使用 Qwen3-ASR；设置 VOICE_DIALOGUE_ASR=legacy 可切回原引擎做 A/B 对比
+        if os.environ.get('VOICE_DIALOGUE_ASR', 'qwen') == 'legacy':
+            self._language_to_asr_mapping = {
+                'zh': 'funasr',  # 中文优先使用FunASR
+                'en': 'whisper',  # 英文优先使用Whisper
+            }
+        else:
+            self._language_to_asr_mapping = {
+                'zh': 'qwen',
+                'en': 'qwen',
+            }
 
     def create_asr(self, language: Literal['auto', 'zh', 'en']) -> ASRInterface:
         """

src/voice_dialogue/asr/models/__init__.py

@@ -19,3 +19,12 @@ except ImportError as e:
     from voice_dialogue.utils.logger import logger
 
     logger.warning(f"Failed to import some Whisper implementations: {e}")
+
+try:
+    from .qwen import QwenASRClient
+
+    __all__.append('QwenASRClient')
+except ImportError as e:
+    from voice_dialogue.utils.logger import logger
+
+    logger.warning(f"Failed to import some Qwen ASR implementations: {e}")

src/voice_dialogue/asr/models/qwen.py

@@ -0,0 +1,70 @@
+import os
+import typing
+
+import numpy as np
+import torch
+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.utils.logger import logger
+
+# Qwen3-ASR 的 language 参数使用语言全名
+LANGUAGE_NAME_MAPPING = {
+    'zh': 'Chinese',
+    'en': 'English',
+}
+
+DEFAULT_MODEL = os.environ.get('QWEN_ASR_MODEL', 'Qwen/Qwen3-ASR-1.7B')
+
+TARGET_SAMPLE_RATE = 16000
+
+
+@asr_tables.register('asr_classes', 'qwen')
+class QwenASRClient(ASRInterface):
+    """Qwen3-ASR 客户端（transformers 后端，macOS 上使用 MPS 加速）"""
+    supported_langs = ['zh', 'en']
+
+    def __init__(self):
+        super().__init__()
+        self.model: typing.Optional[Qwen3ASRModel] = None
+
+    def setup(self, **kwargs) -> None:
+        model_name = kwargs.get('model', DEFAULT_MODEL)
+
+        if torch.backends.mps.is_available():
+            device_map, dtype = 'mps', torch.bfloat16
+        elif torch.cuda.is_available():
+            device_map, dtype = 'cuda:0', torch.bfloat16
+        else:
+            device_map, dtype = 'cpu', torch.float32
+
+        logger.info(f'[INFO] Loading Qwen3-ASR model: {model_name} (device={device_map}, dtype={dtype})')
+        self.model = Qwen3ASRModel.from_pretrained(
+            model_name,
+            dtype=dtype,
+            device_map=device_map,
+            max_inference_batch_size=1,
+            max_new_tokens=256,
+        )
+
+    def warmup(self) -> None:
+        logger.info('[INFO] Warming up Qwen3-ASR model...')
+        try:
+            self.transcribe(self.warmup_audiodata)
+            logger.info('[INFO] Qwen3-ASR model warmed up.')
+        except Exception as e:
+            logger.warning(f'[WARNING] Qwen3-ASR model warmup failed: {e}')
+
+    def transcribe(self, audio_array: np.ndarray, language: str = None) -> str:
+        audio_array = ensure_minimum_audio_duration(audio_array)
+
+        # 未指定语言时交给模型自动检测（Qwen3-ASR 自带语种识别）
+        qwen_language = LANGUAGE_NAME_MAPPING.get(language)
+
+        results = self.model.transcribe(
+            audio=(audio_array, TARGET_SAMPLE_RATE),
+            language=qwen_language,
+        )
+        return ' '.join(result.text for result in results).strip()

src/voice_dialogue/services/asr_service.py

@@ -42,7 +42,7 @@ class ASRService(BaseThread, PerformanceLogMixin):
             voice_task.whisper_start_time = time.time()
 
             user_voice: np.array = voice_task.user_voice
-            transcribed_text = self.client.transcribe(user_voice)
+            transcribed_text = self.client.transcribe(user_voice, language=self.language)
             if not transcribed_text.strip():
                 voice_state_manager.reset_task_id()
                 continue