feat(qwen-audio): add Alibaba Cloud Qwen-Audio Realtime service integration

- Introduced Qwen-Audio Realtime service for speech-to-speech processing in Pipecat.
- Updated interface catalog to include Qwen-Audio Realtime capabilities.
- Enhanced model resource testing to support new service.
- Added configuration options for audio sample rates and turn detection modes.
- Updated documentation to reflect integration details and usage instructions.

backend/services/pipecat/service_factory.py

@@ -178,8 +178,24 @@ def create_tts(cfg: AssistantConfig):
     )
 
 
+def realtime_audio_sample_rates(cfg: AssistantConfig) -> tuple[int, int]:
+    """Resolve provider defaults once for both the service and pipeline.
+
+    Values remain overridable for future model versions with different audio
+    formats, while today's Qwen-Audio contract defaults to 16 kHz in/24 kHz out.
+    """
+    default_input_rate = (
+        16_000 if cfg.realtime_interface_type == "qwen-audio-realtime" else 24_000
+    )
+    return (
+        int(cfg.realtime_values.get("inputSampleRate") or default_input_rate),
+        int(cfg.realtime_values.get("outputSampleRate") or 24_000),
+    )
+
+
 def create_realtime_service(cfg: AssistantConfig, *, instructions: str):
     """Create a speech-to-speech service that owns STT, LLM, and TTS."""
+    input_sample_rate, output_sample_rate = realtime_audio_sample_rates(cfg)
     if cfg.realtime_interface_type == "stepfun-realtime":
         from services.pipecat.stepfun_realtime import StepFunRealtimeService
 
@@ -189,12 +205,8 @@ def create_realtime_service(cfg: AssistantConfig, *, instructions: str):
             base_url=_require(cfg.realtime_base_url, "Realtime apiUrl"),
             instructions=instructions,
             voice=str(cfg.realtime_values.get("voice") or "linjiajiejie"),
-            input_sample_rate=int(
-                cfg.realtime_values.get("inputSampleRate") or 24000
-            ),
-            output_sample_rate=int(
-                cfg.realtime_values.get("outputSampleRate") or 24000
-            ),
+            input_sample_rate=input_sample_rate,
+            output_sample_rate=output_sample_rate,
             prefix_padding_ms=int(
                 cfg.realtime_values.get("prefixPaddingMs") or 500
             ),
@@ -205,4 +217,26 @@ def create_realtime_service(cfg: AssistantConfig, *, instructions: str):
                 cfg.realtime_values.get("energyAwakenessThreshold") or 2500
             ),
         )
+    if cfg.realtime_interface_type == "qwen-audio-realtime":
+        from services.pipecat.qwen_audio_realtime import QwenAudioRealtimeService
+
+        return QwenAudioRealtimeService(
+            api_key=_require(cfg.realtime_api_key, "Realtime apiKey"),
+            model=_require(cfg.realtimeModel, "Realtime modelId"),
+            base_url=_require(cfg.realtime_base_url, "Realtime apiUrl"),
+            instructions=instructions,
+            voice=str(cfg.realtime_values.get("voice") or "longanqian"),
+            input_sample_rate=input_sample_rate,
+            output_sample_rate=output_sample_rate,
+            turn_detection_mode=str(
+                cfg.realtime_values.get("turnDetection") or "server_vad"
+            ),
+            vad_threshold=float(cfg.realtime_values.get("vadThreshold", 0.5)),
+            silence_duration_ms=int(
+                cfg.realtime_values.get("silenceDurationMs") or 800
+            ),
+            max_history_turns=int(
+                cfg.realtime_values.get("maxHistoryTurns") or 20
+            ),
+        )
     raise ValueError(f"不支持的 Realtime 接口类型: {cfg.realtime_interface_type}")