move bargein detector to inference

Author: chenghao-mou

examples/bargein_server.py

@@ -53,13 +53,15 @@ class BargeinRequest(BaseModel):
     agentId: str | None = Field(None, description="Optional agent ID")
     threshold: float = Field(0.95, description="Threshold for bargein detection")
     min_frames: int = Field(2, description="Minimum number of frames for bargein detection")
+    created_at: float = Field(..., description="Timestamp of the audio waveform")
 
 
 class BargeinResponse(BaseModel):
     """Response model for bargein detection."""
 
     is_bargein: bool = Field(..., description="Whether bargein is detected")
     confidence: float | None = Field(None, description="Confidence score (optional)")
+    created_at: float = Field(..., description="Timestamp of the data creation")
 
 
 def load_onnx_model(model_path: str) -> ort.InferenceSession:
@@ -165,7 +167,9 @@ async def detect_bargein(request: BargeinRequest) -> BargeinResponse:
 
         logger.info(f"Bargein detection result: is_bargein={is_bargein}")
 
-        return BargeinResponse(is_bargein=is_bargein, confidence=None)
+        return BargeinResponse(
+            is_bargein=is_bargein, confidence=None, created_at=request.created_at
+        )
 
     except Exception as e:
         logger.error(f"Error during bargein detection: {e}", exc_info=True)
@@ -246,7 +250,12 @@ async def websocket_bargein(websocket: WebSocket) -> None:
                         delta = time.time() - created_at
 
                         await websocket.send_json(
-                            {"type": "inference_done", "delta": delta, "is_bargein": is_bargein}
+                            {
+                                "type": "inference_done",
+                                "delta": delta,
+                                "is_bargein": is_bargein,
+                                "created_at": created_at,
+                            }
                         )
 
                         if is_bargein:

livekit-agents/livekit/agents/inference/bargein.py

@@ -41,7 +41,7 @@
 THRESHOLD = 0.95
 MIN_BARGEIN_DURATION = 0.05  # 25ms per frame
 MAX_WINDOW_SIZE = 3 * 16000  # 3 seconds at 16000 Hz
-STEP_SIZE = int(0.2 * 16000)  # 0.2 second at 16000 Hz
+STEP_SIZE = int(0.1 * 16000)  # 0.1 second at 16000 Hz
 REMOTE_INFERENCE_TIMEOUT = 1
 DEFAULT_BASE_URL = "https://agent-gateway.livekit.cloud/v1"
 
@@ -440,7 +440,13 @@ async def recv_task(ws: aiohttp.ClientWebSocketResponse) -> None:
                     pass
                 elif msg_type == "inference_done":
                     is_bargein_result = data.get("is_bargein", False)
-                    logger.debug("inference done", extra={"is_bargein": is_bargein_result})
+                    logger.debug(
+                        "inference done",
+                        extra={
+                            "is_bargein": is_bargein_result,
+                            "inference_duration": time.time() - data.get("created_at", 0.0),
+                        },
+                    )
                     self._event_ch.send_nowait(
                         BargeinEvent(
                             type=BargeinEventType.INFERENCE_DONE,
@@ -489,7 +495,7 @@ async def _connect_ws(self) -> aiohttp.ClientWebSocketResponse:
         """Connect to the LiveKit STT WebSocket."""
         params: dict[str, Any] = {
             "settings": {
-                "sample_rate": str(self._opts.sample_rate),
+                "sample_rate": self._opts.sample_rate,
             },
         }