refactor StreamingHandler to use LazyModel for resource mgmt

Author: janaab11

src/diart/console/serve.py

@@ -6,7 +6,7 @@
 from diart import argdoc
 from diart import models as m
 from diart import utils
-from diart.handler import StreamingInferenceConfig, StreamingInferenceHandler
+from diart.handler import StreamingHandlerConfig, StreamingHandler
 
 
 def run():
@@ -94,24 +94,23 @@ def run():
     args.segmentation = m.SegmentationModel.from_pretrained(args.segmentation, hf_token)
     args.embedding = m.EmbeddingModel.from_pretrained(args.embedding, hf_token)
 
-    # Resolve pipeline
+    # Resolve pipeline configuration
     pipeline_class = utils.get_pipeline_class(args.pipeline)
-    config = pipeline_class.get_config_class()(**vars(args))
-    pipeline = pipeline_class(config)
+    pipeline_config = pipeline_class.get_config_class()(**vars(args))
 
-    # Create inference configuration
-    inference_config = StreamingInferenceConfig(
-        pipeline=pipeline,
+    # Create handler configuration for inference
+    config = StreamingHandlerConfig(
+        pipeline_class=pipeline_class,
+        pipeline_config=pipeline_config,
         batch_size=1,
         do_profile=False,
         do_plot=False,
         show_progress=False,
     )
 
-    # Initialize handler with new configuration
-    handler = StreamingInferenceHandler(
-        inference_config=inference_config,
-        sample_rate=config.sample_rate,
+    # Initialize handler
+    handler = StreamingHandler(
+        config=config,
         host=args.host,
         port=args.port,
     )
@@ -121,3 +120,18 @@ def run():
 
 if __name__ == "__main__":
     run()
+
+# diart.serve --host 0.0.0.0 --port 7007 --segmentation pyannote/segmentation-3.0 --embedding speechbrain/spkrec-resnet-voxceleb --tau-active 0.55 --rho-update 0.3 --delta-new 0.7 --latency 2
+# diart.serve --host 0.0.0.0 --port 7007 --segmentation pyannote/segmentation-3.0 --embedding speechbrain/spkrec-ecapa-voxceleb --tau-active 0.65 --rho-update 0.3 --delta-new 0.7 --latency 2
+# diart.serve --host 0.0.0.0 --port 7007 --segmentation pyannote/segmentation-3.0 --embedding speechbrain/spkrec-resnet-voxceleb --tau-active 0.53 --rho-update 0.15 --delta-new 0.83 --latency 2
+
+# 2 speakers : 
+# diart.serve --host 0.0.0.0 --port 7007 --segmentation pyannote/segmentation-3.0 --embedding speechbrain/spkrec-resnet-voxceleb --tau-active 0.55 --rho-update 0.15 --delta-new 0.7 --latency 3 --max-speakers 2
+
+# 3 speakers :
+# diart.serve --host 0.0.0.0 --port 7007 --segmentation pyannote/segmentation-3.0 --embedding speechbrain/spkrec-ecapa-voxceleb --tau-active 0.45 --rho-update 0.15 --delta-new 0.6 --latency 5 --max-speakers 3
+# diart.serve --host 0.0.0.0 --port 7007 --segmentation pyannote/segmentation-3.0 --embedding speechbrain/spkrec-ecapa-voxceleb --tau-active 0.25 --rho-update 0.45 --delta-new 0.6 --latency 5 --max-speakers 4
+# diart.serve --host 0.0.0.0 --port 7007 --segmentation pyannote/segmentation-3.0 --embedding speechbrain/spkrec-ecapa-voxceleb --tau-active 0.3 --rho-update 0.3 --delta-new 0.8 --latency 5 --max-speakers 4
+
+# 4 speakers :
+# diart.serve --host 0.0.0.0 --port 7007 --segmentation pyannote/segmentation-3.0 --embedding speechbrain/spkrec-ecapa-voxceleb --tau-active 0.2 --rho-update 0.5 --delta-new 0.7 --latency 5 --max-speakers 4

src/diart/handler.py

@@ -19,29 +19,15 @@
 
 
 @dataclass
-class WebSocketAudioSourceConfig:
-    """Configuration for WebSocket audio source.
-
-    Parameters
-    ----------
-    uri : str
-        WebSocket URI for the audio source
-    sample_rate : int
-        Audio sample rate in Hz
-    """
-
-    uri: str
-    sample_rate: int = 16000
-
-
-@dataclass
-class StreamingInferenceConfig:
+class StreamingHandlerConfig:
     """Configuration for streaming inference.
 
     Parameters
     ----------
-    pipeline : blocks.Pipeline
-        Diarization pipeline configuration
+    pipeline_class : type
+        Pipeline class
+    pipeline_config : blocks.PipelineConfig
+        Pipeline configuration
     batch_size : int
         Number of inputs to process at once
     do_profile : bool
@@ -54,7 +40,8 @@ class StreamingInferenceConfig:
         Custom progress bar implementation
     """
 
-    pipeline: blocks.Pipeline
+    pipeline_class: type
+    pipeline_config: blocks.PipelineConfig
     batch_size: int = 1
     do_profile: bool = True
     do_plot: bool = False
@@ -70,18 +57,16 @@ class ClientState:
     inference: StreamingInference
 
 
-class StreamingInferenceHandler:
+class StreamingHandler:
     """Handles real-time speaker diarization inference for multiple audio sources over WebSocket.
 
     This handler manages WebSocket connections from multiple clients, processing
     audio streams and performing speaker diarization in real-time.
 
     Parameters
     ----------
-    inference_config : StreamingInferenceConfig
+    config : StreamingHandlerConfig
         Streaming inference configuration
-    sample_rate : int, optional
-        Audio sample rate in Hz, by default 16000
     host : str, optional
         WebSocket server host, by default "127.0.0.1"
     port : int, optional
@@ -94,15 +79,13 @@ class StreamingInferenceHandler:
 
     def __init__(
         self,
-        inference_config: StreamingInferenceConfig,
-        sample_rate: int = 16000,
+        config: StreamingHandlerConfig,
         host: Text = "127.0.0.1",
         port: int = 7007,
         key: Optional[Union[Text, Path]] = None,
         certificate: Optional[Union[Text, Path]] = None,
     ):
-        self.inference_config = inference_config
-        self.sample_rate = sample_rate
+        self.config = config
         self.host = host
         self.port = port
 
@@ -135,26 +118,21 @@ def _create_client_state(self, client_id: Text) -> ClientState:
         """
         # Create a new pipeline instance with the same config
         # This ensures each client has its own state while sharing model weights
-        pipeline = self.inference_config.pipeline.__class__(
-            self.inference_config.pipeline.config
-        )
-
-        audio_config = WebSocketAudioSourceConfig(
-            uri=f"{self.uri}:{client_id}", sample_rate=self.sample_rate
-        )
+        pipeline = self.config.pipeline_class(self.config.pipeline_config)
 
         audio_source = src.WebSocketAudioSource(
-            uri=audio_config.uri, sample_rate=audio_config.sample_rate
+            uri=f"{self.uri}:{client_id}",
+            sample_rate=self.config.pipeline_config.sample_rate,
         )
 
         inference = StreamingInference(
             pipeline=pipeline,
             source=audio_source,
-            batch_size=self.inference_config.batch_size,
-            do_profile=self.inference_config.do_profile,
-            do_plot=self.inference_config.do_plot,
-            show_progress=self.inference_config.show_progress,
-            progress_bar=self.inference_config.progress_bar,
+            batch_size=self.config.batch_size,
+            do_profile=self.config.do_profile,
+            do_plot=self.config.do_plot,
+            show_progress=self.config.show_progress,
+            progress_bar=self.config.progress_bar,
         )
 
         return ClientState(audio_source=audio_source, inference=inference)
@@ -174,16 +152,15 @@ def _on_connect(self, client: Dict[Text, Any], server: WebsocketServer) -> None:
 
         if client_id not in self._clients:
             try:
-                client_state = self._create_client_state(client_id)
-                self._clients[client_id] = client_state
+                self._clients[client_id] = self._create_client_state(client_id)
 
                 # Setup RTTM response hook
-                client_state.inference.attach_hooks(
+                self._clients[client_id].inference.attach_hooks(
                     lambda ann_wav: self.send(client_id, ann_wav[0].to_rttm())
                 )
 
                 # Start inference
-                client_state.inference()
+                self._clients[client_id].inference()
                 logger.info(f"Started inference for client: {client_id}")
 
                 # Send ready notification to client