Refactor LLM processor and improve async handling

Move contextual information handling from noise filtering to extraction
and centralize LLM call logic. Wrap blocking transcription and state
update calls in asyncio.to_thread to prevent event loop blocking.
Update transcriber model size to base.

src/pipeline/orchestrator.py

@@ -44,7 +44,7 @@ class PipelineOrchestrator:
 
         # Modules
         self.listener = AudioListener(loop=self.loop)
-        self.transcriber = Transcriber(model_size="small")
+        self.transcriber = Transcriber(model_size="base")
         self.processor = LLMProcessor()
         self.rag_manager = RAGManager()
 
@@ -107,7 +107,9 @@ class PipelineOrchestrator:
                 full_audio = np.concatenate(self.audio_buffer)
 
                 # Transcribe (WhisperX now returns a list of (speaker, text, start, end))
-                results = self.transcriber.transcribe(full_audio)
+                results = await asyncio.to_thread(
+                    self.transcriber.transcribe, full_audio
+                )
 
                 # Filter for only new segments that start after the last processed segment
                 new_segments = [