Fix Unbounded _audio_buffer growth.

core/duplex_pipeline.py

@@ -108,6 +108,8 @@ class DuplexPipeline:
         self._is_bot_speaking = False
         self._current_turn_task: Optional[asyncio.Task] = None
         self._audio_buffer: bytes = b""
+        max_buffer_seconds = settings.max_audio_buffer_seconds if hasattr(settings, "max_audio_buffer_seconds") else 30
+        self._max_audio_buffer_bytes = int(settings.sample_rate * 2 * max_buffer_seconds)
         self._last_vad_status: str = "Silence"
         
         # Interruption handling
@@ -263,6 +265,9 @@ class DuplexPipeline:
             # 3. Buffer audio for ASR
             if vad_status == "Speech" or self.conversation.state == ConversationState.LISTENING:
                 self._audio_buffer += pcm_bytes
+                if len(self._audio_buffer) > self._max_audio_buffer_bytes:
+                    # Keep only the most recent audio to cap memory usage
+                    self._audio_buffer = self._audio_buffer[-self._max_audio_buffer_bytes:]
                 await self.asr_service.send_audio(pcm_bytes)
                 
                 # For SiliconFlow ASR, trigger interim transcription periodically
@@ -669,6 +674,12 @@ class DuplexPipeline:
                 await self._current_turn_task
             except asyncio.CancelledError:
                 pass
+
+        # Ensure underlying services are cancelled to avoid leaking work/audio
+        if self.tts_service:
+            await self.tts_service.cancel()
+        if self.llm_service and hasattr(self.llm_service, 'cancel'):
+            self.llm_service.cancel()
         
         self._is_bot_speaking = False
         self._interrupt_event.clear()