chore: re-addd more silence before speech audio

src/control_backend/agents/perception/vad_agent.py

@@ -285,9 +285,10 @@ class VADAgent(BaseAgent):
                 assert self.audio_out_socket is not None
                 await self.audio_out_socket.send(self.audio_buffer[: -2 * len(chunk)].tobytes())
 
-            # At this point, we know that the speech has ended.
-            # Prepend the last chunk that had no speech, for a more fluent boundary
-            self.audio_buffer = chunk
+            # At this point, we know that there is no speech.
+            # Prepend the last few chunks that had no speech, for a more fluent boundary.
+            self.audio_buffer = np.append(self.audio_buffer, chunk)
+            self.audio_buffer = self.audio_buffer[-begin_silence_length * len(chunk) :]
 
     async def handle_message(self, msg: InternalMessage):
         """