feat: allow no audio input while robot is speaking

The VAD agent will discard its current buffer and retry receiving data.

ref: N25B-213

src/control_backend/agents/vad_agent.py

@@ -53,20 +53,14 @@ class Streaming(CyclicBehaviour):
         self.audio_out_socket = audio_out_socket
 
         self.audio_buffer = np.array([], dtype=np.float32)
-        self.i_since_data = 0  # Used to avoid logging every cycle if audio input stops
         self.i_since_speech = 100  # Used to allow small pauses in speech
 
     async def run(self) -> None:
         data = await self.audio_in_poller.poll()
         if data is None:
-            if self.i_since_data % 10 == 0:
-                logger.debug(
-                    "Failed to receive audio from socket for %d ms.",
-                    self.audio_in_poller.timeout_ms * (self.i_since_data + 1),
-                )
-            self.i_since_data += 1
+            logger.debug("No audio data received. Discarding buffer until new data arrives.")
+            self.audio_buffer = np.array([], dtype=np.float32)
             return
-        self.i_since_data = 0
 
         # copy otherwise Torch will be sad that it's immutable
         chunk = np.frombuffer(data, dtype=np.float32).copy()

test/unit/agents/test_vad_streaming.py

@@ -86,9 +86,7 @@ async def test_no_data(audio_in_socket, audio_out_socket, streaming):
     audio_in_poller.poll.return_value = None
     streaming.audio_in_poller = audio_in_poller
 
-    assert streaming.i_since_data == 0
-
     await streaming.run()
 
     audio_out_socket.send.assert_not_called()
-    assert streaming.i_since_data == 1
+    assert len(streaming.audio_buffer) == 0