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):
         """

test/unit/agents/perception/vad_agent/test_vad_streaming.py

@@ -24,7 +24,9 @@ def audio_out_socket():
 
 @pytest.fixture
 def vad_agent(audio_out_socket):
-    return VADAgent("tcp://localhost:5555", False)
+    agent = VADAgent("tcp://localhost:5555", False)
+    agent._internal_pub_socket = AsyncMock()
+    return agent
 
 
 @pytest.fixture(autouse=True)
@@ -44,6 +46,12 @@ def patch_settings(monkeypatch):
     monkeypatch.setattr(vad_agent.settings.vad_settings, "sample_rate_hz", 16_000, raising=False)
 
 
+@pytest.fixture(autouse=True)
+def mock_experiment_logger():
+    with patch("control_backend.agents.perception.vad_agent.experiment_logger") as logger:
+        yield logger
+
+
 async def simulate_streaming_with_probabilities(streaming, probabilities: list[float]):
     """
     Simulates a streaming scenario with given VAD model probabilities for testing purposes.
@@ -84,14 +92,15 @@ async def test_voice_activity_detected(audio_out_socket, vad_agent):
     Test a scenario where there is voice activity detected between silences.
     """
     speech_chunk_count = 5
-    probabilities = [0.0] * 5 + [1.0] * speech_chunk_count + [0.0] * 5
+    begin_silence_chunks = settings.behaviour_settings.vad_begin_silence_chunks
+    probabilities = [0.0] * 15 + [1.0] * speech_chunk_count + [0.0] * 5
     vad_agent.audio_out_socket = audio_out_socket
     await simulate_streaming_with_probabilities(vad_agent, probabilities)
 
     audio_out_socket.send.assert_called_once()
     data = audio_out_socket.send.call_args[0][0]
     assert isinstance(data, bytes)
-    assert len(data) == 512 * 4 * (speech_chunk_count + 1)
+    assert len(data) == 512 * 4 * (begin_silence_chunks + speech_chunk_count)
 
 
 @pytest.mark.asyncio
@@ -101,8 +110,9 @@ async def test_voice_activity_short_pause(audio_out_socket, vad_agent):
     short pause.
     """
     speech_chunk_count = 5
+    begin_silence_chunks = settings.behaviour_settings.vad_begin_silence_chunks
     probabilities = (
-        [0.0] * 5 + [1.0] * speech_chunk_count + [0.0] + [1.0] * speech_chunk_count + [0.0] * 5
+        [0.0] * 15 + [1.0] * speech_chunk_count + [0.0] + [1.0] * speech_chunk_count + [0.0] * 5
     )
     vad_agent.audio_out_socket = audio_out_socket
     await simulate_streaming_with_probabilities(vad_agent, probabilities)
@@ -110,8 +120,8 @@ async def test_voice_activity_short_pause(audio_out_socket, vad_agent):
     audio_out_socket.send.assert_called_once()
     data = audio_out_socket.send.call_args[0][0]
     assert isinstance(data, bytes)
-    # Expecting 13 chunks (2*5 with speech, 1 pause between, 1 as padding)
-    assert len(data) == 512 * 4 * (speech_chunk_count * 2 + 1 + 1)
+    # Expecting 13 chunks (2*5 with speech, 1 pause between, begin_silence_chunks as padding)
+    assert len(data) == 512 * 4 * (speech_chunk_count * 2 + 1 + begin_silence_chunks)
 
 
 @pytest.mark.asyncio