feat: prepend more silence before speech audio for better transcription beginnings

ref: N25B-429

.env.example

@@ -10,7 +10,7 @@ LLM_SETTINGS__LOCAL_LLM_URL="http://localhost:1234/v1/chat/completions"
 LLM_SETTINGS__LOCAL_LLM_MODEL="gpt-oss"
 
 # Number of non-speech chunks to wait before speech ended. A chunk is approximately 31 ms. Increasing this number allows longer pauses in speech, but also increases response time.
-BEHAVIOUR_SETTINGS__VAD_NON_SPEECH_PATIENCE_CHUNKS=3
+BEHAVIOUR_SETTINGS__VAD_NON_SPEECH_PATIENCE_CHUNKS=15
 
 # Timeout in milliseconds for socket polling. Increase this number if network latency/jitter is high, often the case when using Wi-Fi. Perhaps 500 ms. A symptom of this issue is transcriptions getting cut off.
 BEHAVIOUR_SETTINGS__SOCKET_POLLER_TIMEOUT_MS=100

src/control_backend/agents/perception/vad_agent.py

@@ -229,10 +229,11 @@ class VADAgent(BaseAgent):
             assert self.model is not None
             prob = self.model(torch.from_numpy(chunk), settings.vad_settings.sample_rate_hz).item()
             non_speech_patience = settings.behaviour_settings.vad_non_speech_patience_chunks
+            begin_silence_length = settings.behaviour_settings.vad_begin_silence_chunks
             prob_threshold = settings.behaviour_settings.vad_prob_threshold
 
             if prob > prob_threshold:
-                if self.i_since_speech > non_speech_patience:
+                if self.i_since_speech > non_speech_patience + begin_silence_length:
                     self.logger.debug("Speech started.")
                 self.audio_buffer = np.append(self.audio_buffer, chunk)
                 self.i_since_speech = 0
@@ -246,11 +247,12 @@ class VADAgent(BaseAgent):
                 continue
 
             # Speech probably ended. Make sure we have a usable amount of data.
-            if len(self.audio_buffer) >= 3 * len(chunk):
+            if len(self.audio_buffer) > begin_silence_length * len(chunk):
                 self.logger.debug("Speech ended.")
                 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) :]

src/control_backend/core/config.py

@@ -73,6 +73,7 @@ class BehaviourSettings(BaseModel):
     :ivar vad_prob_threshold: Probability threshold for Voice Activity Detection.
     :ivar vad_initial_since_speech: Initial value for 'since speech' counter in VAD.
     :ivar vad_non_speech_patience_chunks: Number of non-speech chunks to wait before speech ended.
+    :ivar vad_begin_silence_chunks: The number of chunks of silence to prepend to speech chunks.
     :ivar transcription_max_concurrent_tasks: Maximum number of concurrent transcription tasks.
     :ivar transcription_words_per_minute: Estimated words per minute for transcription timing.
     :ivar transcription_words_per_token: Estimated words per token for transcription timing.
@@ -90,6 +91,7 @@ class BehaviourSettings(BaseModel):
     vad_prob_threshold: float = 0.5
     vad_initial_since_speech: int = 100
     vad_non_speech_patience_chunks: int = 15
+    vad_begin_silence_chunks: int = 3
 
     # transcription behaviour
     transcription_max_concurrent_tasks: int = 3