Enable dual audio input channel (#334)

This PR adds the ability to choose the microphone stream that works best with a particular STT engine using the new dual audio input channel feature.

linux_voice_assistant/__main__.py

@@ -74,6 +74,13 @@ async def main() -> None:
     parser.add_argument("--mic-volume", type=int, default=100, choices=list(range(1, 101)), help="Microphone volume level (1 to 100)")
     parser.add_argument("--mic-auto-gain", type=int, default=0, choices=list(range(32)))
     parser.add_argument("--mic-noise-suppression", type=int, default=0, choices=(0, 1, 2, 3, 4))
+    parser.add_argument(
+        "--audio-input-channels",
+        type=int,
+        default=1,
+        choices=(1, 2),
+        help="Number of mic channels to capture and stream (1=mono, 2=dual-channel voice)",
+    )
     parser.add_argument(
         "--wake-word-dir",
         default=[_WAKEWORDS_DIR],
@@ -328,6 +335,7 @@ async def main() -> None:
         mic_volume=preferences.mic_volume,
         mic_auto_gain=preferences.mic_auto_gain,
         mic_noise_suppression=preferences.mic_noise_suppression,
+        audio_input_channels=args.audio_input_channels,
         timer_max_ring_seconds=args.timer_max_ring_seconds,
     )
 
@@ -413,6 +421,7 @@ async def main() -> None:
 
 def process_audio(state: ServerState, mic, block_size: int):
     """Process audio chunks from the microphone."""
+    n_channels = state.audio_input_channels
 
     wake_words: List[Union[MicroWakeWord, OpenWakeWord]] = []
     micro_features: Optional[MicroWakeWordFeatures] = None
@@ -427,12 +436,23 @@ def process_audio(state: ServerState, mic, block_size: int):
 
     try:
         _LOGGER.debug("Opening audio input device: %s", mic.name)
-        with mic.recorder(samplerate=16000, channels=1, blocksize=block_size) as mic_in:
+        with mic.recorder(samplerate=16000, channels=n_channels, blocksize=block_size) as mic_in:
             while True:
-                audio_chunk_array = mic_in.record(block_size).reshape(-1)
-                # little-endian 16-bit signed
+                # Shape: (block_size, n_channels) for stereo, (block_size, 1) for mono.
+                raw = mic_in.record(block_size)  # float32, range [-1, 1]
                 mic_vol_scalar = max(0.1, min(1.0, state.mic_volume / 100.0))
-                audio_chunk = (np.clip(audio_chunk_array * mic_vol_scalar, -1.0, 1.0) * 32767.0).astype("<i2").tobytes()
+
+                # Build per-channel byte arrays.  Channel 0 is the primary
+                # microphone; channel 1 (when present) is the reference/speaker
+                # feed used for server-side AEC.
+                channel_chunks: list[bytes] = []
+                for ch in range(n_channels):
+                    col = raw[:, ch] if n_channels > 1 else raw.reshape(-1)
+                    chunk = (np.clip(col * mic_vol_scalar, -1.0, 1.0) * 32767.0).astype("<i2").tobytes()
+                    channel_chunks.append(chunk)
+
+                # Primary channel drives WebRTC and wake-word detection.
+                audio_chunk = channel_chunks[0]
                 agc = state.preferences.mic_auto_gain or 0
                 ns = state.preferences.mic_noise_suppression or 0
 
@@ -532,7 +552,9 @@ def process_audio(state: ServerState, mic, block_size: int):
                         oww_features = OpenWakeWordFeatures.from_builtin()
 
                 try:
-                    state.satellite.handle_audio(audio_chunk)
+                    # Both channels travel in one message: data=ch0 (enhanced), data2=ch1 (raw reference)
+                    audio_chunk_2 = channel_chunks[1] if n_channels >= 2 else None
+                    state.satellite.handle_audio(audio_chunk, audio_chunk_2)
 
                     assert micro_features is not None
                     micro_inputs.clear()

linux_voice_assistant/models.py

@@ -125,6 +125,7 @@ class ServerState:
     mic_auto_gain: int = 0
     mic_noise_suppression: int = 0
     mic_volume: int = 100  # 1–100, default maximum
+    audio_input_channels: int = 2  # number of mic channels to stream
     timer_max_ring_seconds: float = 900.0
 
     def save_preferences(self) -> None:

linux_voice_assistant/satellite.py

@@ -63,6 +63,8 @@ _LOGGER = logging.getLogger(__name__)
 
 PROTO_TO_MESSAGE_TYPE = {v: k for k, v in MESSAGE_TYPE_TO_PROTO.items()}
 
+_HAS_AUDIO_DATA2 = "data2" in {f.name for f in VoiceAssistantAudio.DESCRIPTOR.fields}
+
 
 class VoiceSatelliteProtocol(APIServer):
 
@@ -82,6 +84,9 @@ class VoiceSatelliteProtocol(APIServer):
             self.supported_features = (
                 VoiceAssistantFeature.VOICE_ASSISTANT | VoiceAssistantFeature.API_AUDIO | VoiceAssistantFeature.ANNOUNCE | VoiceAssistantFeature.START_CONVERSATION | VoiceAssistantFeature.TIMERS
             )
+            # Channel 1 carries echo-reference audio; advertise SPEAKER so HA knows to use it for server-side AEC.
+            if state.audio_input_channels >= 2:
+                self.supported_features |= VoiceAssistantFeature.MULTI_CHANNEL_AUDIO  # pylint: disable=no-member
 
         existing_media_players = [entity for entity in self.state.entities if isinstance(entity, MediaPlayerEntity)]
 
@@ -607,12 +612,14 @@ class VoiceSatelliteProtocol(APIServer):
             self.state.save_preferences()
             self.state.wake_words_changed = True
 
-    def handle_audio(self, audio_chunk: bytes) -> None:
-
+    # handle_audio — both channels in ONE message
+    def handle_audio(self, audio_chunk: bytes, audio_chunk_2: Optional[bytes] = None) -> None:
         if not self._is_streaming_audio or self.state.muted:
             return
-
-        self.send_messages([VoiceAssistantAudio(data=audio_chunk)])
+        if _HAS_AUDIO_DATA2 and audio_chunk_2 is not None:
+            self.send_messages([VoiceAssistantAudio(data=audio_chunk, data2=audio_chunk_2)])
+        else:
+            self.send_messages([VoiceAssistantAudio(data=audio_chunk)])
 
     def wakeup(self, wake_word: Union[MicroWakeWord, OpenWakeWord]) -> None:
         if self._timer_finished: