ASRC: Adds NONBLOCK read from ALSA buffer; controls the amount of frames in the ALSA buffer; Adds resampling to get rid of audio glitches; no latency buildup anymore.

pyproject.toml

@@ -17,7 +17,8 @@ dependencies = [
     "sounddevice (>=0.5.2,<0.6.0)",
     "python-dotenv (>=1.1.1,<2.0.0)",
     "smbus2 (>=0.5.0,<0.6.0)",
-    "samplerate (>=0.2.2,<0.3.0)"
+    "samplerate (>=0.2.2,<0.3.0)",
+    "pyalsaaudio (>=0.9.0,<1.0.0)"
 ]
 
 [project.optional-dependencies]

src/auracast/multicast.py

@@ -30,6 +30,7 @@ import time
 import threading
 
 import numpy as np  # for audio down-mix
+import samplerate
 import os
 
 import lc3  # type: ignore  # pylint: disable=E0401
@@ -56,7 +57,7 @@ from auracast.utils.webrtc_audio_input import WebRTCAudioInput
 
 
 # Patch sounddevice.InputStream globally to use low-latency settings
-import sounddevice as sd
+import alsaaudio
 from collections import deque
 
 
@@ -139,96 +140,139 @@ class AlsaArecordAudioInput(audio_io.AudioInput):
         self._proc = None
 
 
-class ModSoundDeviceAudioInput(audio_io.SoundDeviceAudioInput):
-    """Patched SoundDeviceAudioInput with low-latency capture and adaptive resampling."""
+class PyAlsaAudioInput(audio_io.ThreadedAudioInput):
+    """PyALSA audio input with non-blocking reads - supports mono/stereo."""
 
-    def _open(self):
-        """Create RawInputStream with low-latency parameters and initialize ring buffer."""
-        dev_info = sd.query_devices(self._device)
-        hostapis = sd.query_hostapis()
-        api_index = dev_info.get('hostapi')
-        api_name = hostapis[api_index]['name'] if isinstance(api_index, int) and 0 <= api_index < len(hostapis) else 'unknown'
-        pa_ver = sd.get_portaudio_version()
+    def __init__(self, device, pcm_format: audio_io.PcmFormat):
+        super().__init__()
+        self._device = str(device) if not isinstance(device, str) else device
+        if self._device.isdigit():
+            self._device = 'default' if self._device == '0' else f'hw:{self._device}'
+        self._pcm_format = pcm_format
+        self._pcm = None
+        self._actual_channels = None
+        self._periodsize = None
+        self._hw_channels = None
+        self._first_read = True
+        self._resampler = None
+        self._resampler_buffer = np.empty(0, dtype=np.float32)
 
-        logging.info(
-            "SoundDevice backend=%s device='%s' (id=%s) ch=%s default_low_input_latency=%.4f default_high_input_latency=%.4f portaudio=%s",
-            api_name,
-            dev_info.get('name'),
-            self._device,
-            dev_info.get('max_input_channels'),
-            float(dev_info.get('default_low_input_latency') or 0.0),
-            float(dev_info.get('default_high_input_latency') or 0.0),
-            pa_ver[1] if isinstance(pa_ver, tuple) and len(pa_ver) >= 2 else pa_ver,
-        )
-        # Create RawInputStream with injected low-latency parameters
-        # Target ~2 ms blocksize (48 kHz -> 96 frames). For other rates, keep ~2 ms.
-        _sr = int(self._pcm_format.sample_rate)
-
-        self.counter=0
-        self.max_avail=0
-        self.logfile_name="available_samples.txt"
-        self.blocksize = 120
-
-        if os.path.exists(self.logfile_name):
-            os.remove(self.logfile_name)
-
-        self._stream = sd.RawInputStream(
-            samplerate=self._pcm_format.sample_rate,
+    def _open(self) -> audio_io.PcmFormat:
+        ALSA_PERIODSIZE = 240
+        ALSA_PERIODS = 4
+        ALSA_MODE = alsaaudio.PCM_NONBLOCK
+        
+        requested_rate = int(self._pcm_format.sample_rate)
+        requested_channels = int(self._pcm_format.channels)
+        self._periodsize = ALSA_PERIODSIZE
+        
+        self._pcm = alsaaudio.PCM(
+            type=alsaaudio.PCM_CAPTURE,
+            mode=ALSA_MODE,
             device=self._device,
-            channels=self._pcm_format.channels,
-            dtype='int16',
-            blocksize=self.blocksize,
-            latency=0.004,
+            periods=ALSA_PERIODS,
         )
-        self._stream.start()
+        
+        self._pcm.setchannels(requested_channels)
+        self._pcm.setformat(alsaaudio.PCM_FORMAT_S16_LE)
+        actual_rate = self._pcm.setrate(requested_rate)
+        self._pcm.setperiodsize(ALSA_PERIODSIZE)
+        
+        logging.info("PyALSA: device=%s rate=%d ch=%d periodsize=%d (%.1fms) periods=%d mode=%s", 
+                     self._device, actual_rate, requested_channels, ALSA_PERIODSIZE, 
+                     (ALSA_PERIODSIZE / actual_rate) * 1000, ALSA_PERIODS, ALSA_MODE)
+        
+        if actual_rate != requested_rate:
+            logging.warning("PyALSA: Sample rate mismatch! requested=%d actual=%d", requested_rate, actual_rate)
+        
+        self._actual_channels = requested_channels
+        self._resampler = samplerate.Resampler('sinc_fastest', channels=requested_channels)
+        self._resampler_buffer = np.empty(0, dtype=np.float32)
 
+        self._bang_bang = 0
+        
         return audio_io.PcmFormat(
             audio_io.PcmFormat.Endianness.LITTLE,
             audio_io.PcmFormat.SampleType.INT16,
-            self._pcm_format.sample_rate,
-            1,
+            actual_rate,
+            requested_channels,
         )
 
     def _read(self, frame_size: int) -> bytes:
-        """Read PCM samples from the stream."""
+        try:
+            avail = self._pcm.avail()
+            logging.debug("PyALSA: avail before read: %d", avail)
+            length, data = self._pcm.read_sw(frame_size + self._bang_bang)
+            avail = self._pcm.avail()
+            SETPOINT = 120 
+            TOLERANCE = 40 
+            if avail < SETPOINT - TOLERANCE:
+                self._bang_bang = -1 
+            elif avail > SETPOINT + TOLERANCE:
+                self._bang_bang = 1
+            else:
+                self._bang_bang = 0
+                
+                
+            logging.debug("PyALSA: read length=%d, data length=%d, avail=%d, bang_bang=%d", length, len(data), avail, self._bang_bang)
 
-        #if self.counter % 50 == 0:
-        frame_size = frame_size + 1 # consume samples a little faster to avoid latency akkumulation
+            if length > 0:
+                if self._first_read:
+                    expected_mono = self._periodsize * 2
+                    expected_stereo = self._periodsize * 2 * 2
+                    # self._hw_channels = 2 if len(data) == expected_stereo else 1
+                    self._hw_channels = 2  # TODO fix stereo detection, on first read might detect 0 data
+                    logging.info("PyALSA first read: bytes=%d detected_hw_channels=%d requested_channels=%d", 
+                                len(data), self._hw_channels, self._actual_channels)
+                    self._first_read = False
+                
+                if self._hw_channels == 2 and self._actual_channels == 1:
+                    pcm_stereo = np.frombuffer(data, dtype=np.int16)
+                    pcm_mono = pcm_stereo[::2]
+                    data = pcm_mono.tobytes()
+                
+                actual_samples = len(data) // (2 * self._actual_channels)
+                ratio = frame_size / actual_samples
+                pcm_f32 = np.frombuffer(data, dtype=np.int16).astype(np.float32) / 32768.0
+                if self._actual_channels > 1:
+                    pcm_f32 = pcm_f32.reshape(-1, self._actual_channels)
+                resampled = self._resampler.process(pcm_f32, ratio, end_of_input=False)
+                if self._actual_channels > 1:
+                    resampled = resampled.reshape(-1)
+                self._resampler_buffer = np.concatenate([self._resampler_buffer, resampled])
+            else:
+                logging.warning("PyALSA: No data read from ALSA")
+                self._resampler_buffer = np.concatenate([
+                    self._resampler_buffer,
+                    np.zeros(frame_size * self._actual_channels, dtype=np.float32),
+                ])
+        except alsaaudio.ALSAAudioError as e:
+            logging.error("PyALSA: ALSA read error: %s", e)
+            self._resampler_buffer = np.concatenate([
+                self._resampler_buffer,
+                np.zeros(frame_size * self._actual_channels, dtype=np.float32),
+            ])
 
-        pcm_buffer, overflowed = self._stream.read(frame_size)
-        if overflowed:
-            logging.warning("SoundDeviceAudioInput: overflowed")
+        needed = frame_size * self._actual_channels
+        if len(self._resampler_buffer) < needed:
+            pad = np.zeros(needed - len(self._resampler_buffer), dtype=np.float32)
+            self._resampler_buffer = np.concatenate([self._resampler_buffer, pad])
+            logging.debug("PyALSA: padded buffer with %d samples", needed - len(self._resampler_buffer))
 
-        n_available = self._stream.read_available
+        output = self._resampler_buffer[:needed]
+        self._resampler_buffer = self._resampler_buffer[needed:]
 
-        # adapt = n_available > 20
-        # if adapt:
-        #     pcm_extra, overflowed = self._stream.read(3)
-        #     logging.info('consuming extra samples, available was %d', n_available)
-        #     if overflowed:
-        #         logging.warning("SoundDeviceAudioInput: overflowed")
-            
-        #     out = bytes(pcm_buffer) +  bytes(pcm_extra)
-        # else:
-        out = bytes(pcm_buffer)     
+        logging.debug("PyALSA: resampler_buffer remaining=%d", len(self._resampler_buffer))
+        return np.clip(output * 32767.0, -32768, 32767).astype(np.int16).tobytes()
 
-        self.max_avail = max(self.max_avail, n_available)
+  
 
-        #Diagnostics 
-        #with open(self.logfile_name, "a", encoding="utf-8") as f:
-        #    f.write(f"{n_available}, {adapt}, {round(self._runavg, 2)}, {overflowed}\n")
+    def _close(self) -> None:
+        if self._pcm:
+            self._pcm.close()
+            self._pcm = None
 
-        if self.counter % 500 == 0:
-            logging.info(
-                "read available=%d, max=%d, latency:%d", 
-                n_available, self.max_avail, self._stream.latency
-            )
-            self.max_avail = 0
-
-        self.counter += 1
-        return out
-
-audio_io.SoundDeviceAudioInput = ModSoundDeviceAudioInput
+audio_io.SoundDeviceAudioInput = PyAlsaAudioInput
 
 # modified from bumble
 class ModWaveAudioInput(audio_io.ThreadedAudioInput):
@@ -538,7 +582,7 @@ async def init_broadcast(
 
     def on_flow():
         data_packet_queue = iso_queue.data_packet_queue
-        print(
+        logging.info(
             f'\rPACKETS: pending={data_packet_queue.pending}, '
             f'queued={data_packet_queue.queued}, '
             f'completed={data_packet_queue.completed}',

src/auracast/multicast_control.py

@@ -143,6 +143,10 @@ async def main():
         level=os.environ.get('LOG_LEVEL', logging.DEBUG),
         format='%(module)s.py:%(lineno)d %(levelname)s: %(message)s'
     )
+    
+    # Enable debug logging for bumble
+    # logging.getLogger('bumble').setLevel(logging.DEBUG)
+    
     os.chdir(os.path.dirname(__file__))
 
     global_conf = auracast_config.AuracastGlobalConfig(

src/auracast/server/multicast_server.py

@@ -26,6 +26,10 @@ from auracast.utils.sounddevice_utils import (
 )
 
 load_dotenv()
+
+# Configure bumble debug logging
+# log.getLogger('bumble').setLevel(log.DEBUG)
+
 # make sure pipewire sets latency
 # Primary and secondary persisted settings files
 STREAM_SETTINGS_FILE1 = os.path.join(os.path.dirname(__file__), 'stream_settings.json')