[FEAT]: Add support for stream.read() to record silence from WASAPI speakers

[BungaaFACE]

What problem are you facing?

• audio isn`t recorded
• audio is recorded with artifacts
• problem with "silence"
• other

What is the cause of the error (in your opinion)?

• PyAudio\PortAudio bug
• I just need help(or answer)

I have created stream from speakers and trying to read() it. But when nothing plays on system sound then code stucks on stream.read(1024) until i turn on music or another sound source.

It would be great if you can add parameter fill_silense=True or something like that, so code won't be blocked on stream.read()

Here is code:

import numpy as np
import pyaudiowpatch as pyaudio
from faster_whisper import WhisperModel

def get_stream(p: pyaudio.PyAudio, device='micro'):
    if device == 'micro':
        stream = p.open(format=pyaudio.paInt16, channels=1, rate=16000, input=True, output=True, frames_per_buffer=1024)
        pyaudio.Stream()
    else:
        wasapi_info = p.get_host_api_info_by_type(pyaudio.paWASAPI)
        default_speakers = p.get_device_info_by_index(wasapi_info["defaultOutputDevice"])

        if not default_speakers["isLoopbackDevice"]:
            for loopback in p.get_loopback_device_info_generator():
                if default_speakers["name"] in loopback["name"]:
                    default_speakers = loopback
                    break
            else:
                print("Default loopback output device not found.\n\nRun `python -m pyaudiowpatch` to check available devices.\nExiting...\n")
                return

        print(f"Recording from: ({default_speakers['index']}){default_speakers['name']}")
        stream = p.open(
            format=pyaudio.paInt16,
            channels=default_speakers["maxInputChannels"],
            rate=int(default_speakers["defaultSampleRate"]),
            frames_per_buffer=1024,
            input=True,
            input_device_index=default_speakers["index"]
        )
    return stream

def transcribe_chuck(p, stream, model: WhisperModel, chunck_length=4):
    frames = []
    for _ in range(0, int(stream._rate / stream._frames_per_buffer * chunck_length)):
        data = stream.read(stream._frames_per_buffer)
        frames.append(data)

    audio_data = b''.join(frames)
    np_audio = np.frombuffer(audio_data, dtype=np.int16).astype(np.float32) / 32768.0

    segments, info = model.transcribe(np_audio, beam_size=7)
    transcription = ' '.join(segment.text for segment in segments)
    return transcription

def main():
    model_size = "large"
    model = WhisperModel(model_size, compute_type='float16')  # device="cuda",

    p = pyaudio.PyAudio()
    stream = get_stream(p, device='micro1')

    accumulated_transcription = ''

    try:
        while True:
            transcription = transcribe_chuck(p, stream, model)
            print(transcription)
            accumulated_transcription += transcription + ' '
    except KeyboardInterrupt:
        print('Stopping...')
        with open('log.txt', 'w') as log_file:
            log_file.write(accumulated_transcription)
    finally:
        stream.stop_stream()
        stream.close()
        p.terminate()

if __name__ == "__main__":
    main()

[s0d3s]

Hi! This is not a bug. Actually, it is a limitation of the PyAudio/PortAudio "synchronous" read (stream.read) implementation (under the hood it’s the same as the "callback" read). The skipping of silence and the anticipation of the actual sound are done deliberately.

Solution for your case: Use the callback read approach with queue.Queue (or another queue implementation). This should allow you to record even silence.

Your suggestion might be a good candidate for a new feature, so I will keep this issue open for some time.
If anyone else is interested in this, please mark this comment with a "🚀".

[BungaaFACE]

As workaround i created separate thread which translates silence to speakers. So stream is always got something to read()

The reason i need to read silence too because i need record original sound with correct timing, for example last 10 seconds.

For example system sound exists on 2nd and 4th second of 10 sec timeline, recorded result will be 2 seconds long with no pause between sounds

def stream_silence(p: pyaudio.PyAudio, speaker_stream, LISTEN_EVENT: threading.Event):
    '''
    This is a workaround for speaker_stream.read(1024) function.
    Whene there is no sound playing in system, python code stucks 
    on empty speaker stream reading. This func sending silence to speaker
    so stream is not empty
    '''
    stream = p.open(format=pyaudio.paInt16,
                    channels=1,
                    rate=speaker_stream._rate,
                    output=True)
    silence = b'\x00' * 1024 * SAMPLE_SIZE
    while LISTEN_EVENT.is_set():
        stream.write(silence)
        time.sleep(1)