multiprocessing bug

[jbaumgarten]

Sir, Madam,

I may need some help in order to use webrtcvad in multiprocess environment.
I'm working with pcm alaw audio which are converted in linear pcm for analysis purpose. Before processing the analysis, I'm using webrtcvad library in order to detect silences. Everything is ok if I'm doing one analysis at a time but as long I'm doing several silence detection at a time it fails.

The following code detect the silences:

import asyncio
import audioop
import concurrent.futures
import os
import webrtcvad

n = int(16000 * (30 / 1000.0) * 2)
vad = webrtcvad.Vad(2)

file1_name = "44022"
file2_name = "44024"

def process_chunk(data, file):
    while len(data) >= n:
        audio_30ms = data[0:n]
        data = data[n:]
        is_speech = vad.is_speech(audio_30ms, 16000)
        print(f"{file} - Is speech? {'true' if is_speech else 'false'}")


async def file1(pool):
    loop = asyncio.get_running_loop()
    f = open(f"{file1_name}-1.raw", "rb")
    data = f.read(960)
    while data:
        audioop.alaw2lin(data, 2)
        data = audioop.ratecv(data, 2, 1, 8000, 16000, None)[0]
        await loop.run_in_executor(pool, process_chunk, data, file1_name)
        data = f.read(960)


async def file2(pool):
    loop = asyncio.get_running_loop()
    f = open(f"{file2_name}-1.raw", "rb")
    data = f.read(960)
    while data:
        audioop.alaw2lin(data, 2)
        data = audioop.ratecv(data, 2, 1, 8000, 16000, None)[0]
        await loop.run_in_executor(pool, process_chunk, data, file2_name)
        data = f.read(960)


async def test():
    pool = concurrent.futures.ThreadPoolExecutor((os.cpu_count() or 1))
    await file1(pool)
    await file2(pool)


if __name__ == '__main__':
    asyncio.run(test())

The following doesn't detect the silences

import asyncio
import audioop
import concurrent.futures
import os
import webrtcvad

n = int(16000 * (30 / 1000.0) * 2)
vad = webrtcvad.Vad(2)

file1_name = "44022"
file2_name = "44024"

def process_chunk(data, file):
    while len(data) >= n:
        audio_30ms = data[0:n]
        data = data[n:]
        is_speech = vad.is_speech(audio_30ms, 16000)
        print(f"{file} - Is speech? {'true' if is_speech else 'false'}")


async def file1(pool):
    loop = asyncio.get_running_loop()
    f = open(f"{file1_name}-1.raw", "rb")
    data = f.read(960)
    while data:
        audioop.alaw2lin(data, 2)
        data = audioop.ratecv(data, 2, 1, 8000, 16000, None)[0]
        await loop.run_in_executor(pool, process_chunk, data, file1_name)
        data = f.read(960)


async def file2(pool):
    loop = asyncio.get_running_loop()
    f = open(f"{file2_name}-1.raw", "rb")
    data = f.read(960)
    while data:
        audioop.alaw2lin(data, 2)
        data = audioop.ratecv(data, 2, 1, 8000, 16000, None)[0]
        await loop.run_in_executor(pool, process_chunk, data, file2_name)
        data = f.read(960)


async def test():
    pool = concurrent.futures.ThreadPoolExecutor((os.cpu_count() or 1))
    await asyncio.gather(file1(pool), file2(pool))


if __name__ == '__main__':
    asyncio.run(test())

Did I miss something?

I also added some logs to display the buffer which is analized, and in both case, the buffer is the same

audio.zip

[jbaumgarten]

Another strange thing. The following source code

vad = webrtcvad.Vad(2)

audio_30ms = b'\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5\xd5'
is_speech = vad.is_speech(audio_30ms, 16000)
print(f"Is speech? {'true' if is_speech else 'false'}")
is_speech = vad.is_speech(audio_30ms, 16000)
print(f"Is speech? {'true' if is_speech else 'false'} --> {audio_30ms}")
is_speech = vad.is_speech(audio_30ms, 16000)
print(f"Is speech? {'true' if is_speech else 'false'} --> {audio_30ms}")
is_speech = vad.is_speech(audio_30ms, 16000)
print(f"Is speech? {'true' if is_speech else 'false'} --> {audio_30ms}")
is_speech = vad.is_speech(audio_30ms, 16000)
print(f"Is speech? {'true' if is_speech else 'false'} --> {audio_30ms}")
is_speech = vad.is_speech(audio_30ms, 16000)
print(f"Is speech? {'true' if is_speech else 'false'} --> {audio_30ms}")
is_speech = vad.is_speech(audio_30ms, 16000)
print(f"Is speech? {'true' if is_speech else 'false'} --> {audio_30ms}")

The output

Is speech? true
Is speech? true
Is speech? false
Is speech? false
Is speech? false
Is speech? false