parallel_example.ipynb

Kernel: Python 3 (ipykernel)

Install Dependencies

# !pip install -q torchaudio
SAMPLING_RATE = 16000
import torch
from pprint import pprint
import time
import shutil

torch.set_num_threads(1)
NUM_PROCESS=4 # set to the number of CPU cores in the machine
NUM_COPIES=8
# download wav files, make multiple copies
torch.hub.download_url_to_file('https://models.silero.ai/vad_models/en.wav', f"en_example0.wav")
for idx in range(NUM_COPIES-1):
    shutil.copy(f"en_example0.wav", f"en_example{idx+1}.wav")

Load VAD model from torch hub

model, utils = torch.hub.load(repo_or_dir='snakers4/silero-vad',
                                model='silero_vad',
                                force_reload=True,
                                onnx=False)

(get_speech_timestamps,
save_audio,
read_audio,
VADIterator,
collect_chunks) = utils

Define a vad process function

import multiprocessing

vad_models = dict()

def init_model(model):
  pid = multiprocessing.current_process().pid
  model, _ = torch.hub.load(repo_or_dir='snakers4/silero-vad',
                                model='silero_vad',
                                force_reload=False,
                                onnx=False)
  vad_models[pid] = model

def vad_process(audio_file: str):
    
    pid = multiprocessing.current_process().pid
    
    with torch.no_grad():
        wav = read_audio(audio_file, sampling_rate=SAMPLING_RATE)
        return get_speech_timestamps(
            wav,
            vad_models[pid],
            0.46,  # speech prob threshold
            16000,  # sample rate
            300,  # min speech duration in ms
            20,  # max speech duration in seconds
            600,  # min silence duration
            512,  # window size
            200,  # spech pad ms
        )

Parallelization

from concurrent.futures import ProcessPoolExecutor, as_completed

futures = []

with ProcessPoolExecutor(max_workers=NUM_PROCESS, initializer=init_model, initargs=(model,)) as ex:
    for i in range(NUM_COPIES):
        futures.append(ex.submit(vad_process, f"en_example{idx}.wav"))

for finished in as_completed(futures):
    pprint(finished.result())