Silero VAD

The Common Voice dataset cannot be downloaded via static links. You must visit the Mozilla Common Voice website and download the archives manually.

Note: The provided annotations are compatible with Common Voice Corpus 16.1.

• audio_path: Contains the unique filenames of the .mp3 files obtained after downloading the dataset.

The Silero-VAD datasets are provided as .feather files containing speech annotations. You can load these files using the pandas library.

All data is annotated with a temporal discretization of approximately 30 milliseconds (num_samples = 512).

import pandas as pd
dataframe = pd.read_feather(PATH_TO_FEATHER_FILE)

Install the Silero VAD package using pip to access the high-level Python API.

pip install silero-vad

You can find the optimal input and output thresholds for your fine-tuned model using the search_thresholds script. This script uses the settings defined in your config.yml and applies the specified model to the validation dataset to find the best thresholds.

python search_thresholds

To tune the Silero-VAD model on custom data, ensure the following Python dependencies are installed:

• torchaudio>=0.12.0
• omegaconf>=2.3.0
• sklearn>=1.2.0
• torch>=1.12.0
• pandas>=2.2.2
• tqdm

If you set use_torchhub: False in your configuration, you must also install the silero-vad library via pip.

pip install silero-vad

Use the silero-vad package to load a model, read an audio file, and extract speech timestamps. You can use the return_seconds=True argument to get timestamps in seconds instead of the default sample count.

from silero_vad import load_silero_vad, read_audio, get_speech_timestamps

model = load_silero_vad()
wav = read_audio('path_to_audio_file')
speech_timestamps = get_speech_timestamps(
  wav,
  model,
  return_seconds=True,  # Return speech timestamps in seconds (default is samples)
)

Load Silero VAD directly from GitHub using PyTorch's hub mechanism. This is useful if you want to avoid installing the package via pip and prefer managing models through Torch Hub.

import torch
torch.set_num_threads(1)

model, utils = torch.hub.load(repo_or_dir='snakers4/silero-vad', model='silero_vad')
(get_speech_timestamps, _, read_audio, _, _) = utils

Silero VAD uses torchaudio for audio I/O (torchaudio.info, torchaudio.load, and torchaudio.save). You must ensure a proper audio backend is installed on your system. Choose one of the following options:

• FFmpeg: conda install -c conda-forge 'ffmpeg<7'
• sox_io: apt-get install sox (Tested on libsox 14.4.2)
• soundfile: pip install soundfile

To start the fine-tuning process using your config.yml settings, run the tune.py script. The process runs for the specified num_epochs, and the best checkpoint (based on ROC-AUC on the validation set) will be saved to the path specified in model_save_path in .jit format.

python tune.py

Training and validation datasets must be prepared as .feather files. Each dataframe must contain the following mandatory columns:

• audio_path: Absolute path to the audio file. Files should be PCM data, preferably in .wav or .opus formats. It is recommended to resample audio to 16000 Hz before training.
• speech_ts: A list of dictionaries representing speech segments: [{'start': START_SEC, 'end': END_SEC}, ...], where times are in seconds. For high-quality tuning, use a precision of up to 30 milliseconds.

Note: Audio files longer than max_train_length_sec will be truncated. For better efficiency, it is recommended to pre-cut long audio into segments of length max_train_length_sec.