Work in Progress

Voice Activity Detection with a (small) CRDNN model trained on Libriparty

This repository provides all the necessary tools to perform voice activity detection with SpeechBrain using a model pretrained on Libriparty.

The pre-trained system can process short and long speech recordings and outputs the segments where speech activity is detected. The output of the system looks like this:

segment_001  0.00  2.57 NON_SPEECH
segment_002  2.57  8.20 SPEECH
segment_003  8.20  9.10 NON_SPEECH
segment_004  9.10  10.93 SPEECH
segment_005  10.93  12.00 NON_SPEECH
segment_006  12.00  14.40 SPEECH
segment_007  14.40  15.00 NON_SPEECH
segment_008  15.00  17.70 SPEECH

For a better experience, we encourage you to learn more about SpeechBrain.

Results

The model performance on the LibriParty test set is:

Release	hyperparams file	Test Precision	Test Recall	Test F-Score	Model link	GPUs
2021-09-09	train.yaml	0.9518	0.9437	0.9477	Model	1xV100 16GB

Pipeline description

This system is composed of a CRDNN that outputs posteriors probabilities with a value close to one for speech frames and close to zero for non-speech segments. A threshold is applied on top of the posteriors to detect candidate speech boundaries.

Depending on the active options, these boundaries can be post-processed (e.g, merging close segments, removing short segments, etc) to further improve the performance. See more details below.

Install SpeechBrain

First of all, please install SpeechBrain with the following command:

pip install speechbrain

Please notice that we encourage you to read our tutorials and learn more about SpeechBrain.

Perform Voice Activity Detection

from speechbrain.pretrained import VAD

VAD = VAD.from_hparams(source="speechbrain/vad-crdnn-libriparty", savedir="pretrained_models/vad-crdnn-libriparty")
boundaries = VAD.get_speech_segments("speechbrain/vad-crdnn-libriparty/example_vad.wav")

# Print the output
VAD.save_boundaries(boundaries)

The output is a tensor that contains the beginning/end second of each detected speech segment. You can save the boundaries on a file with:

VAD.save_boundaries(boundaries, save_path='VAD_file.txt')

Sometimes it is useful to jointly visualize the VAD output with the input signal itself. This is helpful to quickly figure out if the VAD is doing or not a good job.

To do it:

import torchaudio
upsampled_boundaries = VAD.upsample_boundaries(boundaries, 'pretrained_model_checkpoints/example_vad.wav')    
torchaudio.save('vad_final.wav', upsampled_boundaries.cpu(), 16000)

This creates a "VAD signal" with the same dimensionality as the original signal.

You can now open vad_final.wav and pretrained_model_checkpoints/example_vad.wav with software like audacity to visualize them jointly.

VAD pipeline details

The pipeline for detecting the speech segments is the following:

Compute posteriors probabilities at the frame level.
Apply a threshold on the posterior probability.
Derive candidate speech segments on top of that.
Apply energy VAD within each candidate segment (optional). This might break down long sentences into short one based on the energy content.
Merge segments that are too close.
Remove segments that are too short.
Double-check speech segments (optional). This could is a final check to make sure the detected segments are actually speech ones.

We designed the VAD such that you can have access to all of these steps (this might help to debug):


# 1- Let's compute frame-level posteriors first
audio_file = 'pretrained_model_checkpoints/example_vad.wav'
prob_chunks = VAD.get_speech_prob_file(audio_file)

# 2- Let's apply a threshold on top of the posteriors
prob_th = VAD.apply_threshold(prob_chunks).float()

# 3- Let's now derive the candidate speech segments
boundaries = VAD.get_boundaries(prob_th)

# 4- Apply energy VAD within each candidate speech segment (optional)

boundaries = VAD.energy_VAD(audio_file,boundaries)

# 5- Merge segments that are too close
boundaries = VAD.merge_short_segments(boundaries, close_th=0.250)

# 6- Re,ove segments that are too short
boundaries = VAD.remove_short_segments(boundaries, len_th=0.250)

# 7- Double-check speech segments (optional).
boundaries = VAD.double_check_speech_segments(boundaries, audio_file,  speech_th=0.5)

Inference on GPU

To perform inference on the GPU, add run_opts={"device":"cuda"} when calling the from_hparams method.

Training

The model was trained with SpeechBrain (ea17d22). To train it from scratch follows these steps:

Clone SpeechBrain:

git clone https://github.com/speechbrain/speechbrain/

Install it:

cd speechbrain
pip install -r requirements.txt
pip install -e .

Run Training: Training heavily relies on data augmentation. Make sure you have downloaded all the datasets needed:

 - LibriParty: https://drive.google.com/file/d/1--cAS5ePojMwNY5fewioXAv9YlYAWzIJ/view?usp=sharing
 - Musan: https://www.openslr.org/resources/17/musan.tar.gz
 - CommonLanguage: https://zenodo.org/record/5036977/files/CommonLanguage.tar.gz?download=1

cd recipes/LibriParty/VAD
python train.py hparams/train.yaml --data_folder=/path/to/LibriParty --musan_folder=/path/to/musan/ --commonlanguage_folder=/path/to/common_voice_kpd

Limitations

The SpeechBrain team does not provide any warranty on the performance achieved by this model when used on other datasets.

Citing SpeechBrain

Please, cite SpeechBrain if you use it for your research or business.

@misc{speechbrain,
  title={{SpeechBrain}: A General-Purpose Speech Toolkit},
  author={Mirco Ravanelli and Titouan Parcollet and Peter Plantinga and Aku Rouhe and Samuele Cornell and Loren Lugosch and Cem Subakan and Nauman Dawalatabad and Abdelwahab Heba and Jianyuan Zhong and Ju-Chieh Chou and Sung-Lin Yeh and Szu-Wei Fu and Chien-Feng Liao and Elena Rastorgueva and François Grondin and William Aris and Hwidong Na and Yan Gao and Renato De Mori and Yoshua Bengio},
  year={2021},
  eprint={2106.04624},
  archivePrefix={arXiv},
  primaryClass={eess.AS},
  note={arXiv:2106.04624}
}