Data driven GPVAD

Repository for the work in TASLP 2021 Voice activity detection in the wild: A data-driven approach using teacher-student training.

Sample predictions against other methods

Noise robustness

Results

Our best model trained on the SRE (V3) dataset obtains the following results:

	Precision	Recall	F1	AUC	FER	Event-F1
aurora_clean	96.844	95.102	95.93	98.66	3.06	74.8
aurora_noisy	90.435	92.871	91.544	97.63	6.68	54.45
dcase18	89.202	88.362	88.717	95.2	10.82	57.85

Usage

We provide most of our pretrained models in this repository, including:

1. Both teachers (T_1, T_2)
2. Unbalanced audioset pretrained model
3. Voxceleb 2 pretrained model
4. Our best submission (SRE V3 trained)

To download and run evaluation just do:

git clone https://github.com/RicherMans/Datadriven-VAD
cd Datadriven-VAD
pip3 install -r requirements.txt
python3 forward.py -w example/example.wav

Running this will print:

|   index | event_label   |   onset |   offset | filename            |
|--------:|:--------------|--------:|---------:|:--------------------|
|       0 | Speech        |    0.28 |     0.94 | example/example.wav |
|       1 | Speech        |    1.04 |     2.22 | example/example.wav |

Predicting voice activity

We support single file and filelist-batching in our script. Obtaining VAD predictions is easy:

python3 forward.py -w example/example.wav

Or if one prefers to do that batch_wise, first prepare a filelist: find . -type f -name *.wav > wavlist.txt' And then just run:

python3 forward.py -l wavlist

Extra parameters

• -model adjusts the pretrained model. Can be one of t1,t2,v2,a2,a2_v2,sre. Refer to the paper for each respective model. By default we use sre.
• -soft instead of predicting human-readable timestamps, the model is now outputting the raw probabilities.
• -hard instead of predicting human-readable timestamps, the model is now outputting the post-processed 0-1 flags indicating speech. Please note this is different from the paper, which thresholded the soft probabilities without post-processing.
• -th adjusts the threshold. If a single threshold is passed (e.g., -th 0.5), we utilize simple binearization. Otherwise use the default double threshold with -th 0.5 0.1.
• -o outputs the results into a new folder.

Training from scratch

If you intend to rerun our work, prepare some data and extract log-Mel spectrogram features. Say, you have downloaded the balanced subset of AudioSet and stored all files in a folder data/balanced/. Then:

cd data;
mkdir hdf5 csv_labels;
find balanced -type f > wavs.txt;
python3 extract_features.py wavs.txt -o hdf5/balanced.h5
h5ls -r hdf5/balanced.h5 | awk -F[/' '] 'BEGIN{print "filename","hdf5path"}NR>1{print $2,"hdf5/balanced.h5"}'> csv_labels/balanced.csv

The input for our label prediction script is a csv file with exactly two columns, filename and hdf5path.

An example csv_labels/balanced.csv would be:

filename hdf5path
--PJHxphWEs_30.000.wav hdf5/balanced.h5                                                                                          
--ZhevVpy1s_50.000.wav hdf5/balanced.h5                                                                                          
--aE2O5G5WE_0.000.wav hdf5/balanced.h5                                                                                           
--aO5cdqSAg_30.000.wav hdf5/balanced.h5                                                                                          

After feature extraction, proceed to predict labels:

mkdir -p softlabels/{hdf5,csv};
python3 prepare_labels.py --pre ../pretrained_models/teacher1/model.pth csv_labels/balanced.csv softlabels/hdf5/balanced.h5 softlabels/csv/balanced.csv

Lastly, just train:

cd ../; #Go to project root
# Change config accoringly with input data
python3 run.py train configs/example.yaml

Citation

If youre using this work, please cite it in your publications.

@article{Dinkel2021,
author = {Dinkel, Heinrich and Wang, Shuai and Xu, Xuenan and Wu, Mengyue and Yu, Kai},
doi = {10.1109/TASLP.2021.3073596},
issn = {2329-9290},
journal = {IEEE/ACM Transactions on Audio, Speech, and Language Processing},
pages = {1542--1555},
title = {{Voice Activity Detection in the Wild: A Data-Driven Approach Using Teacher-Student Training}},
url = {https://ieeexplore.ieee.org/document/9405474/},
volume = {29},
year = {2021}
}

and

@inproceedings{Dinkel2020,
  author={Heinrich Dinkel and Yefei Chen and Mengyue Wu and Kai Yu},
  title={{Voice Activity Detection in the Wild via Weakly Supervised Sound Event Detection}},
  year=2020,
  booktitle={Proc. Interspeech 2020},
  pages={3665--3669},
  doi={10.21437/Interspeech.2020-0995},
  url={http://dx.doi.org/10.21437/Interspeech.2020-0995}
}