PyTorch xvector training

This repository contains xvector definitions and training recipes using PyTorch. Two training recipes are included, for narrowband (8kHz) and wideband (16kHz) data. The repository is organized into two main folders. The scripts folder contains recipes for training xvector extractors. The xvectors folder is the xvectors module, containing model architecture definitions, data loaders, and helper functions for performing inference (xvector extraction) with trained models.

Usage

We recommend creating an anaconda environment, cloning the repository, and installing xvectors as an importable module to utilize this repository effectively. The steps are summarized below:

>> conda create -n speech python=3.8
>> conda activate speech
>> git clone https://github.com/hltcoe/xvectors.git
>> cd xvectors
>> pip install -e .

Model Training

Our setup of xvector training requires that audio features be precomputed. We use filterbank features created using Kaldi, and require that features used for training be accessible through Kaldi file formats. More specifically, the files needed for training are:

1. feats.scp, feats.ark
2. utt2spk

Please refer to the Kaldi documentation on how to prepare data in this format.

Training the Narrowband Model

To train a narrowband model, run the train_nb.sh bash script located in the scripts directory. The data sources we utilized for narrowband xvector training include:

1. LDC2004S07
2. LDC2011S09
3. LDC2006S44 
4. LDC2017S06
5. LDC2011S08
6. LDC98S75
7. LDC2011S04
8. LDC2012S01
9. LDC2002S06
10. LDC2001S13
11. LDC2011S10
12. LDC2011S01
13. LDC99S79
14. LDC2011S05
15. VoxCeleb1
16. VoxCeleb2

All data is also augmented with the MUSAN corpus and telephony codecs.

Training the Wideband Model

To train a wideband model, run the train_wb.sh bash script located in the scripts directory. The data sources we utilized for wideband xvector training include:

1. VoxCeleb1
2. VoxCeleb2
3. LDC2013S03

All data is also augmented with the MUSAN corpus.

Citations

Snyder, David, et al. "X-vectors: Robust dnn embeddings for speaker recognition." 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018.

Contact

If you have any comments or questions, please create a GitHub issue.