wav2letter++ is a fast, open source speech processing toolkit from the Speech team at Facebook AI Research built to facilitate research in end-to-end models for speech recognition. It is written entirely in C++ and uses the ArrayFire tensor library and the flashlight machine learning library for maximum efficiency. Our approach is detailed in this arXiv paper.
This repository also contains pre-trained models and implementations for various ASR results including:
- [NEW] Pratap et al. (2020): Scaling Online Speech Recognition Using ConvNets
- [NEW SOTA] Synnaeve et al. (2019): End-to-end ASR: from Supervised to Semi-Supervised Learning with Modern Architectures
- Kahn et al. (2019): Self-Training for End-to-End Speech Recognition
- Likhomanenko et al. (2019): Who Needs Words? Lexicon-free Speech Recognition
- Hannun et al. (2019): Sequence-to-Sequence Speech Recognition with Time-Depth Separable Convolutions
The previous iteration of wav2letter (written in Lua) can be found in the wav2letter-lua branch.
All details and documentation can be found on the wiki.
To get started with wav2letter++, checkout the tutorials section.
We also provide complete recipes for WSJ, Timit and Librispeech and they can be found in recipes folder.
Finally, we provide Python bindings for a subset of wav2letter++ (featurization, decoder, and ASG criterion) and a standalone inference framework for running online ASR.
If you use the code in your paper, then please cite it as:
@article{pratap2018w2l,
author = {Vineel Pratap, Awni Hannun, Qiantong Xu, Jeff Cai, Jacob Kahn, Gabriel Synnaeve, Vitaliy Liptchinsky, Ronan Collobert},
title = {wav2letter++: The Fastest Open-source Speech Recognition System},
journal = {CoRR},
volume = {abs/1812.07625},
year = {2018},
url = {https://arxiv.org/abs/1812.07625},
}
- Facebook page: https://www.facebook.com/groups/717232008481207/
- Google group: https://groups.google.com/forum/#!forum/wav2letter-users
- Contact: vineelkpratap@fb.com, awni@fb.com, qiantong@fb.com, jcai@fb.com, jacobkahn@fb.com, gab@fb.com, vitaliy888@fb.com, locronan@fb.com
See the CONTRIBUTING file for how to help out.
wav2letter++ is BSD-licensed, as found in the LICENSE file.