wav2letter

Why we included this project

wav2letter is the automatic speech recognition toolkit that Facebook AI Research built for training and running end-to-end models that go straight from raw audio to text. The repo ships reproducible recipes tied to published papers, including streaming convolutional networks, lexicon-free decoding, and semi-supervised self-training, plus pre-trained models you can start from. The catch is that development has moved into the Flashlight project, so this repository is best treated as a reference for proven recipes and results rather than a codebase you build on going forward. For researchers and engineers who want to reproduce published ASR results or see how a production-grade end-to-end system is put together, that reference value is exactly what matters.