The speech toolkit that got swallowed by Flashlight

What it does wav2letter++ was Facebook AI Research’s C++ automatic speech recognition toolkit. This repository now functions mainly as a historical archive: it still holds training recipes, data preparation scripts, and pre-trained models to reproduce a handful of 2019–2020 research papers, but the living codebase has been absorbed into the ASR application inside Flashlight.

The interesting bit The repo is deliberately pinned to the past; exact reproduction of its published results requires Flashlight 0.3.2 or older, so it acts more like a time-capsule for benchmarking specific papers than a framework you would extend today.

Key highlights

• Recipes and pre-trained models for five ASR papers, including streaming ConvNets, self-training, and lexicon-free recognition
• End-to-end C++ implementation from the pre-transformer era of speech recognition
• Locked to legacy Flashlight (≤ 0.3.2) for exact reproducibility
• MIT licensed
• Original Lua-based predecessor available on a separate branch

Caveats

• All future development happens in the Flashlight ASR application, not in this repository
• Exact paper reproduction requires outdated dependency versions
• This is effectively an archived snapshot rather than the current codebase

Verdict Use this only if you need to reproduce those specific papers or their exact pre-trained weights; for new work, go directly to the Flashlight ASR application.