flashlight/wav2letter

Project overview

flashlight/wav2letter is an open source automatic speech recognition toolkit that provides research recipes and pre-trained models for advanced ASR architectures, backed by FAIR. Active development has been consolidated into the Flashlight repository, and the project is intended for researchers and engineers working with C++ build environments.

Project type

Model Development · Model Runtime · Audio & Speech

Deployment

Refer to project documentation

License

License pending

Best for

• Researchers and AI engineers seeking to reproduce advanced ASR research papers using provided recipes and pre-trained models.

Key capabilities

• Includes documented recipes to reproduce advanced speech recognition research papers.
• Provides pre-trained models alongside research reproduction recipes.
• Contains data preparation scripts for training and evaluation.
• Provides CMake files to compile the recipes locally.

Limitations and risks

• Active development of wav2letter has been consolidated into the Flashlight repository.
• Building recipes strictly requires the Flashlight 0.3 branch.
• Exact results reproduction must use Flashlight versions 0.3.2 or older.

Getting started

• Setup is considered difficult because it requires manual installation of dependencies like Flashlight and ArrayFire, and requires compiling C++ source code using CMake.

Evidence and sources

• GitHub project description: Facebook AI Research's Automatic Speech Recognition Toolkit
• README: Future wav2letter development will occur in Flashlight.
• GitHub repository metadata: cpp, deep-learning, end-to-end, speech-recognition, wav2letter
• README: This repository includes recipes to reproduce the following research papers as well as *pre-trained* models.
• README: Data preparation for training and evaluation can be found in [data](data) directory.

Reviewed at: 2026-07-16T15:27:16Z