Kaldi

The industry-standard C++ toolkit for speech recognition, providing finite-state transducer based modeling and deep learning integration.

Kaldi is the definitive open-source framework for speech processing (ASR). Built on OpenFST, it offers a modular C++ codebase that supports linear algebra, acoustic modeling, and extensive feature extraction. Researchers use it to build robust systems like the LibriSpeech and Switchboard recipes, leveraging its flexible integration with CUDA for GPU-accelerated neural network training. It remains the primary engine for speech scientists requiring precise control over the decoding graph and lattice generation.

https://kaldi-asr.org/