kaldi

Why we included this project

Kaldi is the toolkit most speech researchers cut their teeth on, and it still holds up for production work that newer frameworks don't cover. The whole pipeline lives in C++: feature extraction, acoustic model training with both GMM and neural approaches, finite-state-transducer decoding, and speaker identification and verification, all exposed as command-line programs that script cleanly. What really sets it apart are the example recipes, which take you from raw audio corpora to a working recognizer step by step, making it a good place to learn how ASR actually works instead of treating it as a black box. Its Apache license and modular design have let many commercial systems build on top of it, and parts of it, like the decoders and matrix libraries, still show up inside newer toolkits. If you need to train custom acoustic models, adapt an existing one to a new language or domain, or do speaker recognition, Kaldi is a solid, well-documented base to start from.