Kaldi

Kaldi is a free, open-source toolkit for speech recognition research. It provides a comprehensive set of libraries, binaries, and scripts that enable researchers to build, train, and evaluate acoustic models for automatic speech recognition (ASR). The toolkit emphasizes performance and scalability and is widely used in academia and industry. It supports traditional Gaussian Mixture Model–hidden Markov model pipelines as well as modern neural-network–based systems, and it covers data preparation, feature extraction, alignment, training, decoding, and evaluation.

Kaldi originated around 2009 at Johns Hopkins University, with contributions from researchers including Daniel Povey and colleagues. Since then, it has grown through ongoing collaboration across universities and research groups and is maintained as an open-source project with a large ecosystem of example experiments and tutorials. The project emphasizes modularity, reproducibility, and extensibility, encouraging users to customize pipelines for different languages, datasets, and research goals.

Architecturally, Kaldi is primarily written in C++ and relies on scripting for experiment management. It uses the OpenFST library to implement finite-state transducer–based decoding and supports both traditional GMM-HMM representations and neural-network–based approaches via its nnet3 framework. Core workflows encompass data preparation, feature extraction (such as MFCC or PLP), alignment, various neural and non-neural training regimes, decoding with lattices, and lattice rescoring or neural rescoring.

Impact and usage: Kaldi is widely deployed in academic research and some industry projects due to its performance, flexibility, and extensive documentation. Its recipe-driven approach to experiments helps researchers reproduce results, compare methods, and experiment with different modeling choices within a common framework.