HTK

HTK, short for Hidden Markov Model Toolkit, is a software toolkit used to build and evaluate hidden Markov model–based speech recognition systems. It provides a suite of programs and libraries for data preparation, model training, feature processing, and decoding, with an emphasis on acoustic models built from HMMs and Gaussian mixtures.

HTK originated at the Cambridge University Engineering Department in the 1980s and 1990s as part of research into statistical speech recognition. It was released publicly for academic use and quickly became a widely used reference implementation for HMM-based systems. Multiple major releases followed, accompanied by documentation, sample data, and a configurable workflow for training and evaluation.

Key components include HCopy for feature extraction, HCompV for initialization, HInit and HRest for re-estimation, HERest for Baum-Welch re-estimation, HVite for Viterbi decoding, HHEd for editing HMM definitions, HResults for evaluation, and HParse for grammars and dictionaries. HTK supports context-dependent triphone models, multiple feature streams such as MFCCs with delta and delta-delta coefficients, and various training configurations.

Licensing typically permits free use for academic purposes while restricting commercial use and requiring attribution in publications. HTK is distributed through the Cambridge University Engineering Department and is designed to run on Unix-like systems, requiring a C compiler and related libraries.

Despite the rise of neural-network–based approaches, HTK remains influential in education and research and is still used in legacy projects. It provides a foundational reference implementation that helped shape early statistical speech recognition methods.