julius-speech/julius

Implements a 2-pass tree-trellis search decoder supporting word N-grams and context-dependent HMMs with extensive pruning techniques, while remaining modular enough to work with standard HTK and SRILM formats. Supports multiple recognition modes (dictation, grammar-based, isolated word) in a single thread, with recent versions enabling DNN-based acoustic modeling via separate front-end modules. Runs on resource-constrained devices (<32MB memory) through to servers, with server mode, confidence scoring, word graphs, and forced alignment capabilities.

1,930 stars. No commits in the last 6 months.