Byrne-ASR-English

Tiny ~12M character-level CTC speech recognizer with a lexicon + n-gram LM beam-search decoder, on CPU. ngram is a pure-Python 3-gram language model (Kneser-Ney, from a 1M-sentence news corpus) for better word choice. Best on clean wideband English; narrowband (8kHz) audio is out of domain.

Status: 3-gram ARPA LM active.