jimbozhang/kaldi-gop

Computes GMM-HMM based Goodness of Pronunciation scores by comparing forced-aligned phone likelihoods against an unconstrained phone loop via Viterbi decoding, normalizing by utterance duration. Builds on Kaldi's acoustic modeling framework and provides a lightweight alternative to DNN-based approaches, though with lower expected accuracy. Includes example scripts demonstrating end-to-end GOP computation from audio input.

159 stars. No commits in the last 6 months.