Keyword-based Discriminative Training of Acoustic Models1

In this paper, we investigate a new discriminative training technique which focuses on optimizing a keyword error rate, rather than the error rate on all words. We hypothesize that improvements in keyword error rate correlate with improvements in understanding error rates. Keyword-based discriminative training is accomplished by modifying a standard minimum classification error (MCE) training algorithm so that only segments of speech relevant to keyword errors are used in the acoustic model training. When both the standard and keyword-based techniques are used to adjust mixture weights, we find that keyword error rate reduction compared to baseline maximum likelihood (ML) trained models is nearly twice as large for the keyword-based approach. The overall word accuracy is also found to be improved for keyword-based training, and we run several experiments to investigate this phenomenon.