Phonetically sensitive discriminants for improved speech recognition

A phonetically sensitive transformation of speech features has yielded significant improvement in speech-recognition performance. This (linear) transformation of the speech feature vector is designed to discriminate against out-of-class confusion data and is a function of phonetic state. Evaluation of the technique on the TI/NBS connected digit database demonstrates word (sentence) error rates of 0.5% (1.5%) for unknown-length strings and 0.2% (0.6%) for known-length strings. These error rates are two to three times lower than the best previously reported results and suggest that significant improvements in speech-recognition system performance can be achieved by better acoustic-phonetic modeling.>