Speech recognition using stochastic segment neural networks

The authors previously presented (1991) a stochastic explicit-segment modeling (SESM) approach to speech recognition. There were two major stochastic components: boundary and phonetic classifications. The authors extend their earlier framework and incorporate context-dependent techniques into the major stochastic components. The current implementation uses stochastic segment neural networks (SSNN) to deal with these two problems. The authors have experimented with SSNN on a task of recognizing 25 words (city names) recorded from actual customers over the telephone network. Comparisons show that context-dependent modeling can reduce the error rate quite substantially. Specifically, using context-dependent boundary and phonetic classifications, the authors achieved an error rate of 3.3% with no rejections or about 0.35% at a rejection rate of 15%.>