Discriminative training for discrete HMM of a fixed-point DSP Mandarin digits recognition system

Abstract

Voice dialing technology is widely used in mobile phones, but most of them are based on speaker-dependent speech recognition, and cannot support dialing by saying the digits directly. Digit dialing requires speaker-independent speech recognition technology. The paper introduces a 16-bit fixed-point DSP based Mandarin digit recognition system based on discrete HMM, and discriminative training for the parameters. The parameters of the DHMM are reduced and only the truncated output probability matrices are used. A minimum classification error rate method is used to adjust the matrices discriminatively to improve the recognition accuracy further. In an experiment, a 14.4% error rate reduction is achieved.