A study of speech recognition based on fuzzy statistics

This paper describes a fuzzy statistics-based approach to speech recognition. We call it fuzzy statistics-based hidden Markov models (FSHMMs), the parameters of which are calculated by fuzzy statistics instead of by the traditional Baum-Welch algorithm. The benefit of using FSHMMs lies in two aspects: the output distribution density can match the speech data more accurately, and the models can be adapted for new data conveniently while the loss of representation for old data is small. The corresponding training procedure and decision rule of FSHMMs are also described. We do not prove that the training algorithm converges, but experimental evidence suggests that it does. In recognition tests conducted on an isolated complete vocabulary of Mandarin words (speaker-dependent system), a 90.3% correct word recognition was observed for a male speaker and 16% for a female speaker using the male speaker models. After being adapted by a few female data, the speaker-dependent system could raise the recognition accuracy by 127% for a female and only reduce the accuracy by 4.7% for the male speaker