Improving statistical speech recognition

A summary of the theory of the hybrid connectionist HMM (hidden Markov model) continuous speech recognition system is presented. Experimental results indicating that the connectionist methods can significantly improve the performance of a context-independent HMM system to a performance close to that of the state of the art context-dependent system of much higher complexity are given. Experimental results demonstrating that a state of the art context-dependent HMM system can be significantly improved by interpolating context-independent connectionist probability estimates are reported. The development of a principled network decomposition method that allows the efficient and parsimonious modeling of context-dependent phones with no independence assumptions, is reported.<<ETX>>