Speech Segmentation to Phonemes Based on Hybrid Hidden Markov Models

Abstract

In this paper we develop automatic speech segmentation to phonemes using hybrid system based on Hidden Markov Model (HMM) and Artificial Neutral Network (ANN). It was shown that usage of ANN in order to estimate local probability in HMM leads to optimal global probability estimation in the general case, without imposition of additional model conditions. The result of automatic segmentation is close to the manual one, and can be successfully used in real applications for speech data segmentation and speech recognition system training.