Isolated word speech recognition using hidden Markov models

Abstract

This project represents a study on isolated word speech recognition, using hidden Markov models (HMMs). The theory of HMMs with respect to speech modelling is presented. The practical aspects of the implementation of a discrete observation HMM recogniser are described and performed. Furthermore, a new method of performing recognition utilising the so called, observation probability densities, is developed and applied with reasonable results. The topic of vector quantization is also given importance and a limited theoretical background is presented. The Linde-Buzo-Gray (LBG) algorithm is presented in detail. Furthermore, a new algorithm is developed to compensate for the uneven distribution of vectors among the codewords which result as a consequence of the LBG algorithm. The complete vector quantization and HMMs setup is developed into a real-time isolated word speech recognition system. Recognition tests for both male and female utterances are performed and the results presented. A study on the recognition performance of the system(s), using signals with artificially added noise is also performed