Improvement on Automatic Speech Recognition Using Micro-genetic Algorithm

Abstract

In this paper we extend on previous work about the application of Genetic Algorithms (GAs) to optimize the transition structure of phoneme Hidden Markov Models (HMMs) for Automatic Speech Recognition (ASR). We focus on the development of a micro-GA where, in contrast to other GA approaches, each individual in the initial population consists of an element of the transition matrix of an HMM. Each individual's fitness is measured at the phoneme recognition level, which makes the execution of the algorithm faster. Evaluation of performance was performed with test speech data from the Wall Street Journal (WSJ) database. When measuring the performance of the optimized HMMs at the word recognition level, statistically significant improvements were obtained when compared with the performance of a standard speaker adaptation technique.