Applying Genetic Algorithm on Computer Simulated Connected Digit for Speech Recognition Optimization

This paper describes an alternative approach in solving connected digit problem in speech recognition. Instead of depending very much on the validity of the speech signal via segmentation of isolated digit; this work applies a genetic like approach to anticipate missing or unrecognized acoustic information prior to recognize the whole speech string being uttered. The test speech pattern in this work is computer simulated binary number represent four digit string of Malay connected digits. The simulation test pattern format represents the expected pattern extracted using robust classification techniques like Neural Network or Support Vector Machine. The test patterns deliberately consist of corrupted syllables and words to see the ability of the approach in anticipating the expected pattern. The corrupted test patterns were categorized in two (1) The patterns consist of partially wrong syllable and (2) The patterns consist of all wrong syllables. The “wrong” syllables can be either invalid syllables or valid syllables but in wrong position. The experiments conducted show a promising result. For category (1), the strings were recognized easily using language grammar model. However, for patterns in category (2), it is noted that the approach can determine very well the expected string if the “wrong” syllable is invalid compare to valid but in wrong position. Speech Signal of Continuous