Effect of characteristics of speakers on MSA ASR performance

Abstract

The paper deals with speaker-independent Automatic Speech Recognition (ASR) system for continuous speech. The ASR system is developed for Modern Standard Arabic (MSA) using Hidden Markov Models and a phonetically balanced corpus. The paper investigates the effect of two sources of speech variability: gender of speakers and the regional accent between the northern and southern regions of Algeria. The results show that the Word Error Rate (WER) of the ASR varies significantly between different localities according to the regional accent and gender of speakers. Indeed, higher rates are obtained in regions that present a specific pronunciation of some Arabic phonemes. As regard to gender of speakers, recordings of females are more recognized than those produced by male speakers in particular in southern localities.