Improving ASR performance using context‐dependent phoneme models

Abstract

PurposeThe purpose of this paper is to present evidence of the need to have a carefully designed lexical model for speech recognition for dyslexic children reading in Bahasa Melayu (BM).Design/methodology/approachData collection is performed to obtain the most frequent reading error patterns and the reading recordings. Design and development of the lexical model considers the errors for better recognition accuracy.FindingsIt is found that the recognition accuracy is increased to 75 percent when using context‐dependent (CD) phoneme model and phoneme refinement rule. Comparison between context‐independent phoneme models and CD phoneme model is also presented.Research limitations/implicationsThe most frequent errors recognized and obtained from data collection and analysis illustrate and support that phonological deficit is the major factor for reading disabilities in dyslexics.Practical implicationsThis paper provides the first step towards materializing an automated speech recognition (ASR)‐based application to support reading for BM, which is the first language in Malaysia.Originality/valueThe paper contributes to the knowledge of the most frequent error patterns for dyslexic children's reading in BM and to the knowledge that a CD phoneme model together with the phoneme refinement rule can built up a more fine‐tuned lexical model for an ASR specifically for dyslexic children's reading isolated words in BM.