Malay speech intelligibility test for deaf children: phoneme recognition using artificial neural network

Abstract

It is estimated that about 2000 deaf children are born each year in Malaysia. Most deaf Malaysian children have very poor speech intelligibility. Reduced intelligibility severely compromises communication and social interaction for affected individuals. Although speech deficiencies in the deaf are quite difficult to overcome, learning to produce intelligible speech is not an impossible task. Studies have shown that deaf children receiving Cued Speech can acquire reasonable speech intelligibility, surpassing the majority of signing children in verbal language skills. A reliable measure of speech intelligibility for deaf children is required for several reasons: to provide an index of the severity of speech disorder, to assist in treatment decisions, and to quantify changes which may result from intervention or treatment. This thesis investigates the approach to measure speech intelligibility of deaf Malaysian children. The research discussed in this work starts with the experiments on a practical Malay Speech Intelligibility Test (MSIT), suitable for use within deaf Malaysian children training programme. In this study, speech intelligibility of deaf children is measured through the ability to say simple nonsense syllables (consisting of a consonant and a vowel) for all 22 Malay consonants. The MSIT score should indicate how well these children can produce speech; the higher the score, the better their speech intelligibility. The next course of action was to investigate phoneme recognition system that will suit MSIT. Artificial neural network was utilized to effectively model the distribution of feature vectors present in speech signals for classification. A novel approach using speech spectrum image becomes the inputs to a three-layer MLP (Multi-layer Perceptron) neural network. The input feature sets for the intelligent phoneme identification were based on the intrinsic characteristics of Malay syllables shown in the captured speech signal spectrum image. The spectrum images were produced from widely used speech filter algorithm; Mel-frequency Cepstral Coefficients (MFCC), Perceptual Linear Prediction (PLP) and Relative Spectral Transform - Perceptual Linear Prediction (RASTA-PLP). The classifiers have been tested for twenty-two Malay phonemes utterances from two males and two females’ children speaker. The performance of the system for recognition of Malay phonemes is measured and compared with the performance of human listener. The successful development of the phoneme recognition system serves several purposes: (a) it will be one of the first methods employed to objectively measure speech intelligibility of deaf Malaysian children, and (b) it will contribute to better assessment and management of intervention programme for deaf Malaysian children.