Automatic Speech Recognition Model for Dyslexic Children Reading in Bahasa Melayu

Dyslexic children suffer from dyslexia, a condition that profoundly impedes reading and spelling ability due to its phonological origin. Often, these children found reading and spelling difficult, exhaustive, and less interesting, and thus they are self-withdrawn from the learning process. When reading and spelling, they make many mistakes even for simple, common words that they themselves found embarrassing. However, this does not mean that they have lower IQ level than normal children. In fact, dyslexic children have average or high level of IQ and thus have a lot of potential when given the right help and support such as motivational support and suitable teaching techniques. With advancement in technology in education, computer-based applications are used to stimulate the learning process of reading and spelling. Hence, this study is an initiative towards proposing an automatic speech recognition (ASR) model to enable computer to 'listen' should incorrect reading occurs. The scope of this study focuses on modeling and recognizing single Bahasa Melayu (BM) words within the school syllabus for level one (tahap satu) dyslexic pupils of primary schools. To propose the ASR model, a reading and spelling model of dyslexic children reading in BM is first
proposed, which models reading at word recognition level. To propose such model, ethnographic techniques are employed namely informal interviews and observation, in order to obtain the reading and spelling error patterns of dyslexic children. A number of ten dyslexic children, aged between 7 to 14 years old whose reading level is similar, participated in the study. These children are recruited from two public schools that offer special dyslexia classes for the children. A total of 6112 utterances are recorded in audio form resulting in a total of 6051 errors of various types. Among these, the patterns that are most frequently made by these children are of 'Substitutes vowel', 'omits consonant', 'nasals', and 'substitutes consonant'. The ASR model is proposed taking into consideration the error patterns that make lexical model a fundamental element for speech recognition. The lexical model is modeled to treat mispronunciations as alternative pronunciations or variants of target words. To that, a phoneme refinement strategy is applied aiming to increase recognition accuracy. A prototype recognizer is developed based on the proposed model for further evaluation. The evaluation is performed to evaluate the recognizer's performance in terms of accuracy, measured in word error rate (WER) and miscue detection rate (MDR) that is closely related to false alarm rate (FAR). The recognizer scores a satisfying 25% of WER and a relatively high MDR of 80.77% with 16.67% FAR.