DYSLEXIC CHILDREN’S READING: BUILDING A SPEECH RECOGNITION ENGINE USING AUTOMATIC TRANSCRIPTION AND LABELING

Dyslexic children, due to the deficit in phonological processing, read with a lot of highly phonetically similar errors. Highly phonetically similar errors refer to reading mistakes made at phoneme level when they are reading words, e.g. reading 'deb' for 'bed' (in English) or reading 'apa' for 'bapa' (Malay). Although this study is focusing on reading in Malay but regardless of the language, these errors impose challenges for automatic reading tutor to be accurate when 'listening' to their reading. Hence affecting the ability of the tutor to give useful feedback for correction. For the tutor to perform well, speech recognition engine must be trained to ‘listen’ to the highly phonetically similar errors. However, the challenge is not only to speech recognition but also to automatic transcription and phonetic labelling as the basic component to build a speech recognition engine for such automatic reading tutor. The highly phonetically similar errors affected the accuracy of speech transcription and phonetic labelling as well as speech recognition. Hence, this paper presents a speech recognition engine built using manual and automatic transcription and labelling. Although it is known that the manual transcription and labelling are the most accurate, we present both performances of speech recognition engines using both manual and automatic transcription and labelling. The accuracy of both engines are compared to see if the automatic version is fit to be used for building a larger speech recognition engine. The results were quite a surprise when the engine trained on automatic transcripts and labels produce 76.04% with only 0.23% difference with the manual one.