Automatic transcription and phonetic labelling of dyslexic children's reading in Bahasa Melayu

Automatic speech recognition (ASR) is potentially helpful for children who suffer
from dyslexia. Highly phonetically similar errors of dyslexic children‟s reading affect the accuracy of ASR. Thus, this study aims to evaluate acceptable accuracy of ASR using automatic transcription and phonetic labelling of dyslexic children‟s reading in BM. For that, three objectives have been set: first to produce manual transcription and phonetic labelling; second to construct automatic transcription and phonetic labelling using forced alignment; and third to compare between accuracy using automatic transcription and phonetic labelling and manual transcription and
phonetic labelling. Therefore, to accomplish these goals methods have been used including manual speech labelling and segmentation, forced alignment, Hidden Markov Model (HMM) and Artificial Neural Network (ANN) for training, and for measure accuracy of ASR, Word Error Rate (WER) and False Alarm Rate (FAR) were used. A number of 585 speech files are used for manual transcription, forced alignment and training experiment. The recognition ASR engine using automatic transcription and phonetic labelling obtained optimum results is 76.04% with WER as low as 23.96% and FAR is 17.9%. These results are almost similar with ASR
engine using manual transcription namely 76.26%, WER as low as 23.97% and FAR a 17.9%. As conclusion, the accuracy of automatic transcription and phonetic labelling is acceptable to use it for help dyslexic children learning using ASR in Bahasa Melayu (BM)