Automatic transcription and segmentation accuracy of dyslexic children’s speech

Abstract

Highly phonetically similar reading mistakes often occur when dyslexic children read. In respect to automatic speech transcription, these mistakes are challenging, even for manual transcription. The highly phonetically similar reading mistakes are difficult to be recognized, not to mention segmenting and labelling them accordingly for processing prior to training speech recognition (ASR). The need to automate the segmentation and labelling arise especially when we need to build an ASR for assisting dyslexic children’s reading. Hence, the aim of this paper is to investigate the effects that highly phonetically similar errors have upon transcription and segmentation accuracy. A total of 585 speech files are used to produce manual transcription, forced alignment, and training. The recognition of ASR engine using automatic transcription and phonetic labelling obtained 76.04% accuracy with 23.9% word error rate and 18.1% false alarm rate. The results are almost similar with its manual counterpart with 76.26% accuracy, 23.7% word error rate and 17.9% false alarm rate.