‘Listening’ To Dyslexic Children’s Reading: The Transcription And Segmentation Accuracy For ASR

Abstract

Dyslexic children read with a lot of highly phonetically similar error that is a challenge for speech recognition (ASR). Listening to the highly phonetically similar errors are indeed difficult even for a human. To enable a computer to ‘listen’ to dyslexic children’s reading is even more challenging as we have to ‘teach’ the computers to recognize the readings as well as to adapt to the highly phonetically similar errors they make when reading. This is even more difficult when segmenting and labelling the read speech for processing prior to training an ASR. Hence, this paper presents and discusses the effects of highly phonetically similar errors on automatic transcription and segmentation accuracy and how it is somehow influenced by the spoken pronunciations. A number of 585 files of dyslexic children’s reading is used for manual transcription, force alignment, and training. The recognition of ASR engine using automatic transcription and phonetic labelling obtained an optimum result, which is with 23.9% WER and 18.1% FAR. The results are almost similar with ASR engine using manual transcription 23.7% WER and 17.9% FAR.