Foreign accent and automated speech recognition : error analysis of automated transcription of the English of Hebrew speakers

Abstract

This dissertation examines the possible effects a speaker’s foreign accent (namely, Hebrew) may have on the automatic transcription of their speech in English. To assess such a possible effect, automatically generated transcriptions (created by YouTube) of English speech uttered by Hebrew native speakers and English native speakers were evaluated. Comparison and examination of the transcription errors found in the transcriptions of both groups of speakers have given rise to the conclusion that the Hebrew accent does not seem to be a significant factor interfering with the accuracy of automatic transcriptions. Nonetheless, some transcription errors found in the automatic transcriptions of speech of Hebrew native speakers may be attributed to their accents (e.g., errors related to pronunciation, which are less frequent in the transcriptions of speech of native English speakers). Additionally, a language model was utilized in this dissertation with the aim to find a method to predict where transcription errors are likely to occur. Due to confinements to data in the form of written text, such predictions cannot be made presently in this study.