The Development of an Efficient Transcription System for Kuwaiti Broadcast News and Conversational Speech.

Abstract

The research aims to introduce the first efficient speech transcription system for Kuwaiti Arabic (KA) using speech recognition technologies. The system replaces the conventional manual transcription scheme, which improves reliability, achieves the best use of time, and streamlines the process simultaneously. The research also presents two practical solutions for two fundamental challenges facing the development of speech recognition systems for Arabic dialects.  The first challenge is the shortage of dialectal data that is required for efficient modeling. This challenge is addressed by using a combination of available Modern Standard Arabic (MSA) data and the dialectal data when building the acoustic and language models. The second challenge is related to the linguistics specifications of the targeted dialect, which can be seen in the absence of a well-defined orthography system and consequently, the lack of pronunciation dictionaries. The research uses an extended version of the MADAMIRA morphological analyzer that covers KA to automatically generate the pronunciation dictionary needed to build the model. It also uses data from the GALE (phase 3), which contains approximately 22 hours of Kuwaiti speech, varying among broadcast news, talk shows, and conversational programs, as well as 29 hours of TV shows obtained from Kuwaiti TV channels. For MSA data, the researchers retrieved approximately 17.1 hours of speech produced by Gulf speakers from the GALE (phase 3) database. The best performing model reported in this research achieves 7.9% of the word error rate (WER), which is anticipated to deliver a good performance when used in varied applications.