Arabic Automatic Speech Recognition Enhancement

Abstract

In this paper, we propose three approaches for Arabic automatic speech recognition. For pronunciation modeling, we propose a pronunciation variant generation with decision tree. For acoustic modeling, we propose the Hybrid approach to adapt the native acoustic model using another native acoustic model. Regarding the language model, we improve the language model using processed text. The experimental results show that the proposed pronunciation model approach has reduction in WER around 1%. The acoustic modeling reduce the WER by 1.2% and the adapted language modeling show reduction in WER by 1.9%.