Diacritics Effect on Arabic Speech Recognition

Abstract

Arabic is the native language for over 300 million speakers and one of the official languages in United Nations. It has a unique set of diacritics that can alter a word’s meaning. Arabic automatic speech recognition (ASR) received little attention compared to other languages, and researches were oblivious to the diacritics in most cases. Omitting diacritics circumscribes the Arabic ASR system’s usability for several applications such as voice-enabled translation, text to speech, and speech-to-speech. In this paper, we study the effect of diacritics on Arabic ASR systems. Our approach is based on building and comparing diacritized and nondiacritized models for different corpus sizes. In particular, we build Arabic ASR models using state-of-the-art technologies for 1, 2, 5, 10, and 23 h. Each of those models was trained once with a diacritized corpus and another time with a nondiacritized version of the same corpus. KALDI toolkit and SRILM were used to build eight models for each corpus that are GMM-SI, GMM SAT, GMM MPE, GMM MMI, SGMM, SGMM-bMMI, DNN, DNN-MPE. Eighty different models were created using this experimental setup. Our results show that Word Error Rates (WERs) ranged from 4.68% to 42%. Adding diacritics increased WER by 0.59% to 3.29%. Although diacritics increased WERs, it is recommended to include diacritics for ASR systems when integrated with other systems such as voice-enabled translation. We believe that the benefit of the overall accuracy of the integrated system (e.g., translation) outweighs the WER increase for the Arabic ASR system.