Investigation of Dynamic Time Warping and Neural Network for Arabic phonemes recognition based Malay speakers

Abstract

Speech recognition techniques for Arabic language are still in its infant stage and gain much attention recently. This is due to Arabic as the language of the Holy book of Muslims; hence it has attracted attention of native speakers and other Muslims who are non-native Arabic speakers as they need to use Arabic language while performing worships. Therefore, this research investigated Arabic phonemes recognition specifically for Malay speakers. The proposed methods are evaluated and examined utilizing a corpus which contains Arabic phoneme tokens with Mel Frequency Cepstral Coefficients (MFCC) as feature extraction. Next, recognition process is attained using Dynamic Time Warping (DTW) and Pattern Recognition Neural Network (PRNN) for verifying the similarity between the Arabic phonemes. In this study, three methods are used to evaluate the recognition stage. Firstly, DTW and PRNN are evaluated solely followed by combination of both. Results attained showed that the overall recognition rate of this method is 89.92% for DTW individually, 94% for PRNN solely whilst for fusion of DTW and PRNN the recognition rate attained is 98.28% and thus proven that fusion of DTW and PRNN can be utilised for recognition of Arabic phonemes.