MFC peak based segmentation for continuous Arabic audio signal

Abstract

This paper presents an algorithm for segmenting a subset of emphatic and non-emphatic sounds automatically from continuously spoken Arabic speech. The important contribution of this paper is to generate rules for automatic segmentation of these sounds which can be extended to the rest of Arabic sounds. In addition, the findings can be used for other speech analysis problems such as data training for speech recognizers, continuous speech segmentation systems, or to build and label Arabic databases. This study has been done in context of recited principles of the Holy Quran which commonly known as recitation rules. The method developed is based on peaks detection from delta function of Mel Frequency Cepstral Coefficients “MFCC”. The peaks position is used for boundaries locating of the target sounds within the speech signal. Medium vocabulary speech database was used to evaluate the system performance. The test database contained 80 recited words. Each word was recorded from six different speakers constituting total of 480 words. A significant increase in accuracy rate has been achieved compared to the prior work of [1, 2] by enhancing the developed algorithm. Results show that the enhanced algorithm achieved a segmentation accuracy of up to 90%.