Automatic Detection of Shadda in Modern Standard Arabic Continuous Speech

Abstract

The presence of diacritics Shadda in Arabic continuous speech may lead to the reduction of the accuracy of automatic Word Boundary Detection (WBD), which caused one word will be wrongly detected as two words. Therefore, this will affect the accuracy of Automatic Speech Recognition (ASR), if it is based on WBD. Shadda is one of the essential characteristics of the Arabic language which represents a consonant doubling. In this paper, a proposed method of automatic detection of Shadda in Modern Standard Arabic (MSA) continuous speech was introduced to improve the accuracy of WBD in MSA continuous speech. The prosodic features namely Short Time Energy (STE), Fundamental Frequency and Intensity were investigated for its ability as Shadda pattern detection in continuous MSA speech. We have analyzed the proposed features by implementing a separated algorithm for each feature to detect Shadda pattern automatically. In addition, a new proposed method which is a combination of STE and Intensity were introduced. The dataset in this work is a collection of 1-hour TV broadcast news from Aljazeera Arabic TV channel for 2018 - broadcasters. We found that the Shadda pattern is very similar to unvoiced regions of speech, and this represents a big challenge for the improvement of WDB using Shadda. Results showed that the detection of Shadda using Short Time Energy and Intensity outperforms the Fundamental frequency with 55% of accuracy. Intensity achieved 71.5% in accuracy. In addition, a combination between Intensity & STE features was performed and achieved good results with 67.15% in accuracy. The number of false positive too has been reduced compared to Intensity alone.