ROBUST FEATURE EXTRACTION BASED ON SPECTRAL AND PROSODIC FEATURES FOR CLASSICAL ARABIC ACCENTS RECOGNITION

Abstract

The variability of speech patterns produced by individuals is unique. The uniqueness is due to the accent influenced by the individual’s native dialect. Modeling individual variation of spoken language is a challenge under the Automatic Speech Recognition (ASR) field. The individual differences concerning of accent revealed the critical issues in Classical Arabic (CA) recitation among Malay speakers. This problem is caused by the misarticulate phonemes, which affected by the Malay colloquial dialect and native language. Most of ASR researchers are unable to understand the behavior of phonemes and speech patterns in CA, thus degrading the ASR performance. This paper focuses on identifying the accent of Malay speakers on the recitation of Sūrah Al-Fātiḥah with 7 Quranic accents, using the proposed feature extraction technique. In this work, the technique presented is a combination of spectral and prosodic features, which are mainly designed for accent in ASR. Differed with current conventional method, where the spectral feature alone has been applied for feature extraction in many ASR research. The prosodic elements in CA such as pitch, energy and spectral-tilt need to be taken into consideration, thus a significant variety of features for each phoneme able to help in distinguishing one accent from another. Meanwhile, the spectral representation of Mel-Frequency Cepstral Coefficients (MFCC) is utilized for the decorrelating property of the cepstrum. At present, Gaussian Mixture Models (GMM) has been applied for the classification stage. From experimental results, the system performance is the best when the prosodic is integrated with MFCC, alongside the GMM with 81.7%-89.6% of accuracy. It was 5.5%-7.3% increment as compared to MFCC alone.