Spectral and prosodic feature extractions for classical Arabic accents recognition among Malay speakers / Noor Jamaliah Ibrahim

Noor Jamaliah , Ibrahim (2021) Spectral and prosodic feature extractions for classical Arabic accents recognition among Malay speakers / Noor Jamaliah Ibrahim. PhD thesis, Universiti Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (226Kb)
	PDF (Thesis PhD) Download (3307Kb)

Abstract

The variability of speech patterns produced by individuals is unique. The uniqueness is due to the accent being influenced by individual's native dialect. Understanding and modeling the individual variation in spoken language are considered a fundamental challenge in the field of Automatic Speech Recognition (ASR) and accent identification. Differences in individual accent has been one of the critical issues in Classical Arabic (CA) recitation, particularly among the Malay speakers. This issue is often caused by the misarticulated of phonemes, which is influenced by the different style used in mother language and the Malay colloquial dialect. Most ASR researchers are unable to understand the nature and features of phonemes, as well as the speech patterns in CA, which result in reduced performance of ASR system. This research focused on identifying the accent used by Malay speakers in reciting of Surah Al-Fatihah against the seven types of Quranic accents (Qira'at), using the proposed feature extraction and classification technique. The experiment started with a thorough evaluation on phonetic properties of a set of letters and phonemes, which pronunciation are almost similar in different accents. Prior knowledge about the accents provides valuable information for speaker profiling, which can be incorporated into the decision parameter and technique to improve the system performance and efficiency. A combination of spectral and prosodic features in ASR system is proposed in this study, which is primarily designed for accent recognition. A distinguished variation in features of each phoneme or syllable can help in identifying and distinguishing one accent from another. Initial acoustic cues of the confused phonemes are selected by studying the speech production. The proposed system differs than that of conventional method, which only incorporates the spectral features for feature extraction, as reported in most research on ASR system. The prosodic components in CA such as pitch, energy, and spectral-tilt need to be taken into consideration, thus, presenting significant variability in features of each phoneme, which helps in distinguishing one accent from another. On the other hand, the spectral representation of the Mel-Frequency Cepstral Coefficients (MFCC) was used in determining the decorrelating property of the cepstrum. Our aim was to make use of the Gaussian Mixture Models-Universal Background Model (GMM-UBM) for automated classification of speech and make comparison with the conventional Gaussian Mixture Model (GMM). Based on the results, the system performed best when the prosodic features are integrated with the spectral features of MFCC, via the GMM classification with an accuracy of 81.713% (test-set) and 89.697% (train-set); the results translates to an improvement by 7.303% (test-set) and 5.477% (train-set) as compared to that of MFCC alone. Meanwhile, in a comparison between the use of GMM-UBM and conventional GMM classifications, an accuracy of 86.148% (test-set) and 90.255% (train-set) was achieved using the former, which translates into an improvement by 4.435% (test-set) and 0.558% (train-set), as compared to the latter. Therefore, it can be concluded that, integrating the prosodic and spectral features of MFCC, and incorporated with GMM-UBM classification in ASR system, outperforms other algorithms tested in this research.

Actions (For repository staff only : Login required)