Design of intelligent Qira’at identification algorithm

The speaker's native dialect, accents and the socioeconomic background are few factors that influence the speaking style. The mixing of Qira’at is considered forbidden in Islam, especially during salat prayer. The identification of threats that could influence the accuracy of voice recognition and influence decisions in recitation recognition performance of accents recognition. On the other hand, only few studies focus on research of the performance factor or accuracy in the reverberant environment and none yet focusing on the factors that would affect Qira’at speech signals and identification. The main objective of this thesis is to propose the identification process of Quranic recitation but oriented to the identification of various Qira’at with the emphasis on recognition without being affected by echo or noise during live recitation or in recordings. Sequential Windowing Parameterizing of Affine Projection Algorithm (SPAP) is proposed to improve windowing parameterizing during echo cancellation, while recognition accuracy factors are taken into account for further improvement. The process of the SPAP Algorithm is to extend parameters of the Affine Projection Block with two different selections of windowing length that affect the final accuracy on pattern classification. The usage of Feature Selection (FS) contributes to simplify and enhance the quality of the dataset used by selecting significant features. Qira’at audio files for Surah Ad- Dhuha are used in this study to re-sample an audio sample. Clean audio signals from AEC are used with proposed feature selection technique called as X-Ant Colony Optimization, that utilizes the concept of Ant Colony Optimization, and can enhance feature extraction. For the feature vectors that are collected from feature extraction (MFCC) and feature selection (X-ACO), the feature vectors are used as input for the classification phase. A combination of Principal Component Analysis (PPCA) and Gaussian Mixture Model (GMM) is proposedly in used for the classification phase as it is able to reduce any redundancy from the latent variables and carries only the most important information through dispersion of entropy. To evaluate the algorithm, 350 samples for 10 types of Qira’at recitation are in used, and for justifying the best pattern classification, few algorithms are tested in the early preliminary evaluation with K-Nearest Neighbour, GMM and PPCA. And in the final evaluation for PPCA, it achieved high accuracy with 95.15%, while WER and EER are around 7.63%. The current evaluation for SPAP tests another echo database of a Lecture Room that presents a reduction in the accuracy rate of around 92.1% while the WER and EER are around 7.53%. But both of the results are significant compared to achieve results for MFCC without SPAP feature selection technique that just acquired 89.1% in early preliminary test. It proves that the current proposed algorithm achieves better results in Echo Greathall comparable to Echo Lecture Room and finally, these results will be used as foundation for any upcoming related research that may improve the understanding of Qira’at among the Muslim.

Preview Text FK 2018 100 - IR.pdf Download (752kB) | Preview

Actions (login required)

View Item