Options
Rokiah Abdullah
Preferred name
Rokiah Abdullah
Official Name
Rokiah, Abdullah
Alternative Name
Abdullah, Rokiah
Abdullah, R.
Main Affiliation
Scopus Author ID
57208820919
Researcher ID
EJA-4882-2022
Now showing
1 - 1 of 1
-
PublicationDual tree complex wavelet packet transform based features for Malaysian speaker and accent recognition( 2021)An approach using energy and entropy based on Dual Tree Complex Wavelet Packet Transform (DT-CWPT) has been proposed for Malaysian speaker and accent recognition. The frequency transformation method using Fourier transform is not a very useful tool to analyse the signal as the time-information is lost. Time-frequency, using the wavelet approach is a good tool for the analysis of nonstationary signals both in time and frequency scale. In order to test the accuracy of the proposed method, speech signals are decomposed into 5 levels from energy and entropy derived from WPT and DT-CWPT. The performance is compared with conventional Mel Frequency Cepstral Coefficients (MFCC) and Linear Predictive Coding (LPC) features. Three different classifiers, such as k-Nearest Neighbors (k-NN), Support Vector Machine (SVM) and Extreme Learning Machine (ELM) are used to evaluate the performance of speaker and accent recognition. The new database was developed using English digits (0-9) and Malay words uttered by 75 undergraduate students of Universiti Malaysia Perlis (UniMAP), consisting the three main accents in Malaysia, which are Malay, Chinese and Indian. The experiments were carried out individually and by consolidating all the features. It was found through the experimental observations that the results employing the energy and entropy-based wavelet are promising and comparable. The best recognition rate achieved was 91.05%, which was computed from the energy and entropy-based wavelet (DT-CWPT) for speaker recognition using Malay words. For accent recognition, the best recognition rate at 94.84% was obtained from MFCC features using Malay words. For combined features, the speaker recognition using Malay words achieved 97.67%. While for accent recognition, the highest recognition for combined features obtained was 98.13%. Although the recognition result yielded a good result, it suffers from large features and long computation time. The feature selection, namely Binary Genetic Algorithm (GA) was applied to select the best subset from original features to reduce the number of features and long computation time. The results demonstrated that the number of features decreased by more than 70%. The computation time using k-NN, SVM, ELM classifier was reduced by at least by 69.5%, 53% and 14%, respectively. It was observed that GA reduces the computation time with a significant percentage while maintaining the recognition rate at a comparable figure with only a slight difference of within 3%. The accent recognition results show that there are significant differences between the three accents in Malaysia using the Malay language. It can be concluded that between the three accents in Malaysia, the native speakers (Malay accent) performed better followed by the Indian accent and Chinese accent. This is because as the native speakers, Malay speakers can pronounce the Malay words more precise, compared to Chinese and Indian. Overall, the results from Malay words show better performance compared to English digits for speaker and accent recognition.