Asian Research Thesis Index

Abstract

In this thesis, “a text-independent speaker verification system for Pashto speakers using accent and dialect recognition approach” has been designed. The purpose of the designed system is to recognize the region of origin of Pashto native speakers on the basis of their distinct dialects and to verify them using a speaker verification system. Due to the unavailability of the Pashto voice data in the form of different accents and dialects, a Pashto speakers’ database using different dialects of Pashto was developed. In order to develop the data initially, different dialectical variations of Pashto language were studied in detail and then the speech data was collected only from those different regions of Pakistan and Afghanistan where the Pashto is spoken with different dialects. After the database development, it is processed through front end and feature extraction processes where Mel Frequency Cepstral Coefficient (MFCC) features have been extracted from the collected data. After the MFCC feature vectors have been obtained, a Multilayer Perceptron (MLP) based classifier was designed to classify the speakers. Two separate classification experiments were performed (1) Speaker identification followed by dialect identification (2) Text-independent speaker verification followed by dialect identification. Speaker identification followed by dialect identification achieved 96.0 % identification accuracy, whereas, speaker verification followed by dialect identification achieved 100 % verification accuracy. Furthermore, the proposed Gaussian Mixture Model (GMM) based dialect identification system achieved 93.8 % identification accuracy in identifying Pashto native dialects. In order to inspect the noise robustness of the proposed system, the system’s performance was checked with the different degrees of noise level and Signal to Noise Ratio (SNR) was computed for each degree of noise. The performance of the system showed slightly degradation with the increase in the noise level, hence, showed its robustness against noise. A simple Pashto digits recognition (1 to 10 digits of Pashto) was also included in the study using MLP, HMM & SVM classifiers. Comparative analysis showed that the SVM based Pashto digit recognizer with 98.5 % recognition accuracy outperformed both the MLP and HMM based Pashto digit recognizers by showing 1.3 % and 3.3 % improvement in recognition accuracy. In order to benchmark the proposed research, the system’s performance was further tested on classifying some foreign accent of Pashto (Urdu accent of Pashto). In case of classifying the Urdu accent of Pashto, the system achieved 74.4 % recognition accuracy. Finally, the results achieved in the conducted experiments were compared with the recently proposed state of the art dialect identification, speaker verification and Pashto digit recognition systems. Comparative study showed that the proposed system outperformed some recently proposed dialect identification as well as speaker verification systems and showed relative improvement in recognition accuracies.

Add/Update Thesis