SPEAKER VERIFICATION SYSTEM IN VARIOUS EMOTIONS USING ATOM ALIGNED SPARSE REPRESENTATION

• Main Author: Kusuma, Andika
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39464
• Institution: Institut Teknologi Bandung

Automatic Speaker Recognition system is a system that determines speaker identity through sound waves. This system can facilitate various daily services such as bank transaction via telephone. Nowadays, IVector based Automatic Speaker Recognition system for Bahasa has not been able to handle the problem of emotional difference. However, in reality, speaker enrollment and recognition is often done in different emotional condition. This emotional difference frequently degrades the performance of existing systems. Therefore, this research focuses on constructing Automatic Speaker Recognition system for Bahasa that could handle different emotion problem by applying IVector modelling technique and Atom Aligned Sparse Representation (AASR) transformation technique. This research begins with collecting data in the form of sound recordings of several speakers at neutral and emotional condition. The emotion classes used in this study are angry, happiness, sadness, and contentment. Compared to the baseline system that was built using IVector method only, the AASR system shows an increase in performance, namely a decrease in Equal Error Rate (EER) of 3.79% in non-neutral emotion test data. In neutral emotion test data, the AASR system also experience a decrease in EER of 2.24%. Overall, the AASR system improves speaker recognition performance by reducing the EER by 3.46%.