Date of Graduation
Master of Science in Electrical Engineering (MSEE)
Second Committee Member
The proliferation of voice-activated devices and systems and over-the-phone bank transactions has made our daily affairs much easier in recent times. The ease that these systems offer also call for a need for them to be fail-safe against impersonators. Due to the sensitive information that might be shred on these systems, it is imperative that security be an utmost concern during the development stages. Vital systems like these should incorporate a functionality of discriminating between the actual speaker and impersonators. That functionality is the focus of this thesis.
Several methods have been proposed to be used to achieve this system and some success has been recorded so far. However, due to the vital role this system has to play in securing critical information, efforts have been continually made to reduce the probability of error in the systems. Therefore, statistical learning methods or techniques are utilized in this thesis because they have proven to have high accuracy and efficiency in various other applications. The statistical methods used are Gaussian Mixture Models and Support Vector Machines. These methods have become the de facto techniques for designing speaker identification systems. The effectiveness of the support vector machine is dependent on the type of kernel used. Several kernels have been proposed for achieving better results and we also introduce a kernel in this thesis which will serve as an alternative to the already defined ones. Other factors including the number of components used in modeling the Gaussian Mixture Model (GMM) affect the performance of the system and these factors are used in this thesis and exciting results were obtained.
Ojutiku, Alli Ayoola, "Text-Independent Speaker Identification using Statistical Learning" (2015). Theses and Dissertations. 1248.