An application of an auditory periphery model in speaker identification

  • Md. Atiqul Islam

Western Sydney University thesis: Doctoral thesis

Abstract

The number of applications of automatic Speaker Identification (SID) is growing due to the advanced technologies for secure access and authentication in services and devices. In 2016, in a study, the Cascade of Asymmetric Resonators with Fast Acting Compression (CAR FAC) cochlear model achieved the best performance among seven recent cochlear models to fit a set of human auditory physiological data. Motivated by the performance of the CAR-FAC, I apply this cochlear model in an SID task for the first time to produce a similar performance to a human auditory system. This thesis investigates the potential of the CAR-FAC model in an SID task. I investigate the capability of the CAR-FAC in text-dependent and text-independent SID tasks. This thesis also investigates contributions of different parameters, nonlinearities, and stages of the CAR-FAC that enhance SID accuracy. The performance of the CAR-FAC is compared with another recent cochlear model called the Auditory Nerve (AN) model. In addition, three FFT-based auditory features - Mel frequency Cepstral Coefficient (MFCC), Frequency Domain Linear Prediction (FDLP), and Gammatone Frequency Cepstral Coefficient (GFCC), are also included to compare their performance with cochlear features. This comparison allows me to investigate a better front-end for a noise-robust SID system. Three different statistical classifiers: a Gaussian Mixture Model with Universal Background Model (GMM-UBM), a Support Vector Machine (SVM), and an I-vector were used to evaluate the performance. These statistical classifiers allow me to investigate nonlinearities in the cochlear front-ends. The performance is evaluated under clean and noisy conditions for a wide range of noise levels. Techniques to improve the performance of a cochlear algorithm are also investigated in this thesis. It was found that the application of a cube root and DCT on cochlear output enhances the SID accuracy substantially.
Date of Award2021
Original languageEnglish

Keywords

  • automatic speech recognition
  • auditory perception

Cite this

'