Certain explorations on speech enhancement techniques for automatic speaker recognition in noisy environment

Abstract

This Thesis presents a detailed study on speech enhancement algorithm to provide robustness to the Automatic Speaker Recognizer (ASR) in real-life noisy conditions. The main objective of this work is to attenuate the noise component of a noisy speech in order to enhance the quality of the speech processing devices and make them more robust under noisy conditions using wavelet based algorithms and to carry out a comprehensive evaluation and comparison of their performances on speaker recognition task. In this Thesis two new single channel wavelet based speech enhancement methods and a noise robust automatic speaker recognition are developed and reported.. Firstly, a technique using Time Adaptive Discrete Wavelet Thresholding (TADWT) based on Bionic Wavelet Transform (BWT) is proposed. In this approach discrete BWT is used for speech enhancement task, the adaptive nature of the BWT is captured by introducing a time varying linear factor at each scale over time and modified soft thresholding function is used for denoising. This method also provides a good auditory representation (sufficient frequency resolution), good perceptual quality of speech and low computational load. In this thesis, basic spectral subtraction (SS), iterative Wiener filtering (IWF), Ephraim Malah filtering (EMF), Bionic wavelet based thresholding (BWT) techniques and Perceptual wavelet packet transform (PWPT) have been used as baseline methods for speech enhancement tests. Performance evaluation of proposed methods is made based on segmental signal to noise ratio (SSNR), signal to noise ratio (SNR), Itakura-Saito (IS) distance measure and minimum mean square error (MMSE) for the objective speech quality evaluation. The average recognition accuracy of the system is improved while incorporating SE methods as preprocessor while comparing with the recognition rate obtained for degraded speech.