Estimation of Noise Power Spectrum and Automatic Speaker Recognition System

Abstract: Unseen noise estimation is one of the challenging steps to make the speech enhancement algorithm work in adverse conditions. The prior knowledge known about the encountered noise is that it is different from the involved speech.The proposed work consists of two segments, Speech Enhancement and Automatic Speaker Recognition (ASR) system. The speech enhancement comprises of offline training and online enhancement processes. In offline training, the inputs clean speech data and noisy speech magnitude are collected and trained using Support Vector Machine (SVM). In online enhancement, the trained signals are compared and their noise spectrum is estimated using the Modified Spectral Subtraction (MSS) method which is also used for the removal of noises. Then the enhanced speech signal is obtained by transforming the estimated spectrum into time domain. The features are extracted from the obtained enhanced speech using Mel Frequency Cepstral Coefficients (MFCC) and Perceptual Linear Prediction (PLP). Finally the speaker recognition is done using k-NN and Gaussian Mixture Model (GMM) based Multi-SVM. The experimental results are compared and efficient system is obtained.

Keywords: Speech enhancement, SVM, MSS, k-NN, GMM and Multi-SVM.