Speaker recognition using orthogonal LPC parameters in noisy environment

Abstract

Last two decades witnessed a number of speaker recognition algorithms that yield reasonably good to excellent performance with high quality and relatively noise-free speech. However, in practice, noise is unavailable. The presence of noise is found to cause severe deterioration in the performance of these algorithms. In this thesis, the problem of recognizing speaker from noisy speech has been studied. Sambur’s algorithm that utilizes orthogonal linear predictive coding (OLPC) parameters is chosen for this study because of its simplicity and high recognition accuracy. The effect of noise on the performance of Sambur’s algorithm is studied. The algorithm is, then, modified by incorporating various parameter estimation techniques which are capable of yielding relatively better LPC parameters. These techniques may be divided into two groups. The first group contains three parameter estimation procedures, namely, the instrumental variable (IV) method, the autocorrelation subtraction (AS) method and estimation through the use of shifted Yule-Walker (SYW) equations. The techniques in the second group are based upon enhancing speech followed by the conventional LPC estimation technique applied to the enhanced speech. Three enhancement algorithms have been tested, namely, the adaptive noise cancellation (ANC) technique, the linear prediction smoothing (LPS) technique and the adaptive filtering technique (AFT). It was found from comparative study that speaker recognition based on speech enhanced by AFT yielded the best recognition accuracy among these techniques.