The Study Of Speaker Recognition Based On Wavelet Transform And GMM

This paper describes in detail the basic knowledge of wavelet theory, studies the application of wavelet analysis to speech signal processing and presents two new speech feature parameters based on wavelet packet analysis, the two parameters are named SBC and WPP respectively. This paper also describes a frequency bands division of the speech signal by combining the wavelet transform and wavelet packet transform, this division method is similar to the critical bands of human being 's ear, from which we could get speech feature parameters SBC. Wavelet transform is a better decorrelator of subband energies than discrete cosine transform, from which we could get WPP. The experiment results indicate that the new feature parameter WPP is able to outperform SBC and SBC is better than MFCC.The Gaussian mixture modeling(GMM) techniques are increasingly being used in speaker recognition. We have proposed a modification to the standard diagonal GMM. In the new model (named as OGMM), there is an orthogonal transform matrix.,feature vectors are first transformed to the space spanned by eigenvectors of the covariance matrix before applying to the diagonal GMM. It is shown that with the same number of mixtures, the OGMM always gives a better performance. To reach a specific performance level, the OGMM needs only one-fourth the number of mixtures used by the standard GMM.