Wavelet Based Multi-Scale Principal Component Analysis for Speech Enhancement

The goal of speech enhancement varies according to specific applications, such as to reduce listener fatigue, to boost the overall speech quality, to increase intelligibility, and to improve the performance of the voice communication device. This paper presents Multiscale principal component analysis (MSPCA) for denoising of single channel speech signal. Principle Component Analysis (PCA) is a standard tool in modern data analysis because it is simple method for extracting relevant information from complex data matrix using eigenvalues and eigenvectors. The multiscale principal component generalizes the usual PCA of a multivariate signal seen as a matrix by performing simultaneously a PCA on the matrices of details of different levels. In multi scale Principal Component Analysis (MSPCA) decorrelate the variables by extracting a linear relationship and wavelet analysis. In addition, a PCA is performed also on the coarser approximation coefficients matrix in the wavelet domain as well as on the final reconstructed matrix. By selecting conveniently the numbers of retained principal components, interesting simplified signals can be reconstructed Wavelet analysis of speech signal segments the voice information content at different Wavelet scales. At subband levels or scales multivariate data matrix are formed using Wavelet coefficients extracted from the same scales of voice signals. At each subband matrix or scales, PCA is used for noise reduction. Qualitative performance is evaluated and quantitative performance of denoising effect is measured by input/output signal-to-noise ratio (SNR), segmental SNR and IS measure. KeywordsMultiscale Principal Component Analysis (MSPCA), Principal component Analysis (PCA), Speech Enhancement, Signal Denoising, Wavelet Analysis.