Speech Enhancement Using Principal Component Analysis and Variance of the Reconstruction Error Model Identification

In recent years, Automatic Speech Recognition (ASR) systems designed to work in controlled environments using clean speech have reached very high levels of performance. However, the accuracy of speech recognition degrades severely when the systems are operated in noisy environments. In this thesis we address the problem of single-channel speech enhancement. Starting with a study of the state-of-the-art enhancement methods, a comprehensive study of different categories of speech enhancement is presented. As an important class of speech enhancement methods, subspace-based speech enhancement is presented in chapter 2. After a careful study of all forces and drawbacks of this technique, a generalized form of Principal Component Analysis-based (PCA-based) speech enhancement is provided next. As a vital issue in PCA-based enhancement methods, identification of the clean speech signal's model is investigated in chapter 3. Some recent techniques to define the rank of a clean speech signal are presented in this chapter. In the rest of the thesis, a novel technique for rank estimation is developed. We introduce therefore a novel approach for the optimal subspace partitioning using the Variance of the Reconstruction Error (VRE) criterion. This criterion provides consistent parameter estimates and allows us to implement an automatic noise reduction algorithm that can be simply applied to the observed data. This choice also overcomes many limitations encountered with other selection criteria, like overestimation of the signal subspace or the need for empirical parameters. We have also extended our subspace algorithm to take into account the case of colored and babble noise. Informal listening tests and illustrations have confirmed the method to be numerically noise robust regardless of the type of the noise. ii Acknowledgements