Discriminative Preprocessing of Speech

Discriminative Preprocessing of Speech

In this book, a general framework for feature enhancement is proposed which uses a multilayer perceptron (MLP) to achieve optimal speaker dis­crimination. First, to train this MLP a subset of speakers (speaker basis) is used to represent the underlying characteristics of the given acoustic feature space. Second, the size of the speaker basis is found to be among the crucial factors affecting the performance of a speaker recognition system. Third, it is found that the selection of the speaker basis can also influence system performance. Based on this observation, an automatic speaker selection approach is proposed on the basis of the maximal average between-class variance. Tests in a variety of conditions, including clean and noisy as well as telephone speech, show that this approach can improve the performance of speaker recognition systems. Further, an alternative feature representation is proposed in this book, which is derived from what we call speaker voice signatures (SVS). These are trajectories in a Kohonen self organising map (SOM) which has been trained to represent the acoustic space. This feature representation is found to be complementary to the baseline feature set. Finally, this book finishes with a number of potential extensions of the pro­posed approaches. This book is addressed to professionals in speech processing and students in this domain.