ABSTRACT
This chapter presents an overview of the commonalities among different methods proposed. It considers the more basic Euclidean distance method and explores an informative–theoretical measure for kernel design. Quadratic time–frequency (TFRs) representations can be uniquely characterized by an underlying function called a kernel. Class-dependent TF representations use a discriminative approach to kernel design. In earlier time–frequency research, kernels for a number of properties, such as finite-time support and minimal quadratic interference, have been derived. The chapter illustrates how the theoretical results and approach presented earlier can be applied to acoustic signal classification, in general, and to speech recognition, in particular. For the isolated phone recognition task, individual phones were extracted from continuous speech, and subsequently classified. The phones were grouped into several confusable sets, where within each set phone-based speech recognition systems usually experience confusion. The class-dependent kernel was developed to extract from the TF representation the dimensions useful for classification.
