Fast and Efficient Speech Signal Classification with a Novel Nonlinear Transform

Abstract

This paper introduces the RD transform (RDT), inspired from reaction-diffusion mechanisms in a class of cellular nonlinear networks (CNNs). Such CNNs can be efficiently used to implement the transform but here we will introduce RDT as a general purpose algorithm. While having a computational complexity with several orders of magnitude less than traditional (e.g. DCT, Mel Cepstral, etc.) methods, RDT it is shown to be well suited for signal classification, recognition and detection. Several examples are provided for the problem of speech recognition in the case of multiple-class and users showing performance similar to that obtained with traditional methods but with an important reduction of the implementation costs.