A robust speech feature-perceptive scalogram based on wavelet analysis

Abstract

In real world applications, additive noise will contaminate input speech features for speech recognition and representation when speech recognition systems are working in real environments. There have been many attempts made to find a robust speech feature. In this paper, we propose a robust speech feature, the perceptive scalogram, for speech representation and recognition. The new feature is based on some propositions which state that a human's perception of speech is a perception of specific components of sounds, and the components have a specific changing rate of their short-time spectrum. The proposed perceptive scalogram also takes consideration of the fact that speech is non-stationary, and uses wavelets as its signal analysis tool. Simulation results show the robustness of the perceptive scalogram against additive Gaussian noise.