Exponential transient component analysis for speech modulation decomposition

Abstract

A novel approach, the modulation decomposition method, to help solve the segmentation and identity problems in the development of a speech model, is described. The authors attempt to show the existence of exponential functions in the energy envelope of speech segments. These functions are expected to be basic units of acoustic modulations which can be used to model the kinematics of gestures in utterances. The possible relevance of these functions to components of speech signals was explored by comparing different signals with similar phonetic elements. Three different optimization methods are compared: the Powell method, a conjugate gradient method, and a modified Newton-Raphson (Levenberg-Marquardt) method.<>