An optimized time-frequency distribution for speech analysis

Abstract

In order to analyse and interpret speech signals, different time-frequency representations are used (e.g. spectrogram, Wigner-Ville distribution, wavelets). In this paper we construct within Cohen's class of time-frequency distributions the distribution that is optimally suited for the representation of speech signals. Thereby we take advantage of the special time-frequency structure of speech expressed in the Elementary Waveform Speech Model (EWSM, d'Alessandro, 1990). As an application we present an algorithm that extracts a point pattern in the time-frequency plane out of the speech signal using the optimized distribution. Thus we get a very simple representation of the speech signal that is well interpretable both for non-stationary and for stationary speech segments. Furthermore this representation could serve as a base for further analysis (e.g classification).