Energy Separation Algorithm Based Spectrum Estimation for Very Short Duration of Speech

Abstract

In this paper, we propose a novel method of estimation of short-time spectrum for analysis of speech signals in the closed phase regions of glottal activity. This method uses Teager Energy Operator (TEO) and a related Energy Separation Algorithm (ESA) iteratively, along with the design of digital resonator to estimate formants from a very short duration of the speech. The spectrum of cascade of these four resonators is referred to as our proposed ESA spectrum of speech. The novelty of the proposed approach lies in using very short duration of analysis speech frame that is synchronized with glottal closure instant (i.e., about 1-2 ms) to estimate the proposed spectrum in order to ensure that the vocal tract system characteristics do not change much within this interval and to alleviate erroneous estimation of formants due to nonlinear interaction of excitation source with the vocal tract system. To demonstrate the effectiveness of proposed algorithm for formant estimation on speech data, we have used 1.5 ms speech signal corresponding to closed phase glottal cycles derived from a male speaker of CMU-ARCTIC database.