Formant extraction from group delay function

Abstract

This paper presents an approach based on the properties of group delay functions for extracting formants from speech signals. The algorithm is similar to the cepstral smoothing approach for formant extraction using homomorphic deconvolution. The significant differences are (i) the logarithmic operation is replaced by ()^r operation and (ii) the additive and high resolution properties of group delay function are expolited to emphasize formant peaks. The group delay function (or the negative derivative of the Fourier transform phase) is derived for a signal which in turn is derived from the Fourier transform magnitude of the speech signal. If a suitable value of r is used, this method gives highly consistent estimates of formants compared to both the cepstral approach and the model-based linear prediction (LP) approach for smoothing the magnitude spectrum. The effects of the parameters, exponent r and window width p, on the proposed technique for formant extraction are studied.