Abstract

In this paper, a new spectral representation is introduced and applied to speech recognition. As the widely used LPC autocorrelation technique, it arises from an optimization approach that starts from a set of M+ 1 autocorrelations estimated from the signal samples. This new technique models the analytic spectrum (Fourier's transform of the causal autocorrelation sequence) by assuming that its cepstral coefficients are zero beyond M, and uses an extremely simple algorithm to compute the nonzero coefficients. In speech recognition, the same Euclidean cepstral distance measure that is the object of the optimization is also used to calculate the spectral dissimilarity. Preliminary recognition tests with this technique are presentad.