ON INSTANTANEOUS AND TRANSITIONAL SPECTRAL INFORMATION FOR TEXT-DEPENDENT SPEAKER VERIFICATION

We present a study of the effectiveness of instantaneous and transitional spectral information for text-dependent speaker verification. Instantaneous information, represented by contours of cepstral and normalized cepstral coefficients, describes the variation of the short-time speech spectra. Transitional information, represented by first-order orthogonal coefficients which result from the orthogonal polynomials expansion of the cepstral contours, describes the speed of variation of the speech spectra. Cepstral and first-order orthogonal coefficients vectors were employed separately as well as joined into a single feature vector, in classical DTW-based verification algorithms. Investigations on a population of 22 speakers (high-quality speech) showed that the elimination of the time-invariant spectral components from the speech features, taking place when performing cepstral normalization or computing first-order orthogonal coefficients, brings a substantial reliability improvement. Furthermore, transitional information is practically as effective as instantaneous information, whereas combining both kinds of information does not lead to further improvement. © 1990.