On talker verification via orthogonal parameters

The method of orthogonal parameters uses the eigenvalues and eigenvectors of the matrix of covariances of measurements made on a population of utterances from a given talker to transform measurements made on an utterance from an unknown talker. The values include a subset of the principle components (obtained by Hotelling transformation), plus other values dependent on the variances of the principle components. Variations of the method are given. Earlier work applied the method to measurements which were sets of linear prediction coefficients, reflection coefficients, and logarithmic areas derived from the speech waveform, and suggested that highly accurate speaker verification could be achieved, independent of the speech transmission system, and of the text. Present work shows that there is serious dependence on the transmission system, and that accuracies in the range 90-95 percent are typical with utterances of about 2 s duration. Investigation of a new distance measure based on covariances, combined with study of the factors influencing distortion in telephone transmission, resulted in a new method based on covariances of logarithmic spectral estimates. The results of this method, giving accuracies of about 95 percent, were robust in the presence of transmission coloration. Some remaining research problems are defined.