On the use of linear prediction error energy for speech and speaker recognition

Abstract

The error energy of the linear prediction inverse filter seems to give an effective triter ion for speech and speaker recognition. In this task, reference sounds are stored in terms of their filter coefficients and the error energy of their optimal inverse filters. The error energy of an unknown sound from each reference filter is compared with the reference error energy (Method A). Alternately, the error energy of the optimal inverse filter for an unknown sound is compared with the error energy of the same sound from the reference filters (Method B). Method A has its advantage in requiring only the autocorrelation function of unknown sounds. The recognition task was conducted with ten American vowels in the context of “hVd” produced by nine male and seven female speakers. A single speaker system averaged 99% recognition, whereas a multispeaker system in which one of the speakers was chosen as the reference averaged 46%. A proper threshold for the criterion suggested that the method is effective for speaker recognition. In general, Method A was slightly inferior to Method B. This work is part of a study of the effectiveness of inter‐speaker normalization. [The work was supported by the Office of Naval Research.]