Speaker identification using singular value decomposition of LPC spectral magnitudes

An experimental procedure for identification of speakers using singular value decomposition of spectral magnitudes of speech is described. Amplitude spectra at 10 discrete frequencies from 500 Hz to 5000 Hz, which are derived from a tenth order short-time linear predictive coding (LPC) model, are used as pattern vectors characterising speech utterances. The set of stored reference spectra yielding the largest ratio of singular values with the spectra of an unknown utterance is considered closest to the unknown. An overall identification score of 70% to 87.5%, depending on the frequency used, was observed on a set of 40 test utterances