SPECTRAL MODELS OF INDIVIDUAL SPEAKERS

Milan Sigmund

   The basic idea of the presented speaker analysis approach is to evaluate the spectral model corresponding to the anatomy of the vocal tract of the speaker independent of the actually pronounced phoneme. The procedure for determining the speaker-specific average spectrum is based on the linear predictive coding (LPC) approach. Experimental results show the evolution of a long-time spectrum with respect to the duration of the text independent utterance compared with a specially isolated word vocabulary chosen for the Czech language, a long-time spectrum variability between and within speakers, the effect of the emotional state of the speaker on the long-time spectrum and speaker normalisation by the long-time spectrum. The method is suitable for both text-dependent and text-independent tasks.

Keywords: speaker recognition, individual long-time spectrum

[full-paper]