Comparison of Different Kinds of Long-Time Spectra of Voice Estimated by Modified Linear Prediction to Distinguish Speakers

This paper deals with two kinds of long-time spectra of speech estimated by the linear prediction approach. The standard approach usually used in short-time analysis was modified in two ways to achieve the long-time effect - either autocorrelation coefficients (AC) or predictive coefficients (PC) were averaged over a period of 2 minutes. The spectra were computed using order of prediction from 6 to 22 and evaluated in terms of diversity for a group of 17 speakers. To distinguish speakers, the most appropriate frequencies seem to be around 1010 Hz (AC averaged) or around 340 Hz (PC averaged).