Adriano Petry, Adriano Zanuz, Dante Augusto, and Couto Barone
Universidade Federal do Rio Grande do Sul, Brazil

This paper describes a new method to perform a speaker identity classification. The proposed method is based in a probabilistic distance measure—the Bhattacharyya distance. First the speech samples are pre-emphasized, then 16 LP derived cepstral coefficients are extracted and a lifter window is applied to these feature parameters. Analyzing the feature histograms, it is possible to model them by gaussian probability densities. So the distance measure proposed by Bhattacharyya is applied in way to effectively classify an unknown speech sample. This method have been presented high accuracy, when tested with over 3s speech samples. The speech samples have been recorded from 20 different speakers.