Algoritmo Competitivo Aplicado ao Reconhecimento Autom�tico da Identidade Vocal de Locutores

In recent works, two unsupervised algorithms (referred to as SSC and KMVVT) were successfully applied for vector quantization codebook design, leading to good results in signal processing applications. The present paper is concerned with a comparative study of SSC, KMVVT and the traditional LBG algorithm for designing codebooks of acoustic parameters in a speaker identification system based upon vector quantization. It is shown that SSC is a promising technique since it leads to good recognition rates (up to 97.8%), significantly higher than the ones obtained by using KMVVT or LBG. The authors also show that the speaker identity seems to be suitably represented by cepstral coefficientes. In fact, these acoustic parameters lead to higher recognition rates when compared to those provided by -cepstral coefficients.