Avaliação dos trechos sonoros e não-sonoros do sinal de fala para identificação do locutor independente de texto

Abstract

Verify which acoustic parameters lead to better discrimination among speakers; for this topic the investigated parameters were mel-cepstrum, pitch and energy-profile, as well as their first and second derivatives; to indentify which classifier performs better in termsof recognition rate; a benchmark among Hidden Markov Models, neural networks, vectorial quantization and gaussianas mixture models was performed to investigate the influence of the voiced and unvoiced parts of speech in the performance of speaker recognition systems.