Título: Reconhecimento de Locutores em Língua Portuguêsa com Modelos de Redes Neurais e Gaussianos
Autores: Caricatti, André Machado; Weigang, Li
Resumo: To study the speaker recognition problem, the mel-cepstral coefficients and their derivatives, deltamel-cepstral coefficients are used as the classification parameters. A database is formed from 14 speakers in a closed environment using Portuguese with noise an echo. To execute text-independent speaker recognition, self-organizing maps (SOM) and Gaussian Mixture Models (GMM) are used together with relative entropy calculations among test and reference speaker models. Specially in the GMM application, 20 seconds is enough to form the models with 32 mixtures per speaker, and the correct verifications arrives at a rate of 64% in the best case.
Código DOI: 10.21528/CBRN2001-005
Artigo em pdf: 5cbrn_005.pdf
Arquivo BibTex: 5cbrn_005.bib