Text-independent speaker verification is an interesting task where the use of Gaussian Mixture Models is almost a must. Nevertheless, some preliminar encouraging results obtained in previous works using ANN in speaker verification have led us to consider to perform a direct comparison between these different methods. In this sense, this paper is only focused on the classification stage of both GMM-based and ANNbased speaker verification systems. Experiments are accomplish making use of the AHUMADA/GAUDI spanish speech database, specially oriented for speaker-recognition tasks as it contains multisession and multichannel data of about 500 speakers. Results confirm a better performance when using GMM-based system and microphonic speech but, on the other hand, when testing in specific conditions and with real telephone speech ANN outperforms GMM results.

Full Paper

Bibliographic reference.  Vivaracho, Carlos E. / Ortega-García, Javier / Alonso, Luis / Moro, Quiliano I. (2001): "A comparative study of MLP-based artificial neural networks in text-independent speaker verification against GMM-based systems", In EUROSPEECH-2001, 1753-1757.