Combining Information from Multiple Classiiers for Speaker Veriication

in Proc. of Speaker Recognition and its Commercial and Forensic Applications, France, April 1998. Combining Information from Multiple Classi ers for Speaker Veri cation Sangita Sharma1, Pieter Vermeulen1, Hynek Hermansky1;2 1Oregon Graduate Institute of Science and Technology, Portland, Oregon, USA. 2International Computer Science Institute, Berkeley, California, USA. sangita,pieter,[email protected] Abstract Given two independent speaker veri cation systems, it is reasonable to expect some uncorrelated errors. If some trends in this behavior can be detected, one should be able to improve the performance of the combined system beyond that of the best individual system. In this paper, we use a non-linear combiner to combine results from two independent speaker veri cation systems. Experiments conducted on a subset of the 97 NIST Speaker Evaluation Task, show that the combination improves equal error rate (EER) by up to 9% when the two systems have comparable performance. When one system outperforms the other, the performance of the combiner is similar to that of the best system. Experiments using simulated data show that the combination yields large improvement (21%) in performance when the two systems outperform each other under di erent operating conditions. R esum e Donn e deux syst emes ind ependants de v eri cation de locuteur, ils devraient avoir quelques erreurs noncorr elatives. Si de telles tendances dans les erreurs peuvent être d etect ees, alors une combinaison de ces syst emes donnez l'erreur inf erieure que le meilleur syst eme individuel. En cet article, nous etudions l'utilisation d'un combinateur non lin eaire de combiner les r esultats de deux syst emes ind ependants de v eri cation de locuteur. Les r esultats indiquent les conditions dans lesquelles le combinateur peut r eduire l'erreur (EER).