M. Tistarelli and M.S. Nixon (Eds.): ICB 2009, LNCS 5558, pp. 484–493, 2009. © Springer-Verlag Berlin Heidelberg 2009 Support Vector Machine Regression for Robust Speaker Verification in Mismatching and Forensic Conditions Ismael Mateos-Garcia, Daniel Ramos, Ignacio Lopez-Moreno, and Joaquin Gonzalez-Rodriguez ATVS – Biometric Recognition Group, Escuela Politecnica Superior, Universidad Autonoma de Madrid, C. Francisco Tomás y Valiente 11, 28049 Madrid, Spain {ismael.mateos,daniel.ramos,ignacio.lopez, joaquin.gonzalez}@uam.es Abstract. In this paper we propose the use of Support Vector Machine Regres- sion (SVR) for robust speaker verification in two scenarios: i) strong mismatch in speech conditions and ii) forensic environment. The proposed approach seeks robustness to situations where a proper background database is reduced or not present, a situation typical in forensic cases which has been called database mismatch. For the mismatching condition scenario, we use the NIST SRE 2008 core task as a highly variable environment, but with a mostly representative background set coming from past NIST evaluations. For the forensic scenario, we use the Ahumada III database, a public corpus in Spanish coming from real authored forensic cases collected by Spanish Guardia Civil. We show experi- ments illustrating the robustness of a SVR scheme using a GLDS kernel under strong session variability, even when no session variability is applied, and espe- cially in the forensic scenario, under database mismatch. Keywords: Speaker verification, forensic, GLDS, SVM classification, SVM regression, session variability compensation, robustness. 1 Introduction Speaker verification is currently a mature technology which aims at determine whether a given speech segment of unknown source belongs to the identity of a claimed individual or not. Among the most important challenges of a speaker verifica- tion system is the robustness to the mismatch in conditions between training and test- ing utterances, being its compensation a main factor for the improvement of system performance. Recently, this task has been carried out by the use of data-driven session variability compensation techniques based on factor analysis, which have become the state of the art in these technologies as can be seen in the periodic NIST Speaker Recognition Evaluations (SRE) [1]. Such techniques can be applied to the best- performing systems working at the spectral level, mainly based on Gaussian Mixture Models (GMM) [2] and Support Vector Machines (SVM) [3], increasing their robustness and accuracy. Among all the different compensation variants, the Nuisance