Support vector regression in NIST SRE 2008 multichannel core task
Ismael Mateos García, Daniel Ramos, Ignacio López Moreno, Joaquín González-Rodríguez · 2008
This paper explores two alternatives for speaker verification using Generalized Linear Discriminant Sequence (GLDS) kernel: classical Support Vector Classification (SVC), and Support Vector Regression (SVR), recently proposed by the authors as a more robust approach for telephone speech. In this work we address a more challenging environment, the NIST SRE 2008 multichannel core task, where strong mismatch is introduced by the use of different microphones and recordings from interviews. Channel compensation based in Nuisance Attribute Projection (NAP) has also been investigated in order to analyze its impact for both approaches. Experiments show that, although both techniques show a significant improvement over SVC-GLDS when NAP is used, SVR is also robust to channel mismatch even when channel compensation is not used. This avoids the need of a considerable set of training data adapted to the operational scenario, whose availability is not frequent in general. Results show a similar performance for SVR-GLDS without NAP and SVC-GLDS with NAP. Moreover, SVR-GLDS results are promising, since other configurations and methods for channel compensation can further improve performance.