Odyssey 2012

Robustness due to mismatched train/test conditions is one of the biggest challenges facing speaker recognition today, with transmission channel/handset and additive noise distortion being the most prominent factors. One limitation of the recent speaker recognition systems is that they are based on a latent factor analysis modeling of the GMM mean super-vectors alone. Motivated by the covariance structure of cepstral features, in this study, we develop a factor analysis model in the acoustic feature space instead of the super-vector domain. The proposed technique computes a mixture dependent feature dimensionality reduction transform and is directly applied to the first order Baum-Welch statistics for effective integration with a conventional i-vector-PLDA system. Experimental results on the telephone trials of the NIST SRE 2010 demonstrate the superiority of the proposed scheme.