The combination of iVector extraction and Probabilistic Linear Discriminant Analysis (PLDA) model forms a basis of the current state of the art speaker verification. The PLDA model makes an assumption that the within-speaker (or inter-session) variability in the iVector space is independent of speaker identity. In this work we propose a new model, which can be seen as an extension of PLDA, relaxing this assumption and allowing the within-speaker variability to be different for different locations of speakers in the iVector space. The potential of the proposed model is demonstrated in preliminary experiments.
Bibliographic reference. Lei, Yun / Burget, Lukáš / Scheffer, Nicolas (2012): "Bilinear factor analysis for i-vector based speaker verification", In INTERSPEECH-2012, 1588-1591.