Simplified residual factor analysis for text-independent speaker verification
Lei Zhu, Rong Jian Zheng, Bo Xu · 2010
Joint factor analysis (JFA) is a model used to treat the problem of speaker and session variability in GMM. In this paper, we add an additional direction to the joint factor analysis to model the homogeneous variability between the train utterance and the test utterance directly. This work is a natural extension of earlier work by P. Kenny [1] that shifted the model to test utterance along the channel subspace. We referred the new work as residual factor analysis (RFA). Then we proposed a new strategy called simplified residual factor analysis (S-RFA) to speedup the RFA procedure. We tested the proposed technique on the core condition of NIST 2008 speaker recognition evaluation (SRE08) and obtained about 4.4% relatively reduction in equal error rate (EER), compared to the JFA system with merely no extra CPU time.