Total variability factors combination for speaker verification

Intersession compensations in the total variability space have become the mainstream approach in the field of speaker verification. However, this space training through the standard algorithm needs lots of time and memory. An effective algorithm is proposed to obtain the total variability loading matrix in this paper. We orthogonally concatenate the speaker and the channel loading matrices from joint factor analysis (JFA) system, which can save much more time and memory. Furthermore, if the total variability factors from the standard algorithm and the proposed algorithm are combined, better recognition result can be obtained. In the telephone data of the core condition trials of the NIST SRE 2008, compared with the baseline system, we yield 9.94% and 5.70% relative equal error rate (EER) reduction for the male and female respectively.