Fusion of deep shallow features and models for speaker recognition

We propose a features fusion and a models fusion approach for speaker recognition to further improve the performance of speaker recognition. The proposed method of deep and shallow features fusion describes the speaker information more comprehensively because of the complementarity between different level features; the other method fusions the I-Vector extracted from different speaker recognition systems and can combine the advantages of different speaker recognition system. The experimental results show that, the relative improvements from the proposed framework compared to a state-of-the-art system are of 54.8% and 69.5% relative at the equal error rate when evaluated on the CASIA North and South dialect corpus. Proved that the proposed method is effective.