Integrating Online i-vector into GMM-UBM for Text-dependent Speaker Verification


GMM-UBM is widely used for the text-dependent task for its simplicity and effectiveness, while i-vector provides a compact representation for speaker information. Thus it is promising to fuse these two frameworks. In this paper, a variation of traditional i-vector extracted at frame level is appended with MFCC as tandem features. Incorporating this feature into GMM-UBM system achieves 26% and 41% performance gain compared with DNN i-vector baseline on the RSR2015 and RedDots evaluation set, respectively. Moreover, the performance of the proposed system that trained on 86 hours data is on par with that of the DNN i-vector baseline trained on a much larger dataset (5000 hours).

In Proceedings of APSIPA Annual Summit and Conference, Kuala Lumpur, Malaysia, 2017