Combining Spectral and Prosodic Features in HMM-based Single Utterance Speaker Verification
Osman Büyük, Levent M. Arslan · 2015
In this paper, we combine spectral and prosodic features together in order to improve the verification performance on a text-dependent single utterance speaker verification task. The baseline spectral system makes use of a whole-phrase sentence HMM topology for the fixed utterance. We extract prosodic features using time alignment information obtained from the HMM states. In our experiments we observe that, although the prosodic features individually do not yield high performance, they provide complementary information to the spectral features. We achieve approximately 10% relative reduction in EER when the information sources are combined with a multi-layer neural network.