Separated Modeling of Speaker Dependent and Independent Vocal Characteristics for Text-Independent Speaker Verification
Wei-Ho Tsai · International Journal of Information and Electronics Engineering · 2014
This study proposes an approach to model a speaker's voice by assuming that the voice can be decomposed into a speaker-dependent (SD) component and a speaker-independent (SI) component. Using the Universal Background Model (UBM) to approximate the characteristics of the SI component, we derive the SD component by finding the optimal combination of the two models that matches the given speech data. Results of the experiments conducted using the 2001 NIST speaker recognition evaluation corpus demonstrate the superiority of the proposed approach over the conventional speaker verification based on GMM-UBM.