Compensating Utterance Information in Fixed Phrase Speaker Verification

Rohan Kumar Das, Maulik C. Madhavi, Haizhou Li

Abstract: This work investigates on explicitly utilizing utterance information for fixed phrase speaker verification (SV). In this scenario, the same phrase is spoken by the speakers during the training and testing sessions. In other words, the speaker model possesses both speaker as well as utterance information. Therefore, there is a potential to improve the speaker characterization by compensating the utterance information. In this work, we propose a framework to compensate the utterance information, which is used to normalize the lexical content. A hidden Markov model (HMM) based triphone model is considered as a universal background model (UBM). It is used for adapting the speaker-utterance model and the background utterance model in the proposed utterance compensation framework. Given a test utterance and a claimed speaker-utterance model, the UBM as well as background utterance model is utilized for compensating the average speaker information and the lexical content information, respectively. The studies are conducted on RSR2015 database, which reveal the importance of the proposed utterance compensation framework as compared to the framework without utterance compensation.