The role of 'delta' features in speaker verification

Our previous experiments in Text-Dependent and -Independent Speaker Verification (TD-SV and TI-SV) using trajectory-based models, showed that non-stationary segments benefit TD-SV but not TI-SV, because in TI-SV maximum likelihood (ML) training results mainly in stationary segments. This result questions the role of non-stationary, ‘delta’ parameters in conventional GMM-based TI-SV. In this paper we develop and study a number of GMM-based TI-SV systems for Switchboard which use combinations of static and dynamic parameters. We show that in our segmental GMM and the AFRL GMM system, the trajectory slopes and deltas focus the verification process onto the stationary regions. In our GMM systems, however, the deltas are modelling some speech dynamics. The different functions of deltas may be due to different system settings and frontend processing (e.g. RASTA, speech noise detector). This indicates that the role of delta parameters in GMM-based speaker verification systems is more complex than simply “modelling dynamics”. Our results also show that the superior performance obtained with front-end parameterizations which combine static and delta parameters only emerges after RASTA filtering; without RASTA filtering a ‘delta-only’ front-end performs best.