Gaussian dynamic warping (GDW) method applied to text-dependent speaker detection and verification

Abstract

This paper introduces a new acoustic modeling method called Gaussian Dynamic Warping (GDW). It is targeting real world applications such as voice-based entrance door security systems, the example presented in this paper. The proposed approach uses a hierarchical statistical framework with three levels of specialization for the acoustic modeling. The highest level of specialization is in addition responsible for the modeling of the temporal constraints via a specific Temporal Structure Information (TSI) component. The preliminary results show the ability of the GDW method to elegantly take into account the acoustic variability of speech while capturing important temporal constraints.