Exploration of small enrollment speaker verification on handheld devices

Abstract

This thesis explores the problem of robust speaker verification for handheld devices under the context of extremely limited training data. Although speaker verification technology is an area of great promise for security applications, the implementation of such a system on handheld devices presents its own unique challenges arising from the highly mobile nature of the devices. This work first independently analyzes the impact of a number of key factors, such as speech features, basic modeling techniques, as well as highly variable environmental/microphone conditions on speaker verification accuracy. We then present and evaluate methods for improving speaker verification robustness. In particular, we focus on normalization techniques, such as handset normalization (H-norm), zero normalization (Z-norm) as well as model training methodologies (multistyle training) to minimize the detrimental impact of highly variable environment and microphone conditions on speaker verification robustness.