Application of fusion techniques to speaker authentication over IP networks

The authors explore the benefits of using fusion techniques for combining the outputs of different speaker authentication systems in a particular scenario which corresponds to simulated IP networks. IP networks are characterised by packet losses. Two approaches for IP network modelling are compared: a two-state Gilbert model and a bottleneck network simulation. The aim is to investigate the effect of isolated packet losses against bursts of losses. The authentication systems considered are speaker verification (accepting or rejecting the claimed identity of a speaker by processing samples of his/her voice) and utterance verification (using a set of speaker-independent speech models to recognise a certain utterance). Both systems are embedded in a client/server architecture. Results show the speaker verifier is less affected by packet losses than the utterance verifier. While the former behaves similarly under both IP models, the latter shows more sensitivity to bursts of losses. In any IP model, fusion techniques improve the robustness of the combined authentication system and yield a balanced working point.