From top left
Speech processing, biometrics and man-machine communication
The research activities of the Speech Processing and Biometrics Group (Groupe de traitement de la parole et de biométrie (GTPB)) include several projects in three main research fields of speech processing, biometrics and man-machine communication. The projects are supported not only by the Swiss National Science Foundation (FNS) and the EPFL but also by the European Commission, the Swiss Commission for Technology and Innovation (CTI), the Swiss Federal Office for Education and Science (OFES), industrial partners and end users.
The first group of research projects, the GTP has been focused on, is on speech processing in adverse environments for communication and forensic application purposes.
The aim of the second group of projects is to develop a formulation of the basic concepts of robust biometric recognition, in particular speaker, dynamic signature, face and fingerprint recognition over wired and wireless Internet data networks in the context of missing data theory, multi-modal feature extraction and statistical modelling, and to make such systems computationally efficient. The GTPB works on various theoretical and practical issues concerned with estimation of missing and unreliable biometric data, and how such estimates may be used in biometric recognition algorithms. The unreliable data processing approach is also adapted to the multi-modal biometric data integration.
The third group of research projects concerns man-machine multi-modal communication, in particular voice enabled interfaces for autonomous mobile robots, and aims at developing a novel and comprehensive theory of man-machine speech communication in the case of mobile interfaces. The objective is to integrate smart speech signal acquisition, automatic speech recognition and understanding, speech synthesis and dialogue systems, into the realistic multi-modal, man-machine interface for mobile service robots. The innovative focus of the projects is the application of probabilistic (Bayesian network) inference and learning to the problems of behavioural information acquisition (speech, vision (video camera) and laser scan) and control in autonomous robots when communicating with users. In particular, a new probabilistic model based architecture for error handling in combined human-robot spoken dialogue and interaction systems has been investigated.