Partenaires et organisations internationales
(Anglais)
|
Ubilab (CH), Swisscom (CH), ENST (F), IRISA (F), PTT-Telecom (NL), KPNResearch (NL), KUN (NL), Fortis (NL), KTH (S), Telia (S), and Vocalis (UK)
|
Résumé des résultats (Abstract)
(Anglais)
|
PICASSO develops and tests prototypes of secure telematics transaction services using caller authentication by voice. The services are accessible via the worldwide telephone network, either the Plain Old Telephone System (POTS) or cellular networks. These services can include actions that incur financial obligations (e.g. calling card calls, tele-shopping, and other kinds of electronic commerce). Likewise, services may directly involve financial transactions (moving money between accounts, possibly of different owners) or provide access to private information (e.g. a multi-media mail box in a telecommunication service). To provide intuitive and easy to use interfaces that are secure against intruders, PICASSO integrates speech recognition and speaker verification/identification technology.
This is why the PICASSO project has two principal objectives. Firstly, it aims at evaluating and demonstrating the recognized state of the art technology developed in the preceding CAVE project (also supported by OFES). This evaluation is being done in several telecom, and banking application environments. Secondly, the CAVE technology is to be further improved. The developments pursued are along the following main directions: i) customized password, ii) incremental enrollment, iii) extension of the technology to cellular networks, iv) improvement of the decision strategy, and v) robustness when used in real-life environments.
In this project, Ubilab has been mainly active in technical and algorithmic research and development. During the second year, the main contribution of Ubilab was the development of new incremental enrollment techniques for Automatic Speaker Verification. This work was intended to cope with the major issues related to unsupervised incremental enrollment. In particular, a comparison between batch and adaptation methods highlighted their advantages/drawbacks and demonstrated the need for adaptation-based incremental enrollment when the initial client model training is trained on a very limited amount of data.
Along these lines, Ubilab continued the integration of algorithms developed by the PICASSO consortium into a common research platform.
|
