FaBiole

This project takes part in the biometric speech authentication area. More precisely, it concerns the interest of “phonetic information” in this context. Since about 15 years, Automatic Speaker Recognition systems are evaluated by the NIST thnaks to the Speaker Recognition Evaluation (SRE) campaigns. The story of these evaluations shows an important improvement in terms of performance, which authorizes to think about numerous applications of speaker recognition, and particularly in the forensic area. The potential consequences of such an application lead us to think about the reliability of the performance criterion currently used in order to evaluate the systems. Indeed, these performance criterion are estimated globally on a large set of biometric voice tests (voice comparisons). The criterions are mainly a global Equal Error Rate and a global decision cost obtained by a Decision Cost Function (DCF). The measures are computed as an average obtained from a large set of tests which are coming from different speakers and conditions. These measures don’t really reflect the practical application context where an answer must be returned by a system in a specific situation. In such a situation, only two specific recordings belonging from one or two specific speakers and with a specific content are available. More precisely, these measures don’t take care about both the differences between a speech recording and a speaker as well as and the differences between the speakers. Some recent works demonstrate the limits of these performance measures. This project has two objectives. Firstly, the acoustic and phonetic factors which are linked to inter speaker variability will be studied. The aim is not longer to speak about a global performance but to highlight the part of the inter speaker variability explained by each of these acoustic/phonetic factors. Lastly, these results will be used to propose a confidence measure dedicated to biometric voice authentication/comparison area. These confidence measure will work only on the two speech recordings linked to a given voice comparison, independently of the speaker recognition approach . Basically, the speaker characteristic information contained in the two recordings will be quantified and their coherence estimated in order to compute the confidence measure. Even if this project will provide very useful information on the speaker specific characteristics, the main objective will be to push the biometric voice comparison domain to make a big step in direction of reliability evaluation, instead of staying in the paradigm of performance evaluation. The role of automatic speaker recognition systems will change after this project. Now, they have to take a decision and to evaluate the confidence on this decision. After, they will have only to propose the binary decision when the reliability of the decision will be provided by the outcomes of this project. The potential effect of this project on the real life is very large. The results will authorize a better comprehension of the potential and limits of forensic voice comparison but also to increase drastically the reliability of speaker recognition applications in the commercial field. The independency between the confidence measure and the speaker recognition system itself will virtually authorize to integrate the results of this project in all existing applications…