Project Overview
The European CAVE project is working on the development and testing of speaker verification systems for use in telephone applications like calling-card or financial services. Although the telephone is now a widely-accepted and convenient medium for performing various sorts of transactions, there can be a major risk of fraud if access to the service is controlled only with a relatively insecure method such as the entry of a personal identification number (PIN) code using telephone keypresses. For example, fraud in the US telephone calling card market costs the industry, and individual cardholders, billions of dollars every year.
In addition, the functionality and ease of use of current telephone banking services is restricted by the authentication problem, resulting in insecure or lengthy "log-in" procedures, and the deferment of some transactions until the account holder can be notified in writing. Speaker Verification, by which a telephone caller's identity claim (currently represented by the combination of a account number and PIN-code) can be augmented with information derived from the sound of the caller's voice, is a potential solution to these problems, and could allow for fast and secure authentication of callers.
The work of CAVE
While the underlying technology performs well under laboratory conditions, publically accessible telephone services incorporating speaker verification are almost entirely unheard of. Instead, SV has mostly found its way into niche applications, such as door access control, where user populations are relatively small. Arguably one of the reasons for this is that telephone applications of speech recognition are themselves not sufficiently mature and consequently just as rare. However, speech recognition is not necessarily a prerequisite for speaker verification; in fact, there are many different methods by which SV could be deployed in existing and future telephone services.
In the initial phase of its work, CAVE investigated these possible methods, through an in-depth market study of our target business sectors of banking and telephone calling services, and by surveying the precise needs of telephone service managers in terms of a number of factors such as performance, cost, and integration issues. The public versions of each of these reports can be downloaded by clicking here.
In parallel with this work, we drew up the functional and technical specifications of two automatic demonstration systems, a telephone banking service and a calling card service. These systems are currently being built at the premises of Ubilab and KPN Research respectively. Each system will enter its initial, laboratory-test phase, with a relatively small number of users, at the beginning of 1997; after going through two further iterations, major field-tests of these systems, with hundreds of callers will be carried out in Summer 1997.
We are confident that our planned tests will give us a major new insight into the issues surrounding the deployment of speaker verification technology, such as how to handle the question of user enrolment into the SV system, and generally how best to trade off technological concerns against human-factors issues in order to provide the most secure and usable systems we can. You can download the public versions of the specifications for our systems by clicking here.
In order to provide the core technology for our systems, CAVE has also been working on laboratory research in the area of speaker verification. Our research partners have built a laboratory reference system for SV on top of Entropic Inc's popular HTK speech processing toolkit and, by running a systematic test campaign, have been achieving state-of-the-art performance on two different databases of speech recordings for SV, one of which is the widely used YOHO database.
The results of the laboratory-based work are feeding directly into the SV system being implemented by Vocalis on their "industrial-strength" speech processing platform, for our field tests. You can click here to download a conference paper giving details of our experiments. Soon, you will also be able to download a longer, more detailed report about these activities.
