A flexible software and hardware platform has been realised. From this
platform, three main demonstrators have been developed, evaluated and
installed at several end-user sites :
Level 1 : A network-based voice mail system with access control using robust speech
recognition technology and rejection.
Level 2 : The same application reinforced by two additional vocal modalities
(text-dependent and text-independent verification techniques in addition of speech
recognition)
Level 3 : An access control to buildings realised using a badge reader (or keyboard) for
identification and the level 1 system complemented with one face or profile recogniser
for verification.
Field tests in real condition have been realised on early version of all demonstrators. General trends
showed that Level 1 system was well accepted by users but not Level 2 and Level 3 systems. Users felt
that the constraints were too high (response time, performances) compared to the service provided
(Secured access). These demonstrators were based on background technologies and thus clearly
showed the need for innovative multi modal verification techniques.
Innovative Multimodal Verification Techniques
Concurrently to the development of the different platforms, work was
carried out to develop innovative multi-modal verification techniques.
The key techniques introduced include :
Frontal face recognition algorithms with very low error rate:
4.9% to 7% Equal Error rate (ERR) on the
M2VTS multimodal database,
depending on the technique used. Most of these techniques run very
efficiently (less than a few seconds on modern processors).
Profile recognition (7% ERR)
Lip tracking techniques
Speech verification techniques leading to very low error rates:
less than 5% ERR.
Fusion techniques that permit to combine verification scores of
several modalities and that achieve multimodal verification with Equal
Error Rate as low as 0.7%
(tested on the M2VTS multimodal database).
Algorithm Optimisations and Integration
Some of these techniques have now been integrated in
the fast prototyping platforms and are currently tested in real conditions.
These tests already permit to iterate a back and forth collaboration between
industrial and academic experts in order to optimise and to enhance
the robustness of the algorithms in real conditions. Since these algorithms
already achieved far better performance than the background technology,
there is no doubt that the acceptance by users will be significantly higher.
Multimodal Databases
To better assess the performances of the verification algorithms, an
additional multimodal database is currently being recorded
(300 persons, 20-30 s of video sequence (head rotation from left to right),
10 s of speech) in clean conditions (blue background,
controlled lighting, etc). Currently real conditions databases
are also built during field tests. In particular, situations
such as non uniform lighting, smiling faces, or scale are represented.
Applications
Prototypical applications have been developed
for secured access to information
systems (teleservice), secured access to buildings
and Alarm verification. For example, in a
teleservice application, the system includes a camera
to acquire some frontal images of the face of
the user, and a microphone to record the voice of the user.
