Topic: Biometric authentication using eye-movements

Contact persons: Tomi Kinnunen and Roman Bednarik

Description: Eyes have been used for a long time to recognize persons: iris,
retina scan, location of the eyes in face images. However, there is very
little studies about using dynamics of the eyes as a biometric cue. In this
topic, the student should build a highly novel biometric authentication
technique based on the eye-movement tracking (EMT) signal. The study
consists of developing simple algorithms for fixation/saccade segmentation,
digital filtering of EMT signals, building a state-of-the-art machine
learning classification algorithm and doing extensive testing on
self-collected data. The topic gives a chance to continue to PhD studies.

Salary: N/A at this phase, funding applications pending on several sources
(possible later)
==========================================================================

Topic: Voice conversion
Contact persons: Tomi Kinnunen

Description: Voice conversion refers to "imitation by computer" - that is,
we want to transform the voice characteristics of one person so that his
voice timbre is mapped to another person. It has numerous applications in
entertainment, movie industries (replacing voice actors), and can also be
used for speech enhancement in dedicated applications. In this study, the
student is expected to implement a voice conversion system using a so-called
STRAIGHT vocoder and Gaussian mixture model based regression and experiment
with several new research ideas. The topic will include both subjective and
objective evaluation of the performance. The topic gives a chance to
continue to PhD studies.

Salary: N/A at this phase, funding applications pending on several sources
(possible later)

==========================================================================

Topic: Acoustic feature extraction for speech biometrics
Contact persons: Tomi Kinnunen

Description: Voice-based biometrics, or speaker recognition, has a huge
market potential in telephone-based services, but at the same time the
problem continues to be extremely challenging due to many noise sources:
background noises, varying microphone, intra-person variability, just to
scratch the surface. Most of the state-of-the-art speaker recognition
systems parameterize the speech signal into sequences of acoustic feature
vectors known as mel-frequency cepstral coefficients (MFCCs). These features
have been more or less fixed the past 20 years already, even though they are
very sensitive to noises. A large number of alternative features have been
tried but none have really replaced the MFCCs yet. In this topic, the
student will survey, implement, and compare several alternative feature
extraction methods for voice biometrics and also implement new features. The
topic will include extensive experimental evaluation on the state-of-the-art
speaker recognition benchmark datasets and statistical machine learning
algorithms. Some of the infrastructure to enable this evaluation already
exists so the main focus will be on the front-end optimization. The topic
gives a chance to continue to PhD studies.

Salary: N/A at this phase, funding applications pending on several sources
(possible later)

===========================================================================

Topic: Session compensation for voice biometrics

Contact persons: Tomi Kinnunen

Description: Voice-based biometrics, or speaker recognition, has a huge
market potential in telephone-based services, but at the same time the
problem continues to be extremely challenging due to many noise sources:
background noises, varying microphone, intra-person variability, just to
scratch the surface. Most of the state-of-the-art speaker recognition
systems parameterize the speech signal into sequences of acoustic feature
vectors known as mel-frequency cepstral coefficients (MFCCs). These features
have been more or less fixed the past 20 years already, even though they are
very sensitive to noises. In this topic, the student will survey and compare
several state-of-the-art session compensation methods such as joint factor
analysis (JFA) and nuisance attribute projection (NAP). The topic will
include extensive experimental evaluation on the state-of-the-art speaker
recognition benchmark datasets and statistical machine learning algorithms
and it requires a solid mathematical background, or at minimum, a fearless
attitude towards statistical modeling. The topic gives a chance to continue
to PhD studies.

Salary: N/A at this phase, funding applications pending on several sources
(possible later)

========================================================================

Topic: Use of Kantele (the traditional Finnish string instrument)  
as user-interface for computer

Contact person: Tomi Kinnunen

Salary: N/A