Classification of visual speech features

• Word=time sequence of visemes (mouth states) => temporal evolution of visual speech features is important for recognition:

  • need for dynamic classifiers, to model transitions between mouth states = model temporal evolution of data:
  • HMM (Hidden Markov Models): the most frequently used:
  • TDNN (Time-Delayed Neural Networks): can perform recognition based on the temporal variation of visual speech features, not only static values.

Department of Informatics

Aristotle University of Thessaloniki

• Word recognition:

Previous slide

Next slide

Back to first slide

View graphic version