Classification of visual speech features
Recognition not affected by temporal transitions between the frames of the videosequence; one frame is enough for letter recognition
? use static classifiers; no need to model temporal evolution of data:
- Neural network classifier:
used for vowel recognition in [Shinchi];
structure of the NN: 8 units-input layer;
20 units-hidden layer; 5 units-output layer.
- Probabilistic classifier: used for A-Z letters recognition in [Matthews]; decision rule:
-
Department of Informatics
Aristotle University of Thessaloniki
- Isolated letter recognition:
M=letter; M*=most plausible letter
OA=audio obs.; OV=video obs.