5th International Conference on Spoken Language Processing

Sydney, Australia
November 30 - December 4, 1998

A Fast Method of Producing Talking Head Mouth Shapes from Real Speech

Andrew P. Breen, O. Gloaguen, P. Stern

BT Labs, Ecole Nationale Superieure des Telecommunications de Bretagne, France

The subject of computer generated virtual characters is a diverse and rapidly developing field, with a wide variety of applications in industries as varied as entertainment, education and advertising. Many of these applications require or would be greatly enhanced by having the virtual characters speak with the recorded voice of a real person. Such an ability is particularly useful in applications where users are interacting via avatars in real time in a virtual world. There are three basic problems which need to be addressed when developing an interface which has this functionality. *) The process must be capable of animating mouth shapes in real time. *) The process should not mouth extraneous sounds such as music, doors slamming etc. To do so would diminish the effectiveness of the illusion. *) The mouth shapes produced by the avatar should approximate that of the speaker. This paper describes a series of experiments which attempt to address each of the points outlined above. The experimental procedures are based around a real time low computation approach which relies on a particular variety of neural network known as the Single Layer Look Up Perceptron (SLLUP).

Full Paper

Bibliographic reference.  Breen, Andrew P. / Gloaguen, O. / Stern, P. (1998): "A fast method of producing talking head mouth shapes from real speech", In ICSLP-1998, paper 0390.