Since speech is both auditory and visual, visual cues could compensate to a certain extent the deficiency of auditory ones, in order to improve man-machine communication and telecommunication tools. This paper deals with a noise reduction technique based on speech enhancement with adaptive filters estimated from the speaker's lip pattern. We first present the selected filtering techniques, and then the tool we used to predict the filter pattern from the lip shape. The whole noise reduction system is implemented in the context of stationary vowels including a first kick into the problem of non-visible gestures. The results of perceptual tests are presented in order to quantify the performances of the system. These results are quite promising.
Bibliographic reference. Girin, L. / Feng, Gang / Schwartz, Jean-Luc (1995): "Noisy speech enhancement with filters estimated from the speaker's lips", In EUROSPEECH-1995, 1559-1562.