Two different types of training have been performed in order to check the efficiency of the audio only and the audio-visual speech recognizer. We first trained the HMMs using clear acoustic data, whatever the S/N in the test condition . Second, we trained the HMMs onto corrupted acoustic data by adding a white noise whose intensity N corresponded to that of each S/N condition during the test.
Esprit Project 8579/MIAMI (Schomaker et al., '95)