Investigating the Visual Lombard Effect with Gabor Based Features

The Lombard Effect shows that speakers increase their vocal effort in the presence of noise, and research into acoustic speech, has demonstrated varying effects, depending on the noise level and speaker, with several differences, including timing and vocal effort. Research also identified several differences, including between gender, and noise type. However, most research has focused on the audio domain, with very limited focus on the visual effect. This paper presents a detailed study of the visual Lombard Effect, using a pilot Lombard Speech corpus developed for our needs, and a recently developed Gabor based lip feature extraction approach. Using Kernel Density Estimation, we identify clear differences between genders, and also show that speakers handle different noise types differently.