Comparisons of visual activity primitives for voice activity detection

Voice activity detection (VAD) with solely visual cues have usually performed by detecting lip motion, which is not always feasible. On the other hand, visual activity (e.g., head, hand or whole body motion) is also correlated with speech, and can be used for VAD. Convolutional Neural Networks (CNNs) have demonstrated significantly good results for many applications including visual activity-related tasks. It can be possible to exploit CNN's effectiveness to visual-VAD when whole body visual activity is used. The way visual activity is represented (called visual activity primitives) to be given to a CNN as input, might be important to perform an effective VAD. Some primitives might result in better detection and provide consistent VAD performance such that the detector works equally well for all speakers. This is investigated, for the first time, in this paper. Regarding that, we compare visual activity primitives quantitatively in terms of the overall performance and the standard deviation of the performance, and qualitatively by visualizing the discriminative image regions determined by CNN trained to identify VAD classes. We perform a data-driven VAD with a person-invariant training i.e., without using any labels or features of the test data. This is unlike the state-of-the-art (SOA), which realizes a person-specific VAD with hand-crafted features. Improved performances with much lower standard deviation as compared to SOA are demonstrated.