Fully Convolutional Event-camera Voice Activity Detection Based on Event Intensity

The use of visual signals to detect vocally active duration is quite helpful when there is severe acoustic noise or even can be the only option if the audio channel is missing. There has been significant progress in video-based voice activity detection (VAD). On the other hand, while recently emerging event camera (EC) technology has demonstrated great benefits for applications in robotics, drones, autonomous vehicles, and mobile devices, including visual speech recognition topics, it has not been explored to be used as a vision-only VAD front-end. In this work, we propose an event intensity-based method by designing a fully convolutional network to efficiently realize an EC-VAD that segments vocally active duration. Efficiency is due to pooling the data over the mouth area reducing the dimensions by totally collapsing local spatial information as well as due to one-stage detection by a fully temporal convolutional network. Experimental evaluations show successful detection of voice activity with about 0.91 area under the receiver operating curve over a dataset including high speech content variability and different types of facial actions.