Voice activity detection by upper body motion analysis and unsupervised domain adaptation

We present a novel vision-based voice activity detection (VAD) method that relies only on automatic upper body motion (UBM) analysis. Traditionally, VAD is performed using audio features only, but the use of visual cues instead of audio can be desirable especially when audio is not available such as due to technical, ethical or legal issues. Psychology literature confirms that the way people move while speaking is different from while they are not speaking. This motivates us to claim that an effective representation of UBM can be used to detect 'Who is Speaking and When'. On the other hand, the way people move during their speech varies a lot from culture to culture, and even person to person in the same culture. This results in unrelated UBM representations, such that the distribution of training and test data becomes disparate. To overcome this, we combine stacked sparse autoencoders and simple subspace alignment methods while a classifier is jointly learned using the VAD labels of the training data only. This yields new domain invariant feature representations for training and test data, showing improved VAD results. Our approach is applicable to any person without requiring re-training. The tests applied on a publicly available real-life VAD dataset show better results as compared to the state-of-the-art video-only VAD methods. Moreover, the ablation study justifies the superiority of the proposed method and demonstrates the positive contribution of each component.