Multi-view semi-supervised discriminant analysis: A new approach to audio-visual person recognition