A deep neural network for audio-visual person recognition

This paper presents applications of special types of deep neural networks (DNNs) for audio-visual biometrics. A common example is the DBN-DNN that uses the generative weights of deep belief networks (DBNs) to initialize the feature detecting layers of deterministic feed forward DNNs. In this paper, we propose the DBM-DNN that uses the generative weights of deep Boltzmann machines (DBMs) for initialization of DNNs. Then, a softmax layer is added on top and the DNNs are trained discriminatively. Our experimental results show that lower error rates can be achieved using the DBM-DNN compared to the support vector machine (SVM), linear regression-based classifier (LRC) and the DBN-DNN. Experiments were carried out on two publicly available audio-visual datasets: the VidTIMIT and MOBIO.