Deep temporal architecture for audiovisual speech recognition

The Audiovisual Speech Recognition (AVSR) is one of the applications of multimodal machine learning related to speech recognition, lipreading systems and video classification. In recent and related work, increasing efforts are made in Deep Neural Network (DNN) for AVSR, moreover some DNN models including Multimodal Deep Autoencoder, Multimodal Deep Belief Network and Multimodal Deep Boltzmann Machine perform well in experiments owing to the better generalization and nonlinear transformation. However, these DNN models have several disadvantages: (1) They mainly deal with modal fusion while ignoring temporal fusion. (2) Traditional methods fail to consider the connection among frames in the modal fusion. (3) These models aren’t end-to-end structure. We propose a deep temporal architecture, which has not only classical modal fusion, but temporal modal fusion and temporal fusion. Furthermore, the overfitting and learning with small size samples in the AVSR are also studied, so that we propose a set of useful training strategies. The experiments show the superiority of our model and necessity of the training strategies in three datasets: AVLetters, AVLetters2, AVDigits. In the end, we conclude the work.