Recovering audio-visual synchronization is an important task in the field of visual speech processing.
In this paper, we present a multi-modal regression model that uses a convolutional neural network (CNN) for recovering audio-visual synchronization of single-person speech videos. The proposed model takes audio and visual features of multiple frames as the input and predicts a drifted frame number of the audio-visual pair which we input. We treat this synchronization task as a regression problem. Thus, the model does not need to search with a sliding window which would increase the computational cost. Experimental results show that the proposed method outperforms other baseline methods for recovered accuracy and computational cost.