Multi-level Stereo Attention Model for Center Channel Extraction

In recent years, the spatial audio reproduction of digital media has become popular. Despite the demand for such spatial audio content, very little content is produced with multi-channel audio. Moreover, it is difficult to provide interactive services to users owing to the lack of object-based content. In this paper, we propose a center channel extraction method based on a multi-level convolutional neural network structure to generate object-based content. In addition, we present a novel stereo attention model which considers each channel's characteristics. By applying the proposed method to stereo audio content, we achieve better extraction performance than existing commercial application.