A consideration on improvement of music removal accuracy from environmental sounds

Recently, the demand for video content has been increasing due to the growing number of diverse social networking users. On the other hand, it is necessary to be aware of copyrights when uploading videos. If copyrighted music is included in the video, the music must be removed from the video. There is a method using audio enhancement techniques to remove music from video, but this method removes the environmental sounds in the video, which reduces the sense of realism of the video. Therefore, a method has proposed to remove music only from multiple mixtures of sounds using machine learning. However, the music removal accuracy of the method is not sufficient. In contrast, augmenting the training data is expected to improve accuracy, however, there is no suitable augmentation methods for music data. Therefore, we propose applying pitch shifting as a data augmentation method for music data to improve the music removal accuracy. Then, we conduct a simulation using both the original data without augmentation and the data augmented with our proposed method, and compare the music removal accuracy. As a result, we confirm that music is removed with higher accuracy when learning with augmented data. Therefore, the data augmentation method proposed in this paper is effective.