Improvement of Music Removal Accuracy from Environmental Sounds

Recently, by proliferating video-sharing sites and social networking services, the posting-videos has increased on the internet. However, it needs to be careful when the video contains copyrighted music. One method of music removal is the method to use of speech enhancement. However, this method removes environmental sounds other than music, and the sense of realism for the video is lost. Therefore, the method is proposed to remove only music from speech data containing various sounds by using machine learning. However, the removal accuracy is not sufficient. Then, in this paper, to improve music removal accuracy, we propose a method to feed for the learning model to augmented the existing data. The augmenting is assumed to keep the length of the data in order to unify the data length given to the learning model. Specifically, we use the pitch shift method which changes only the pitch of the sound while maintaining the length of the data. Then, we compare music removal precision on the non-augmentation data and the augmentation data. As a result, it is confirmed that the augmented data is removing music more accurately than the non-augmentation data. Thus, in this paper, the augmentation method proposed is effective.