SINGLE-CHANNEL MUSIC SOURCE SEPARATION BY HARMONIC STRUCTURE MODEL AND SUPPORT VECTOR MACHINE

Music source separation is an interesting but challenging task. Methods for separation are more difficult if the separation is under a single-channel condition. In principle, each musical instrument has its own characteristic, and a harmonic instrument can gener-ate its unique harmonic structure. Therefore, the harmonic structure model has been adopted for music source separation. In this paper, two mixed music sources are separated by harmonic structure model for each partitioned frame, and then the inter-frames are connected individually by the same harmonic structure. Second, from the average harmonic structure (AHS), these two instruments can be identified by comparing harmonic structures within the database. The frequency and amplitude of these two instruments from the database can be learned by support vector machine (SVM). With these two fea-tures, the inharmonic part of the mixed sources can be separated and classified. Finally, the harmonic part and inharmonic parts are added together for each frame, and the mixed music sources on a single channel can be separated. Experimental results show that the proposed method can improve the performance of signal to distortion ratio (SDR) under single channel music source separation. © 2022, Chinese Institute of Electrical Engineering. All rights reserved.