Research On The Separartion Algorithm Of Music Instruments And Singing Vioce

Nowadays, the fact that the quantity of music information has increased greatly, has led to an increasing demand of music information retrieval, identification and handling. However, background music and singing are related to each other in most cases, and the mutual interference between them has brought huge obstacles to music information processing. Music separation technology leads the research on these questions new exploration directions.This thesis mainly studies the music and singing separation. Based on statistical technology and computational auditory scene analysis technology, we have been thorough studying on NMF, music characteristics information analysis technology, acoustic model. On this basis, we have focused on the deep research of music repeating structure establishment and music background extraction. Finally, we proposed some improvements on the music and singing separation algorithm based on repeating. The main innovations are shown as the two following aspects:(1) The algorithm of music separation based on multi-repeating structureIn the separation algorithm based on music repeating, the building of repeating structure pattern has a core position. Although music is very complex, different ones have different repeating patterns. In the past algorithm, the extraction of repeating patterns is not able to reflect the music itself. In the paper, we use MFCC represented the repeating for the whole music structure, to put the fragment with consistent similarity together and build different repeating patterns for groups with different similarities. The experimental results show that the method can reflect the repeating nature of music itself.(2) The algorithm of music separation based on the analysis of the audio sourceIn the original repeating algorithm, if the rhythms of sound source are weak and slow, its repeating patterns are easy to be ignored. In order to better extract this audio type, we use the Harmonic Percussive Source Separation(HPSS). We use repeating patterns and HPSS to extract different audio sources in the background music respectively. The experimental results show that this measurement can greatly reduce the residual of the song, and improve the separation performance.