Representative melodies retrieval using digital signal processing of audio

Abstract

Extracting a melody to have representativeness of music so that contents-based music information retrieval system lets recently reduce response time about a search of a user, and be indexes do, and used this to a search. A way the existing study used MIDI data, and was proposed to extract a representative melody. However, there is a disadvantage limited to MIDI data. Therefore these papers propose the representative melody search way how application is possible to all audios file format as using a digital signal process technique of audio data. A representative melody is a section each audio data showing characteristics most conspicuous. This is the section where appeal is largest to persons listening to audio data and we was able to find from pieces having the maximum during wave forms of total data. Have got a node of audio data and tempos, and measure the frequency that high value (value more than 90% of a wave form highest of data) appears of a wave form from each nodes from node information got for representative melody searches. At this time a 8 node interval is representative melody scope of audio data from the scopes where high values was gathering most around of wave forms. Selected 1000 pieces of guns, and by 250 pieces by genres represented audio data of four kinds of Ballad, Dance, Hip-hop, Rock by experiment to verify effectiveness of a proposal way, and extracted a melody, and therefore showed Ballad 79.6%, Dance 83%, Hip-hop 89%, precision of Rock 62.8%. © 2006 IEEE.