Computational estimation of released decades for Japanese women idol music from musical audio

Abstract

This study realizes computational estimation of released decades for Japanese women idol (or diva) music from musical audio. Even though the released year of music is explicitly rebelled on the music, the feeling of decade, for example “1980’s, or old-fashioned music” is completely subjective so it is thought as difficult for computers to evaluate, yet people may feel it when listening to music. Such feeling is thought to be derived from musical arrangement, instruments, and so forth, yet previous study clarified the possibility to estimate based on acoustic feature of music. If such estimation is realized, it could contribute to enhance conventional recommendation system of music by giving a query: “newly released but old-fashion music.” This study conducted to establish explicit criteria for audio parameters by using logistic regression. Proposed system uses 25 types of acoustic parameters for each musical audio as training data set, and estimates single decades by simple logistic under a 10-fold cross validation scheme. As a result, estimation result of accuracy rate is 58.56% and kappa statistic is 0.48, being almost comparative to that of subjective evaluation. The estimation system is then implemented in a lap-top compute, which will be demonstrated at presentation.