Muusika žanri avastamine kasutades Naïve Bayes klassifikaatori

Abstract

Tänapäeval hoitakse muusikat peamiselt digitaalvormis. Muusika faile on nii palju, et neid tuleb kuidagi klassi tseerida. Kompositsioone on võimalik grupeerida žanrideks heliseva iseloomu järgi. Töö baseerub G. Tzanetakis, G. Essl ja P. Cooki artiklil [GT07], mis käsitleb muusika žanri avastamisalgoritmi loomise temaatikat. Peamiseks ideeks on esitada muusika faile numbrilises formaadis ja võta välja sellest informatsioonist mõned tunnused, mis kirjeldaksid muusika helisust ja aitaks žanrideks klassi tseerimisel. Esiteks, me realiseerisime artiklis pakutud tunnuste arvutamismeetodit ja hinnasime nende töö meie andmestikul, mis koosneb 300 muusika failidest iga žanri (klassika, pop, punk, rap/hip-hop, rokk ja trance) esitavad 50 kompositsiooni. Seejärel valisime klassi kaatorit ja pakkusime välja oma ideed. Tulemuseks me saime 13-elemendilist tunnuste vektori, mis pooleli koosneb baseeruvas artiklis esitatud tunnustest ja pooleli meie ideedest. Tunnuste vektor koos valitud algoritmiga võimaldavad klassi tseerida 6 žanri lood 61,6% täpsusega, mis on peaaegu neli korda parem kui juhuslik. Peale seda, tulemus on 5% parem kui see mida said baseeruva artikli autorid.

Music in digital form is widely spread nowadays. Musical pieces can be grouped into genres according to their sounding characteristics. For most people classi cation of a given composition is a reasonably easy task. Automating this classi cation process is, however, not so trivial. Fortunately, we can state the task of digital music classi cation as a machine learning problem. We consider a set of musical compositions with manually assigned genres as a training set and use it to devise an automatic genre classifier. The traditional approach requires us rstly to extract meaningful features from the acoustic signals, and then apply a general-purpose machine learning algorithm on the transformed data. For feature extraction we used the ideas proposed in the paper by G. Tzanetakis,G. Essl and P. Cook. In their research the authors propose some features that represent music surface and rhythmic structure of audio signals. We reevaluated these features on our own dataset and suggested some additions. Finally, we selected the best performing feature set combining both the original features and our proposed additions. As long as features are selected properly, the choice of the algorithm is largely irrelevant. In this work we selected the Naïve Bayes algorithm due to its conceptual simplicity and eficiency. as a result of our work, we constructed new feature set of 13 elements that classifi es music of six genres with the accuracy of 61,6% that is almost four times better than random.