CLASSIFICATION OF AUDIO SIGNALS INTO MUSIC GENRES

V magistrskem delu predstavimo sistem, ki na mobilni platformi iOS samodejno klasificira glasbeni žanr zvočnih posnetkov na podlagi zajetih vrednosti značilnic. Sistem za posamezen posnetek izračuna vektor značilnic, ki opredeljujejo ritmične, tonske in energetske lastnosti posnetka. Na osnovi učne množice vektorjev z označenim glasbenim žanrom se sistem nauči značilnosti posameznih žanrov, na podlagi le teh pa kasneje opravlja klasifikacijo testnih posnetkov, ki nimajo označenega glasbenega žanra. Klasifikacijo smo izvedli z metodo podpornih vektorjev in pri tem na 1.000 testnih posnetkih dosegli 64 % natančnost ločevanja med naslednjimi desetimi žanri: blues, klasična glasba, country, disco, hip hop, jazz, metal, pop, reggae in rock.In this work, we present a mobile application for iOS, which determines musical genre of audio signal based on the features extracted from the signal. For each audio signal a feature vector that represents timbral, rhythmic and energetic properties of the signal is calculated. Based on a training set of feature vectors with labelled musical genre, system learns the characteristics of specific genre and uses this information to classify unlabelled audio signals into multiple musical genres. Classification is performed by support vector machine unsupervised machine learning algorithm. When tested on 1,000 audio signals with 10 different genres, the implemented classifier yielded accuracy of 64 %