Automatic Classi cation of Musical Sounds in the context of Freesound

TitleAutomatic Classi cation of Musical Sounds in the context of Freesound
Publication TypeMaster Thesis
Year of Publication2014
AuthorsHespanhol, N.
AbstractFreesound.org is an online sound database with over 200.000 sounds now, which everyday serves the purpose of sound designers, composers and creative artists, providing them with Creative Commons license sounds for them to build musical compositions, interactive installations, build applications, movie soundtracks, etc. The need for optimized ways to serve musical content in this platform, as in others, is much needed, in order not to avoid leaving many sounds to be unexplored and never retrieved. What this work proposes is to develop a content based approach to classify sounds into a two level taxonomy. The first level consists of the classes Speech, Sound Effects, Soundscape, Mu- sic and Instrument. The second level, consists of musically meaningful concepts, such as Chord, Single Note, Melody, among others. Two different set of features are evaluated, MFCC and spectral features against low level features, and two different classification schemes too, at against hierarchical. For classification, K-Nearest Neighbour, Support Vector Machines and Decision Trees classifiers are used. Evidence is shown the MFCCs and spectral features combined with a hierarchical model works the best. The dataset used is also made available.
intranet