This paper examines the potential of high-level features extracted from symbolic musical representations in regards to musical classification. Twenty features are implemented and tested by using them to classify 225 MIDI files by genre. This system differs from previous automatic genre classification systems, which have focused on low-level features extracted from audio data. Files are classified into three parent genres and nine sub-genres, with average success rates of 84.8% for the former and 57.8% for the latter. Classification is performed by a novel configuration of feed-forward neural networks that independently classify files by parent genre and sub-genre and combine the results using weighted averages.
CITATION STYLE
McKay, C. (2004). Automatic Genre Classification as a Study of the Viability of High-Level Features for Music Classification. In International Computer Music Conference, ICMC Proceedings. International Computer Music Association.
Mendeley helps you to discover research relevant for your work.