Multiscale approaches to music audio feature learning

Content-based music information retrieval tasks are typically solved with a two-stage approach: features are extracted from music audio signals, and are then used as input to a regressor or classifier. These features can be engineered or learned from data. Although the former approach was dominant in the past, feature learning has started to receive more attention from the MIR community in recent years. Recent results in feature learning indicate that simple algorithms such as K-means can be very effective, sometimes surpassing more complicated approaches based on restricted Boltzmann machines, autoencoders or sparse coding. Furthermore, there has been increased interest in multiscale representations of music audio recently. Such representations are more versatile because music audio exhibits structure on multiple timescales, which are relevant for different MIR tasks to varying degrees. We develop and compare three approaches to multiscale audio feature learning using the spherical K-means algorithm. We evaluate them in an automatic tagging task and a similarity metric learning task on the Magnatagatune dataset.

ISBN:9780615900650

Publication year:2013

Accessibility:Open