Up: Self-Organizing Maps for Content-Based
We presented an approach to automatically organize music by content, i.e. based on its genre and sound characteristics.
The Self-Organizing Map, a prominent unsupervised neural network, is used to cluster feature vectors representing the musical sound based on frequency spectra.
In a first step, music segments are organized to obtain a fine-grained representation of segment-wise similarities, based upon which a clustering of the complete songs can be obtained.
With this approach similar pieces of music are found in neighboring regions of the map.
While the presented approach provides a good organization of music on the two-dimensional map, further improvements may be gained by capturing additional features during the vector creation process.
These features may include beat information as well as representations capturing the dynamics of the various frequency bands.
Furthermore, weighting functions may be used to assign higher importance to specific frequency bands.