||Content Based Music Similarity Estimation
||Hansen, Lars Kai (Intelligent Signal Processing, Department of Informatics and Mathematical Modeling, Technical University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark)
||Technical University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark
||This Bachelor project is an attempt to investigate the requirements for creating
a system capable of estimating similarity between music tracks. A change detection
algorithm analyses each music track, and divides the tracks into segments
flecting the musical contents. The purpose of a dynamic segmentation scheme
is that it should be possible to represent each segment by a single statistical
model, capable of modelling temporal correlations.
The algorithm used for segmenting the tracks is based on an algorithm developed
for detecting speaker changes in news stream broadcasts. An attempt to
optimise the algorithm for better performance with music was made. Furthermore
an attempt to modify the structure of the algorithm in order to obtain
better performance was made. A manually segmented training set was used to
optimise the algorithms and measure performance.
Each segment is represented by a Multivariate Autoregressive model, calculated
on the MEL-scaled cepstral coefficients. Investigations of the usability of
MAR for this particular application were carried out.
System performance, evaluated as the ability to compare music on a segment
basis, was measured on a data set of approximately 50 hours of music, divided
into 11 genres.
||Technical University of Denmark (DTU) : Kgs. Lyngby, Denmark
Creation date: 2009-06-16
Update date: 2010-08-25