Dance Music Segmentation
Segmenting electronic dance music streams based on self-similarity
- Tim Scarfe (http://www.developer-x.com)
- Wouter M. Koolen (http://wouterkoolen.info/)
- Yuri Kalnishkan (http://www.clrc.rhul.ac.uk/people/yura/)
We present an unsupervised, deterministic algorithm for segmenting DJ-mixed Electronic Dance Music (EDM) streams (for example; pod-casts, radio shows, live events) into their respective tracks. We attempt to reconstruct boundaries as close as possible to what a human domain expert would create in respect of the same task. The goal of DJ-mixing is to render track boundaries effectively invisible from the standpoint of human perception which makes the problem difficult.
We use dynamic programming to optimally segment a cost matrix derived from a similarity matrix. The similarity matrix is based on the cosines of a time series of kernel-transformed Fourier based features designed with this domain in mind. Our method is applied to EDM streams. Its formulation incorporates long-term self similarity as a first class concept combined with dynamic programming and it is qualitatively assessed on a large corpus of long streams that have been hand labelled by a domain expert.
In laymans terms, the purpose of this software is to automatically generate a cue sheet in the situation that you have downloaded a radio show like A State of Trance and you have only the track list. Web sites already exist where humans manually create cue sheets i.e. http://cuenation.com/ The problem is you have to rely on them (you can't automate the process).
See http://www.developer-x.com/papers/segmentationextended for more information and associated papers.
On this github project we supply the working code with a sample test set in ./Matlab/examples. The project is written in Matlab although some helper functions i.e. for pre-processing the dataset, extracting cue sheet times etc are included as part of a Visual Studio project.
Simply execute execute_show(1) to see it work.
Or for more cool stuff:
s = 1; % change to 1,2,...,6 for the github test set shows
execute_show( s, config_getbest(1,1), config_getdefaultsegcalculation, config_getbestnoveltyconfig );
I have included training set binaries:
- A State of Trance (With Armin van Buuren) 453 + 462
- Trance Around World (with Above and Beyond) 364 + 372
- Magic Island (with Roger Shah) 98 + 112
This is free software. Feel free to fork or contribute. It would be particularly useful if somebody is interested in creating an implementation of this in a more serious language like C, Java, C# etc. I would be happy to collaborate. Get in touch.
I would like to personally thank Dennis Goncharov for providing me with a dataset complete with cues for my research and also Mikael Lindgren from CueNation. Both provided me with insight on how they find the optimal track indices. The dataset I have from Dennis is quite large but sharable on Google Drive.
Some information from Denis about how he makes the cuesheets -- https://docs.google.com/a/developer-x.com/document/d/1vgr7x0nwzoWrVewLFyRS9xO0ehZowdeyKdCGDvSfQN8/edit
