The preparatory data consisted of 7 instances that have been given to the participants together with their tempo values in order to compare whether algorithms yield the same outputs when ran in participants' labs and on our machines, and to check proper formatting of algorithm outputs. No training data was provided.

The test data consisted of 3199 tempo-annotated instances. The instances range from 2 to 30 seconds, and from 24 to 242 BPM (only 15 have a tempo less than 50 BPM). They all have approximately constant tempi, the format is the same for all: mono, .wav, 44100Hz sampling frequency, 16 bit resolution. The total duration of the test set is approximately 45140 s (i.e. 12 h 36 mn). This data was not available to participants.


If you want to refer to this contest in a publication, or use the data provided on these web pages, please cite the related article. The full citation is "An experimental comparison of audio tempo induction algorithms" Gouyon F., Klapuri A., Dixon S., Alonso M., Tzanetakis G., Uhle C. and Cano P., IEEE Transactions on Speech and Audio Processing, 14(5) 2006

BPM statistics of the three datasets are available here (may be useful e.g. for setting prior probabilities in Bayesian approaches, thanks to an anonymous reviewer for this suggestion).

Fabien Gouyon 2006-03-17