22 February 2008, Beat Induction: Finding the Meter, Patrick de Kok, Gijs Kruitbosch and Nadya Peek
When listening to music, humans can quickly find the beat to tap along to. In the field of Beat Induction, researchers attempt to simulate how the meter of a fragment of music is found. We have proposed a new computational model for finding the rhythm of a musical fragment. Our implementation uses a statistically oriented beat induction method, which harnesses a naive Bayes classifier to categorize beats and time signatures. Using a training corpus of folk songs, we match note placement and duration to different beats and time signatures using the WEKA machine learning toolkit. We evaluated several conditions for classification: to test how soon (that is, how much of a music fragment is needed before) our system can generally correctly classify a beat; how soon our system can find a time signature; and finally to deal with possible difficulties with upbeats. When using our system to classify data, we find it performs quite well in distinguishing between ternary and binary time signatures, but does not do very well in classifying precise time signatures. In our talk, we will offer some of our explanations for this and show some of the other results we have obtained using our classifier. We will also discuss possible improvements and applications of our learning system.
This work has been carried out in the context of the BSc Artificial Intelligence Honours Programme at the Universiteit van Amsterdam, within the Music Cognition Group (http://cf.hum.uva.nl/mmm/).
For more information, contact Leigh Smith (lsmith at science.uva.nl).