The first step in rhythm analysis is data reduction.  The goal is to eliminate high-frequency information (pitch) so as to study low-frequency information (rhythm).  Our data reduction is based on methods used by Sethares and Sheirer in the following articles.  We are especially indebted to William Sethares for sharing his Matlab data reduction routines.

• William A. Sethares and Thomas W. Staley,  ``Meter and Periodicity in Musical Performance,''  preprint, 2001.
• William A. Sethares and Thomas W. Staley,  ``Periodicity Transforms,''  Transactions on Signal Processing, vol. 47, no. 11, November 1999.
• Eric D. Scheirer, ``Tempo and Beat Analysis of Acoustic Musical Signals,''   Journal of the Acoustical Society of America, vol. 103, no. 1, January 1998.

An Example

Here is an example of how the data reduction works.  The original track is the first 10 seconds of ``Hristianova Kopanitsa'' from the CD Balkanology by Ivo Papasov and his Orchestra (Hannibal HNCD 1363, 1991).

One method of data reduction is to filter out the high frequencies (anything perceived as pitch, say 20 Hz and above).  The resulting low-frequency envelope is applied to white noise to produce a rather unpleasant sound.  The rhythm is barely discernible.

Scheirer's article explains how the sound can be first separated into 21 bands of roughly half an octave.  A rhythm vector is formed for each band and used as an envelope for a typical sound of that frequency range.  The bands are numbered from low to high.
 

Finally, let's hear the original track combined with the rhythm track above.