This page contains updated results and errata of the following paper:
Holzapfel, A., Krebs, F., and Srinivasamurthy, A.
Tracking the 'odd': meter inference in a culturally diverse music corpus
Proceedings of the 15th International Society for Music Information Retrieval Conference (ISMIR), Taipei, Taiwan, 2014.
Unfortunately, the original version of the paper contains some errata, which we list and resolve in the following:
The code to reproduce the above mentioned results will be made public available in October 2015.
with R=1 are substantially higher as reported because they were generated with a degenerate transition matrix. This also
means, that using one rhythmic pattern per rhythm class is enough to achieve a good performance. Therefore, the unified HMM,
which incorporates all styles, only has to models R=8 rhythmic pattern states, instead of R=17 in the original paper. We re-ran the experiments after having resolved the bug.
you find the corrected results. Note that the results for R>1 can still be different from the original paper, as we re-trained also the observation model and this implies fitting Gaussian Mixture Models that rely on (random) initialisation.
- The original experiments were performed with 1536 positions per whole note instead of 1600 as reported in the paper.
Here you find the results with 1600 positions per whole note.
- The tempo limits of the original models were determined by taking the median beat interval of each song and then taking the maximum and minimum tempo across all songs of a certain rhythm class. Recently, we found that only throwing away the slowest and fastest 5% and then finding the min/max gives wider tempo ranges and seems to work slightly better.
Here you find the results with 1600 positions per whole note and tempo ranges determined using the 5th percentile.
last edited by fk at
Jul 15, 2015