Challenges / ISMIS 2011 Contest: Music Information Retrieval/Music Instruments

Introduction

It is fairly easy to automatically recognize single instruments. However, when more than one instrument plays at the same time, the task becomes much more difficult. In this track, the goal is to build a model based on data collected for both single instruments and example mixtures in order to recognize pairs of instruments.

Data

Training data consists of two datasets: large one containing data for single instruments and much smaller one containing mixtures of pairs of different instruments. Both sets have the same attributes (the single instruments set contains some additional information, though).

The test set contains only data for mixtures. The pairs of instruments that are playing are to be predicted. Note that test and training sets contain different pairs of instruments (i.e. the pairs from the training set do not occur in the test set). Moreover, not all instruments from the training data must also occur in the test part. There also may be some instruments from the test set that only appear in the single instruments part of the training set.

The data consists of the following attributes:

• BandsCoef1-33, bandsCoefSum - Flatness coefficients
• MFCC1-13 - MFCC coefficients
• HamoPk1-28 - Harmonic peaks
• Prj1-33, prjmin, prjmax, prjsum, prjdis, prjstd - Spectrum projection coefficients
• SpecCentroid, specSpread, energy, log spectral centroid, log spectral spread, flux, rolloff, zerocrossing - The other acoustic spectral features
• LogAttackTime, temporalCentroid - Temporal features

Additionally, the single instruments set contains also:

• Frameid - Each frame is 40ms long signal
• Note - Pitch information
• Playmethod - One schema of musical instrument classifiation according to the way they are played
• Class1,class2 - Another schema of musical instrument classificaiton according to Hornbostel-saches

Both sets contain also the actual labels, i.e. the instruments. For the mixture data, there are two instruments, thus - two labels.

The datasets can be found in the Repository:

• Training dataset - single instruments and mixtures
• Test dataset

Note: you must be registered to this challenge in order to access the files.

Solutions and evaluation

Solution should be a text file containing one pair of labels per line. The names of the instruments in each line can be in any order and should be separated by a comma. Different number of labels in a line will result in an error. Labels are not case sensitive.

The baseline solution can be found in the Repository. It was generated by a modified Nearest Neighbour algorithm (selects two closest neighbours), without any data preprocessing.

The evaluation metric is modified accuracy:

• If no recognized instrument matches the actual ones, 0.0 score is assigned
• If only one instrument is correctly recognized, 0.5 is assigned
• If both instruments match the target ones, 1.0 is assigned

The final score is equal to arithmetic mean of individual scores.

Authors

Task 2, Music Instruments, was prepared by prof. Zbigniew Raś and dr Wenxin (Mike) Jiang from University of North Carolina, Charlotte, USA and Warsaw University of Technology, Poland.