• wav format: 22,050 kHz, mono, 16 bit.
• text format: each row contains an onset time in seconds.

Each file is a Format 1 MIDI file that contains two tracks:

• The so called master track or track 'zero' where the following MIDI metaevents are optionally encoded:
  • Track Name (often contains the song name)
  • Time Signature
  • Tempo
  • Key Signature
• The melody track (track 'one') containing the melody line. This melody line it is not guaranted to be fully monophonic, althought it is so in most files.

Most files have the melody concatenated three times, and one or two silence bars at the beginning. Some melody tracks are polyphonic (double octaves, violin parts in a single track, etc...). If you desperately need the tracks to be monophonic, you can reduce them to monophonic tracks using the 'smf2txt/txt2smf toolchain' avalaible at the following address: http://grfia.dlsi.ua.es/gen.php?id=resources

A command line example that reduces all tracks of a MIDI file to monophonic tracks:

$ smf2txt -p 1 polyphonic.mid | txt2smf -f monophonic.mid

The '-p' option argument means whether you want to preserve the top (1) or bottom (2) line. Be aware, however, that the result may be not exactly what you would expect. A slight, perceptually not significant, overlapping of two otherwise consecutive notes could result in the second note being discarded.

• As explain in Ponce de León and Iñesta (2007), a ten-fold cross validation scheme is proposed for evaluation. Average accuracy and standard deviation are proposed measures. For multiple results comparison, an ANOVA test with Bonferroni correction is suggested.
• In the paper above, the authors provide extensive results about classification done by extracting fixed length segments from melody tracks by means of a sliding window procedure. An exploration of varying both the length of the window and the overlap between consecutive segments was performed, suggesting values for the window length above 30 bars for obtaining good classification results.
• Other publications where this corpus has been used are Pérez-Sancho et al. (2006), Ponce de León et al. (2006) and Moreno-Seco et al. (2006)

4 files are provided, three of them for particular music genres and a forth one merging the other three, in order to study the possible specificities of melody in different music genres:

• CL200: classical music. 703 instances (16 unlabeled)
• JZ200: jazz. 769 instances (11 unlabeled)
• KR200: karaoke (mainly pop music). 1668 instances (338 unlabeled)
• MEL200.arff: the merge of the 3 files.

In Rizo et al. (2006) the names above were used for the training sets, but in Ponce de León et al. (2007) the merged file was named "SMALL".

3140 samples are provided, 365 unlabeled, so 2775 labeled samples are provided according to the description in Ponce de León et al. (2007), although slightly different from that reported in Rizo et al. (2006). The unlabeled samples belong to the files where no melody tracks were easy to find.

• Name of the MIDI file the track was extrated from.
• Path name, corresponding to the genre of the particular MIDI file.
• Song number (starting form zero for each genre) the track belongs to.
• Instance number, starting from 1 for each MIDI file.

The problem is stated at two different levels:

1. Given a particular MIDI track (extracted from a multitrack MIDI file), does it contain a melody? (two class problem fully labelled, just partitions for cross-validation need to be done)
2. Given a multitrack MIDI file, which track contains the melody? (no melody is a possible situation). (song number metadata is needed to be taken into account for classfying the melody track among those belonging to the same song)

• As explained in Rizo et al. (2006), for the task a) a 10-fold cross-validation is proposed for evaluation. Success rate, precision, recall, and F-measure are suggested. The ten folds were done automatically by weka (seed = 1).
• For the task b) leave-one-song-out was utilized due to the limited number of samples. The same performance estimators are proposed according to the instructions in Rizo et al. (2006)

• CL200.arff (Classical files)
• JZ200.arff (Jazz files)
• KR200.arff (Karaoke-pop files)
• ALL200.arff (whole set)