GTZAN music genre dataset [closed]

Question

I want to test my music genre algorithm in the public dataset to compare with other preexisted algorithms. In case of MIREX, all the data are not available. I found that GTZAN dataset are available in a link(marsyas.­info/­download/­data_sets) But, it is not available to me now.

Do you know how can I get this data? Because I use other feature than MFCC, so I need genre annotation as well as music.

Thanks in advance.

Answer 1

EDIT: now, it seems, that Marsyas page is hosted at marsyasweb.appspot.com and you can find links to GTZAN database in data sets subpage.

I was also looking for GTZAN dataset for my university project and I found out that http://marsyas.info is down, so I used google web cache for marsyas.info/download/data_sets (check it for more info). Luckily, GTZAN data sets are hosted at http://opihi.cs.uvic.ca and you can download them!

But be aware of licensing before using datasets (info from cached marsyas download page):

This dataset was used for the well known paper in genre classification " Musical genre classification of audio signals " by G. Tzanetakis and P. Cook in IEEE Transactions on Audio and Speech Processing 2002.

Unfortunately the database was collected gradually and very early on in my research so I have no titles (and obviously no copyright permission etc). The files were collected in 2000-2001 from a variety of sources including personal CDs, radio, microphone recordings, in order to represent a variety of recording conditions. Nevetheless I have been providing it to researchers upon request mainly for comparison purposes etc. Please contact George Tzanetakis ([email protected]) if you intend to publish experimental results using this dataset.

The dataset consists of 1000 audio tracks each 30 seconds long. It contains 10 genres, each represented by 100 tracks. The tracks are all 22050Hz Mono 16-bit audio files in .wav format.

Maybe you will be also interested in other datasets such as Magnatagatune - http://tagatune.org/Magnatagatune.html.

Answer 2

I realize this has been answered a long time ago, but I felt it necessary to update this thread with my answer.

Depending on your genre classification implementation, which I am assuming is a content-based one (spectrograms are popularly used in deep learning approaches, however, you do mention MFCCs), you may find it easy to find datasets with features already extracted from the audio. This is probably the easiest way to directly compare your results against other approaches.

For this you can use the Latin Music Database, which you can get from here, the Free Music Archive, which you can get from here, the Greek Music or Audio Dataset, which you can get from here, the CAL10k/CAL500 dataset, which you can get from here, and pretty much whatever other datasets you can scrape from this website.

Otherwise, what I would recommend doing is to extract these features yourself, using some feature extraction libraries or toolboxes, that way the GTZAN, or Ballroom datasets can be used (or whatever else has audio files). betatester07's link is still working at the moment too!

The feature extraction methods I find easiest to use are jAudio, or LibROSA, for Java and Python respectively. With jAudio you can also call it from the command line, or just use the gui, it also accepts batchfiles so you can process a whole bunch of audio files at once. Using these extraction methods, you can get all of the features to match those standard ones if they rear their ugly heads in a paper that you are trying to implement or compare against.

Be careful using these standard "benchmark" datasets, as there are still faults with them. For this you want to check out Sturm's work.

Hope that helps :)