A list of datasets for various music information retrieval tasks. Here, a “dataset” is loosely defined as a group of audio files accompanied by some amount of semantic information that can be useful when training and evaluating algorithms intended to automatically reproduce the information. Ideally, the datasets are freely available, or at least standardized in some way so that the results of an experiment on a given dataset can be reproduced. Whenever necessary, legal, and possible, datasets are hosted here. Similar lists of MIR datasets can be found here and here.
Here is a table giving basic details for each dataset discussed here:
|Isophonics Annotations||17+ albums||Yes||Yes||No||here|
|KSN Database||50 songs||Non-commercial use||Yes||No||here|
|Apple Jam Packs||40,000 files||No||Yes||Yes||here|
|Ableton EIC2||20,000 files||No||Yes||Yes||here|
|200 Drum Machines||9,000 files||Yes||Yes||Yes|
|Million Song Dataset/Echo Nest||1-30 million songs||Non-commercial use||Partially||Partially||here and here|
|DJ BPM Studio||80,000+ songs||Yes||No||No||here|
|Magnatagatune||26,000 songs||Non-commercial use||Partially||Yes||here|
|Last.fm||7 million+ songs||Non-commercial use||Yes||No||here|
The following is a table describing the MIR tasks each dataset could ostensibly be used for. Key: IR = Instrument recognition, TE = Tempo estimation, BT = Beat tracking, MT = Measure tracking, MS = Music segmentation, OD = Onset detection, KD = Key detection, TS = Time signature estimation, SS = Source separation, G = Genre, TM = Tags/Mood, SA = Similar artists (music recommendation), AT = Automatic transcription (including chord tracking), QH = Query by humming.
|Apple Jam Packs Loops||X||X||X||X||X||X||X||X||X|
|Apple Jam Packs One-shots||X|
|200 Drum Machines||X|
|Million Song Dataset/Echo Nest||X||X||X||X||X||X||X||X||X||X||X|
|DJ BPM Studio||X||X|
|Soundcloud Public Domain||X||X||X||X|
Each dataset is described in more detail below.
The Centre for Digital Music (C4DM) provides annotation data for a selection of albums by The Beatles, Queen, Carole King, Michael Jackson, and Zweieck. The annotations are done by hand by students at the C4DM and include chord, segmentation, key, and beat information. The data is accompanied by a rough confidence score for the annotation accuracy. .txt and .rdf representations are available freely here.
The University of Tokyo created this collection of chord and key annotations of 50 classical music pieces. The harmonic information is described using “KS-notation”. Both the dataset and this notation format is described in their paper.
Apple sells a series of "Jam Packs", which consist of loops (musical phrases) and one-shot samples (single notes) for use with their GarageBand and Logic software. Each Jam Pack focuses on a different class of musical instruments or styles. The loops are encoded in Apple's Core Audio Format (.caf), which allows for metadata to be included in the file directly. The Apple loops include hand-annotated tempo, beat, bar, tatum, time signature, genre, style, instrumentation, and key information. The extraction of this metadata from the .caf files is described here. The loops are also grouped together with similar loops intended to be used together, so by combining (and attempting to re-extract) separate similar loops would produce a useful dataset for source separation. The one-shot samples in the Jam Packs are in .aif format and have no metadata, but are grouped according to their instrument and as a result could be used for instrument recognition tasks.
Ableton distributes a collection of recordings of single notes from single instruments intended to be used as samples with their software instruments. Each file is a high-quality recording of many subsequent notes, and is stored as an encrypted .aif file and organized according to the instrument being recorded. In order to be used as one-shot samples, the files need to be separated note-by-note.
A collection of one-shot samples recorded from 200 different drum machines is available here. All files are uncompressed mono, 44.1 kHz PCM, and are organized according to the drum type (kick, cymbal, snare, tom), with about 1,000 samples of each drum type.
A large database of tempo values for about 80,000 popular songs is available at http://djbpmstudio.com/. The tempo estimations are almost certainly generated algorithmically and not by hand, but no information is available as to the details of the algorithm. Only tempo values are available, not the actual audio files. A .csv dump of the dataset from 4/13/2011 is available here.