Currently the following Musly music similarity methods are available:
As published in M. Mandel and D. Ellis: Song-level
features and support vector machines for music classification. (In the
proceedings of the 6th International Conference on Music Information Retrieval,
ISMIR, 2005). This measure computes the MFCC representation of each song to estimate
a single Gaussian. The Gaussians are compared with the Kullback-Leiber divergence
and make up the similarity.
Timbre (Musly default)
Tweaks the Mandel-Ellis similarity measure for best
results. We use a 25 MFCCs representation, the Jensen-Shannon divergence
and, most importantly, we normalize the similarities with Mutual Proximity
(D. Schnitzer et al.: Using
mutual proximity to improve
content-based audio similarity. In the proceedings of the 12th
International Society for Music Information Retrieval
Conference, ISMIR, 2011).
A commercial music similarity measure of high similarity quality and scalability. Can be used with millions of tracks. More information about it can be found on the OFAI website.
Evaluation of the Methods (January 2014)
We perform automatic music genre classification experiments to
evaluate Musly. A standardized leave-one-out nearest neighbor
classification experiment is performed with multiple standard music
collections used in MIR research. The results for Musly 0.1 are
displayed in the graph above. If available, we apply an artist filter.
All classification experiments can be easily repeated by using the
Musly command line application with the respective collection
- 1517 Artists: 3180 Songs, Genres: Soundtracks & More, R&B & Soul, Classical, Religious, Country, Easy Listening & Vocals, Folk, HipHop, Comedy & Spoken Word, New Age, Rock & Pop, Reggae, Children’s, Blues, Jazz, Alternative & Punk, Latin, World, Electronic & Dance.
The 1517 Artists collection was introduced by Klaus Seyerlehner. It consists
of 3180 freely available songs from 1517 artists. Each song is assigned to one of the 19 genres. It is notable that there is a non-music genre included: Comedy & Spoken Word. The dataset is available freely from: http://www.seyerlehner.info/.
- Ballroom: 698 Songs, Genres: Quickstep, Rumba, Samba, VienneseWaltz,
Waltz, Jive, ChaCha, Tango.
The Ballroom Dances collection was introduced by Gouyon. It is a collection with
music pieces assigned to classical ballroom dance genres. Each music piece has a duration of 30 s.
The collection was subsequently used for the ISMIR 2004 rhythm classification contest, where it
can be downloaded: http://mtg.upf.edu/ismir2004/contest/rhythmContest/.
- GTzan: 1000 Songs, Genres: country, rock, reggae, blues, disco, hiphop, jazz, pop, classical, metal.
The GTzan collection was assembled in 2002 by George Tzanetakis. It consists of 1000
audio tracks (each 30s length) evenly spread over 10 music genres. According to Tzanetakis, the files
are collected from a number of different sources (including CDs, radio and microphone recordings).
It is available freely from the Marsyas webpage: http://marsyas.info/download/data_sets.
- Homburg: 1886 Songs, Genres: funksoulrnb, alternative, rock, raphiphop, folkcountry, blues, electronic, jazz, pop.
The Homburg collection consists of 1886 songs. Each music piece is only 10s, which
is quite short. The short length is sometimes problematic for algorithms trying to estimate rhythm
features. The music snippets are available for download on the website of the Artificial Intelligence
Group of the University of Dortmund: http://www-ai.cs.uni-dortmund.de/audio.html.
- ISMIR 2004: 729 Songs, Genres: metal_punk, pop_rock, electronic, classical, world, jazz_blues.
The ISMIR 2004 collection is one of the most widely used music collections to evaluate music
similarity algorithms. The collection was created for the genre classification contest of the ISMIR
2004 conference where it still can be downloaded from the website: http://ismir2004.ismir.net/genre_contest/index.htm.
- Latin Music DB: 3637 Songs, Genres: Pagode, Forro, Bachata,
Tango, Merengue, Bolero, Salsa, Gaucha, Sertaneja, Axe.
The Latin Music (LMD) database was created by Carlos N. Silla. 3637 songs are classified into ten Latin dance genres. According to Silla, the songs were carefully selected by professionals. More information about the LMD can be found on the website of the authors: http://www.ppgia.pucpr.br/~silla/lmd/.