Pandora: Music discovery Service

I’ve just found the site Relatively recent, the web site wants to be a music discoverer by the listening of a radio fitting to our music tastes. The user has to create a radio station in entering an author’s or song’s name that he appreciates and the radio automatically plays songs which should be similar to your initial choice. The user can influence the futur choices by voting on the previous songs (I like/ I don’t like).

NB: Currently registration is only allowed for US citizen because of a problem of music license. A USA ZIP Code is needed for registration.

How does it work: the Music Genome Project.

According to the web site

Pandora’s music discovery services are based on The Music Genome Project™, the most sophisticated taxonomy of musical information ever collected. It represents years of analysis by our trained team of musicologists, and spans sixty years of music (10, 000 artists). Pandora does not use machine-listening or other forms of automated data extraction. Each song in the Music Genome Project is analyzed using up to 400 distinct musical characteristics by a trained music analyst. These attributes capture not only the musical identity of a song, but also the many significant qualities that are relevant to understanding the musical preferences of listeners.

Together our team of thirty musician-analysts have been listening to music, one song at a time, studying and collecting literally hundreds of musical details on every song. It takes 20-30 minutes per song to capture all of the little details that give each recording its magical sound melody, harmony, instrumentation, rhythm, vocals, lyrics … .

NB: in wikipedia , we can see the list of the attributes use in the Music Genome Project [3].

Event If we understand the methodology to index the musics , we have no information about the algorithm that choose the music to play. Nor a clear notion of the music similarity (a basic euclidian distance among the music’s attributes, or as we could imagine a formula taking in account a more complex relation among the attributes). Mistery…

After 2 hours of listening

Personnaly I’m a bit disappointed but it works: I discovered one or two songs. The problem is the algorithm of selection of musics that choose too quickly unrelated with the first choice, far from the initial song/artist you entered. Maybe it’s wished by the team, helping the user to discover new songs, like the good idea to play in preference unpopular songs. Idem, Voting don’t seem to have a clear impact on the futur choices.

For the moment, the system of music selection is not perfect but I’m pretty sure the team will improve it in a few time. Moreover I hope their business model is robust because paying the music license to be authorized to play freely and entirely the songs for the users should cost a lot of money. Good luck guys..

Too bad I’m not in the project: That kind of problem of similarity in music and of music selection to help the user to discover music is a very exacting research project 😉