C[omp]ute

Welcome to my blog, which was once a mailing list of the same name and is still generated by mail. Please reply via the "comment" links.

Always interested in offers/projects/new ideas. Eclectic experience in fields like: numerical computing; Python web; Java enterprise; functional languages; GPGPU; SQL databases; etc. Based in Santiago, Chile; telecommute worldwide. CV; email.

Personal Projects

Choochoo Training Diary

Last 100 entries

Identifying Artists with EchoNest

From: andrew cooke <andrew@...>

Date: Wed, 6 Feb 2013 18:55:09 -0300

I am trying to identify artists using EchoNest's API
- http://developer.echonest.com/

The simplest way is to simply call artists/search, but I am not 100% confident
in the mp3 metadata that I am using.  So I wanted something more robust, if
possible.

My data (music) is grouped into albums, but unfortunately EchoNest don't
expose albums in their interface.  And I have enough data that using
fingerprinting seemed like it would be inefficient.  So I have ended up with
the following heuristic, which seems to work well:

 * Note whether an album is single-artist (ie if the same artist is on
   all tracks).

 * Try to find the tracks on the album using song/search, specifying both
   artist name and song title.

   * For single-artist albums, you can set the artist for the entire
     album on the first successful track identification.

   * For multi-artist albums you need to take each track individually
     (I could be smarter, as I don't handle the case where there are
     multiple artists, but each artist has several tracks).

 * If any artists remain unidentified, use artist/search.

 * For single-artist albums, where the artist is *still* unidentified,
   use song/search and request the top 15 results for each track name.
   For each result, filter so that any artists returned are mentioned just
   once.  Then combine results and take the most popular artist (if
   mentioned more than once).

The reasons this works well are:

 * The first step uses track info to get a more reliable artist ID.

 * The last step identified artists when the mp3 metadata is confused 
   (eg Jazz albums where the artist is a long list of names).

Andrew

Example Voting Results

From: andrew cooke <andrew@...>

Date: Wed, 6 Feb 2013 19:36:35 -0300

Both good and bad, for your entertainment.  ID3 first, followed by the
identification (in the order generated by my file scan).

 Safar Barlik -> Fairuz
   Not sure what happened here - I don't know wnough about Egyptian music
   to make much sense of this one.

 Duke Ellington, Charles Mingus, Max Roach -> Duke Ellington
 Mulatu Astatke and his Ethiopian Quintet -> Mulatu Astatke
 Bappi, Uthup, Mithun -> Bappi Lahiri
 Jaco Pastorius, Pat Metheny, Bruce Ditmas & Paul Bley -> Paul Bley
   All these are working nicely.  The final aim is to link related
   music, so this level of accurcay is fine.

 Robert Fripp, Andrew Keeling & David Singleton -> Sarah Brightman
   Surreal.  That's going to produce some "interesting" playlists.

 2Sentidos -> Pereza
   Wrong - perhaps suffering from poor Spanish language support in the
   EchoNest database?

 Bill Frisell, Ron Carter & Paul Motian -> Bill Frisell
   Another success.

 Panavizion series -> Mychael Danna
 Hammeriver -> Michael Jackson
   Depressing.

 Keren Ann & Erik Truffaz -> Keren Ann
 Jan Garbarek Brahem Hussain -> Jan Garbarek
   Seems to work best with lists of jazz artists...

 Tape & Bill Wells -> Johann Pachelbel
   What can you expect when all tracks are called "Fugue"?!

 Edu Lobo & Nara Leao & Tamba Trio -> Elis Regina
  Again, I wonder if EchoNest is lacking foreign language tracks?
  These are both brazillian.

 Edie Brickell & New Bohemians -> Edie Brickell
 Ricardo Modegro y Paco de Lucia -> Ricardo Modrego

So, that's 9/16 - about 50% success.  Which isn't bad when the vast majority
of albums are handled by the simpler heuristiucs.

Andrew

Comment on this post