A few months ago, when I had time to do some private development of music related software ( http://www.samsaffron.com/projects.html ) I started investigating how to get proper audio fingerprinting for my music collection. My problem was simple, I wanted to detect all the duplicate songs in my collection. The technology is out there (Philips and Moodlogic have proper implementations) however nothing is open (even relatable is closed and its performance is rather poor)

So…

During my investigation I managed to Google the current developments in computer audition. It turns out that it is pretty easy to get a list of songs that sound like another song (query by example). More interestingly, some cutting edge research is able to predict genres for a particular piece of music fairly accurately (at least as good as we can).

A very interesting read is:

“Beyond the Query-by-Example Paradigm: New Query Interfaces for Music Information Retrieval”
At: http://www-2.cs.cmu.edu/~gtzan/work/publications.html

Anyway,

The code for doing all of these very cool things is open and on source forge (see the project marsyas on sf.net)

I think it would be ultra cool to have the empeg (or karma or whatever) have the functionality that allows you to cycle through your various genres (not based on id3 tags) create automatic playlists based on an example song or tell the player to play “faster” or “slower” songs.

I think that as our music collections grow we need better ways of doing context aware browsing, handling 20gig collections is very different to handling 500gig ones.

Lastly I think it is pretty critical an open source song finger printer is developed. This technology will allow to correct meta data on files (eg. MusicBrainz) and more importantly collect user preference metadata (such as global rating or people who like this also like this).

What do the people on the group think?

When are we going to see a first open source context aware music browser? Will there ever be an open source audio fingerprinter?

Lastly I think it is pretty critical an open source song finger printer is developed. This technology will allow to correct meta data on files (eg. MusicBrainz) [...] Will there ever be an open source audio fingerprinter?

I haven't used it, but why doesn't Musicbrainz fit the bill?

Peter

Music brainz uses a closed source finger printer. Basically it calculatets one feature vector accross the first 30 seconds of a song and sends it to a central server which assigns it a trm id. The server code is closed, so the finger printing is closed. Also to get proper fingerprinting that performs well (Accross multiple bitrates etc.) you should calculate multiple feature vectors (see the philips paper at http://www.research.philips.com/InformationCenter/Global/FArticleSummary.asp?lNodeId=927 ) . The music brainz song matching performance is pretty poor.

Music brainz is an excellent source for clean meta data, its a shame they dont have a better fingerprinter, I suspect that if they had one we would all be using music brainz ...

Music brainz uses a closed source finger printer. Basically it calculatets one feature vector accross the first 30 seconds of a song and sends it to a central server which assigns it a trm id. The server code is closed, so the finger printing is closed.

I guess they must have changed that policy since you last looked (I don't remember the server or TRM generator being GPL last time I looked, either).

Peter

hmmm, just had a look at the source, its a light wrapper which contacts the trm server at music brainz to generate the trm id.

i dont think relatable will open the source to their core business unless they are about to go bust.

its a moot point anyway - the relatable implementaion of audio fingerprinting is very poor, even if they opened it it would be pretty useless. the key is comminicating much more data from the client to the fingerprinting server using the minimal amount of bytes. it is possible to generate a very accurate fingerprint using about 300 bytes. (client sends about 300 bytes to server and in return gets say a 16 byte unique identifier for the song)

the trm implementation sends about 500 bytes (but thats only one feature vector - you probably want to compress it all and send information describing a few hundred feature vectors.)

Right on.. I've been wanting to arrange my music by tempo and mood but haven't had the patience. Something like this would be really awesome for all kinds of applications.

hmmm, just had a look at the source, its a light wrapper which contacts the trm server at music brainz to generate the trm id.

Oops, sorry, so it is. Never mind then.

Peter

Didn't we have a BBS member, a couple years ago, announce that his company was opensourcing their music fingerprinter? Anyone have a link to that old thread?

Well, I don't know that any tuneprint people read this, but they were open sourcing theirs. Of course, tuneprint.com hasn't been touched in a while.

It was ClemsonJeep a long, long time ago.

-jk

Yup, that's the one, thanks.

It's not good. freetantrum.org doesn't exist anymore and etantrum.com has been registered by some Japanese guy and now points to a search engine style holding page.

The sourceforge project still exists here however. I've not looked at the code so I don't know if it would work without their servers though.