Musical fingerprinting launched by Last.FM
In June I wrote about technology that is analysing music for record companies, and suggested that it would also be really useful in search and recommendation for end-users like us. Last.FM have just launched a version of their software that uses the same principals for track recognition, and the details are at Blognation. This isn’t the best application for the technology.
What I’d like is for Last.FM to create an audio fingerprint for each track they stream to users, and then use that to create better recommendations. When users “love” or “ban” tracks, Last.FM could use the preferences and the fingerprints to try to find music with similar characteristics, to play more tracks like the loved ones, and less tracks like the banned ones.
That would help me discover music I had no chance of finding before, as a natural extension to the collaborative filtering they currently use. Playing me tracks that other people love, when those people love the same tracks as me, is helpful, but it needs huge quantities of data to be really useful, and isn’t therefore much help with the long tail. If a track only gets played 5 times a day, the fingerprint will be a lot more interesting than the ratings, and recommending similar tracks with similar fingerprints would be great.
Instead, what they’re doing is using the fingerprint to identify a track, then picking up the user’s tagging data to improve their own. Users play their own music (in itunes, for example) and Last.FM fingerprint it. They then upload the user’s tags from their own music, and apply them to music on their database. Effectively, they’re using the technology to enhance their own music tagging data.
A nice idea, but not the one I’d really like them to build.
What you’re talking about is called feature extraction, not fingerprinting. Fingerprints generated by most systems bear no direct resemblance to how the music sounds.
We are already working on feature extraction behind the scenes to improve our recommendation systems, although it still needs a lot of work.
What you’re writing about is what MusicIP already does.
That’s very interesting. My original article referred to Platinum Blue, but it does look from their website as if MusicIP is doing something like what I described. Thanks for pointing it out.