I liked this article which was originally published on Last.fm by Mark Levy on Thursday, 12 May 2011. I found it useful for fellow deejays and music enthusiasts, so here's the article:We need your help. For several years scientists have been trying to extract meaningful information about music directly from the audio data in a track. Encouraged by successes in automatic speech recognition, for a while researchers hoped that it would be fairly easy for a computer to analyze an audio track and transcribe it directly to sheet music notation, recognising instruments, voices, and all the notes they were playing.
This was quickly found to be a much harder problem than expected, and a new field of research grew out of the failure to conquer this challenge.
Since then scientists have battled to recognise specific aspects of a song, such as its tempo, key, harmony, rhythm, instrumentation, or even just its genre, by analyzing its audio content. Starting in 2004, this battle has been formalised into an annual competition called MIREX.
In each round of MIREX, competitors submit their programs to a carefully controlled competition server, where each program is run on the same set of tracks. The winner is the one whose results agree most closely with a matching set of human judgements.
But gathering human opinions about very specific aspects of large numbers of songs isn’t easy. You need to find people with a keen interest in music, and ideally some musical expertise. You need to persuade them to answer questions that can seem obvious to them (even though they are still difficult for a computer). Finally you need to have enough helpers to allow for the fact that some interesting musical questions can have more than one ‘right’ answer.
We’ve been thinking about this problem recently, while evaluating some rival algorithms to estimate the tempo, or beats per minute, of songs directly from audio… and we thought of you! We know from many years’ experience that Last.fm users have the interest, the knowledge, and the sheer staying-power to make a huge contribution to scientific research in this field. As an extra incentive, improving these methods could help us offer you even better radio and recommendations in the long run.
To test this idea in practice, we’ve built a very simple application on our labs website where you can listen to music and help us improve the state of the art in tempo estimation. If you have fun with this, we plan to add to it over the coming weeks and months.
So please lend us your ears for a few minutes. Thank you!