Machine Learning Could Lead to True Music Search EnginePandora is a nifty music radio service with a killer feature: it can predict which songs you will like and which you will loathe. The company did this by hiring professional music classifiers to come in and tag each and every song in its library with information about the music’s tempo, vocals, structure and personality.
Its a nifty trick, and one that could and should be generalized, but hiring all those professionals costs a lot of money.
Thanks to new research, however, we might soon see the rise of a true music search engine. A search engine that can take any song and provide a list of related songs based on arbitrary metrics.
The University of California wanted to know if a group of untrained, unpaid volunteers could tag music as well as paid professionals, with the goal of getting a sample of tagged music for a computer to learn off of. And they found that, yes, unpaid music lovers can do it just as well.
Combine this with a computer designed to learn from these samples and you have automatic tagging of songs. With that automatic tagging comes the ability to search for songs with a simple query, for example “‘funky’ or ‘laid-back’ ‘folktronica’”.
The real question is how to get people to tag songs like this. The easiest solution: gamification of the process. Through simple reward mechanisms, players are encouraged to keep tagging music until most of the music on the web is covered. What’s more, the authors of the paper made it possible for the computer controlling the data collection to alter the games based on what they had learned. As things keep going, the computer learns what it should be asking the players to get better tags.
Said Gert Lanckriet, a professor of electrical engineering at the UC San Diego Jacobs School of Engineering and lead author on the paper,
"It's like a baby. You teach it a little bit and the baby comes back and asks more questions."