Academic Minute
5:00 am
Wed August 1, 2012

Dr. Gert Lanckriet, University of California San Diego – Audio Search Engines

In today’s Academic Minute, Dr. Gert Lanckriet of the University of California San Diego explains efforts to create a search engine for music.

Gert Lanckriet is an associate professor of electrical and computer engineering in the Jacobs School of Engineering at the University of California San Diego.  He currently heads the UCSD Computer Audition Laboratory (CALab) and leads an interdepartmental group on Computational Statistics and Machine Learning (COSMAL). He received his M.Sc. and Ph.D. degree in Electrical Engineering and Computer Science from the University of California, Berkeley.

About Dr. Lanckriet

Dr. Gert Lanckriet – Audio Search Engines

There are probably a million bands you’ve never heard of, but that you might like, if you heard their music. Many of them have music online. The challenge, however, is how to weed through all that online content and discover an amazing band. It seems next to impossible. But what if you could train a computer to tag all of the songs on the Internet, using the kind of free word associations that humans use all the time to describe music to their friends? Like, “Good for dancing.” “Romantic.”  
 
Could a computer reliably tag every song out there? Could anyone from anywhere in the world find any band or song, even if they never heard it before, just by typing in a few key words? The answer is yes and our solution is game-powered machine learning.

First, we asked music fans from around the world to play an online game called “Herd It.” They listen to a song and tag it with labels such as romantic or jazz or music that’s good for early morning. This is what we mean by “game-powered.” The computer then analyzes the sets of examples provided through the game and identifies common patterns in their audio content. It then uses what it has learned to label millions of other songs online. This is machine learning.  We need the automation of machine learning since humans labeling songs around the clock could never keep up with the volume of music being uploaded to the Internet. So we are training computers to understand music the way humans do.

Based on this research, we created a music search engine. Eventually, our search engine will provide personalized radio stations that select songs for you that are adapted to your mood or activity throughout the day. By that I mean, the system would know you like to listen to hip hop when you’re working out and play it automatically at those times. And thanks to the enormous database of music created through game-powered machine learning, you would never have to hear the same song over and over again.

Related program: