Software Tells Hip Hop from Funk, No Keywords Needed
Screenshot from the Facebook game Herd It. Researchers are using data from Herd It to teach a computer program to classify song files for search engines.
CREDIT: Jacobs School of Engineering
What if a person could read, but couldn’t understand or recall any sounds he heard or images he saw? That's the way search engines work, because they can scour text, but can't easily classify audio files, for example, if the files don't come with some text description. Three engineers from the University of California, San Diego and Ithaca College in New York have figured out a workaround. If they applied their idea to all the song files of the Internet, people could type in "funky," "spooky electronica" or their favorite genre to bring up matching songs, even if the songs aren't popular hits or don't have good descriptions. The researchers published their idea and preliminary results April 24 in the Proceedings of the National Academy of Sciences of the USA.
Because a computer program can't do it, the researchers got real people to classify some songs by playing a Facebook game called Herd It. Herd It players labeled songs with keywords such as "romantic," "jazz," "saxophone" or "happy," according to the University of California, San Diego's engineering school. The researchers found that the game-players' classifications were as accurate as those made by paid experts.
People could never classify all the songs on the Internet, however. The researchers needed an automated solution. So they handed off their categorized songs to a computer program they wrote. The program analyzes the waveforms for the songs – the wavy lines people may see when they open an audio-editing program – and finds which waveforms are associated with which labels. The process of taking a sample of human-classified files and using that example to figure out how to classify additional files is how computer programs learn.
Like any good learner, the music-classifying program knows its weak spots, too. It can recognize if it doesn't have enough human-labeled songs to recognize a category such as jazz. In that case, it will tweak Herd It to get people to label more jazz songs for it to learn from. "It's like a baby. You teach it a little bit and the baby comes back and asks more questions," said Gert Lanckriet, a University of California, San Diego electrical engineer who led the research.
In the future, Lanckriet hopes his music search method will help people create personalized radio stations that can learn what genres of music they like to hear at different times of day. "What I would like long-term is just one single radio station that starts in the morning and it adapts to you throughout the day," he said. "The user doesn’t have to tell the system, 'Hey, it's afternoon now, I prefer to listen to hip hop in the afternoon.' The system knows because it has learned the cell phone user’s preferences."