Tuesday, April 10, 2007

Tracking Down a Tune

A new website is tackling a tough problem: searching for songs based on hums.

By Julie J. Rehmeyer

Credit: Technology Review

Over and over, you hum a melody, trying to identify the song. You sing it to a friend, but either he doesn't know it, or he can't make out your tuneless drone. Until recently, you were out of luck. But now, a new website called Midomi.com can hunt down a tune for you when you hum or sing it into your computer's microphone. And it will even automatically correct for your mistakes, says Keyvan Mohajer, CEO of Melodis, the maker of Midomi.

Technology that searches for music has lagged behind text-based search partly because the latter was deemed a more important service, so more research resources were put into it. Moreover, sound of any kind--musical, vocal, and so on--is composed of many complex signals, which makes it difficult for a computer to analyze and categorize.

But as personal and commercial digital music libraries burgeon, there is a growing need for better ways to search for and discover music. And even without the ability to analyze all aspects of a piece of music, researchers and companies such as Midomi are testing ways to make music search commercially feasible for the first time.

"We're right on the edge of being ready for prime time," says Bryan Pardo, professor of electrical engineering and computer science at Northwestern University, in Evanston, IL. Pardo, an expert in music-recognition research at Northwestern, is not affiliated with Midomi. "We're about where the search engines were in 1990 or 1992."

Midomi--which launched in January--is one of the first companies to identify songs from their sung or hummed melodies. Indeed, other companies are also tackling the musical search problem, but in different ways. For example, the British company Shazam identifies samples of music played into a cell phone or PDA microphone. And AudioRadar is software that makes maps of collections of music to help people find music that is similar to what they like. (See "Audio Software for the Moody Listener.")

But these companies don't search for sung or hummed melodies, as Midomi does. When users come to Midomi's website, they sing or hum a melody into a microphone on their computer. The website compares the recording to sung melodies that other users have entered in its database. If it finds a hit, it displays samples of the original recording, along with an offer to sell the full recording. Users can also listen to the versions of the melody that other people have sung.

Midomi's approach relies on searching through a database of hums, and thereby avoids one of the complex problems encountered when trying to match hums to recorded songs: picking out a melody amidst complex instrumentation and singing and matching it to the input hum. It's simply easier to match a hum to another hum rather than match it to an original recording, says Mohajer. Although the details are proprietary, he says that Midomi's software makes its matches by comparing the melody, rhythm, and words of the sung tune with others in the database.

The company's ambition is "to build the most comprehensive database of searchable music in the world," says Mohajer. The best way to do that, he says, is by turning to its users. On Midomi, users can look up the titles of songs they like to sing and record themselves singing the melodies. The database currently has only a few tens of thousands of songs, although Mohajer says it is growing rapidly.

One of the challenges of search-by-humming is that most people aren't very good singers. But Midomi claims to have beaten the problem of sour notes. "As long as a human can tell what you're singing, we think our technology can tell what you're singing," says Mohajer.