Singing at your computer is the new way to identify that
song that’s been stuck in your head. As long as it’s one of the roughly 3,000
currently in the database of the prototype “query-by-humming” search
engine Tunebot,
brought to you by Arefin Huq, Mark Cartwright and Bryan Pardo of the
Interactive Audio Lab at Northwestern University.
Tunebot, the guts of which are described in a paper just released at the 2010 Sound and
Music Computing Conference, tackles a problem so difficult that it makes
beating the world’s greatest chess grandmaster look like a party trick: How to transform off-key caterwauling of a visitor to your website (or the user of a forthcoming iPhone app) into a melodic signature that can be compared to a
database of thousands (and eventually millions) of songs, yielding something
that even approximates the result your visitor was looking for.
Advertisement
Just populating the requisite database of hummed or sung
tunes is a challenge. You could either have a small army of volunteers sing
multiple renditions of all three million tracks in iTunes database, or you
could turn it into a hit game.
This story is only available to subscribers.
Don’t settle for half the story.
Get paywall-free access to technology news for the here and now.
Karaoke Callout is a game in which a player has to sing, a-capella, a random song chosen by the
phone. The player’s rendition is run through a melody extractor – which determines the
fundamental frequency of the note they’re singing, once every 20 milliseconds. It is then compared to the nearest match in the database for that song,
and their performance is rated by how close it is to the “real
thing.”
Karaoke Callout is what’s known as a Game With a Purpose, which means it’s a way for a
computer scientist to outsource a huge amount of manual training by transforming the task into a leisure-time activity.
Huq et al. are confident that a straightforward
comparison of the melodic signature of songs sung or hummed into their search
engine with songs already collected from Karaoke Callout is an adequate
solution to their problem.
It’s also an elegant solution to the problem of melody recognition. Rather than trying to solve
the much harder problem of comparing a hummed tune to an existing recording or
some simplified version of a song, such as a MIDI encoding, Huq et al. are
comparing apples and apples: two different human beings singing the same song.
And the
more people who sing a particular track into Karaoke Callout, the more
accurately it will be able to identify that song. In other words, whatever hit
rate Tunebot ultimately achieves for Margaritaville – that’s as good as it
gets.