'Father of the MP3' Teaches Machines to Parse Music

Computers have been telling us what to listen to for years -- at least since Amazon first told us that if we like X, we might like Y. But humans always have a hand in those recommendations: Pandora hires music experts, while Apple's Genius and CBS's Last.fm use collaborative filtering that notices when songs, artists or albums appear in the same collections.

On the other hand Mufin -- the core of which was developed by a team run by “father of the MP3" Karlheinz Brandenburg (pictured) -- ignores artist, album, genre and the other music people have in their collections and recommend songs based solely on raw computational analysis. The site recently launched as a private beta with nearly four million tracks. The results are intriguing -- but not always what you might expect due to the rational nature of these algorithms, which dispenses with all notions of style or trendiness.

“There are tons of music sites out there, but as far as I know, we're the only one that depends on a fully automatic recommendation engine," said Petar Djekic, Mufin's head of marketing. Mufin's audio analysis engine finds music's sonic siblings based on 40 attributes per song including percussion, style, speech, sound density, vocals, tempo, sound color, instruments, volume, dynamics and loudness. Other attributes are at play here too, but Djekic says they cannot be expressed as mathematical manifestations of musical concepts because they consist of statistical elements gleaned from humans' reactions to certain music.

“We first create a tag set where, first of all, people manually tag the music," said Djekic. “We have this set of music files that have been manually tagged, and then we see how our algorithm tags the music so we can improve." Once these rules have been set into the algorithms, there's no human intervention in Mufin's music analysis as it encounters new music.

The core technology behind this automated music parsing is AudioID, developed Brandenburg, whose PhD and undergraduate theses formed the basis of MP3 and other forms of audio compression. His team of audio analysts within Germany's massive Fraunhofer Gesellschaft organization won a company award for AudioID in 2002. Two years later, because it lacked the ability or inclination to exploit the technology, Fraunhofer spun the unit off as a commercial division called m2any that was purchased by Magix in 2007.

Since then, AudioID has formed the basis of Magix Mufin MusicFinder, a $20 downloadable app that helps users find “sound-alike" music on their own computers. It's taken Mufin about a year and a half to apply the same process to the nearly four million of tracks on the Mufin website -- and potentially on Amazon, iTunes and beyond, assuming these potential partners see value in these automated recommendations.