Orpheus among the algorithms (FF pt. 3)

It's a golden "Chamber of Commerce" day in Orlando in February. In a vast darkened hall at FASTForward08, 1,200 people from northern climes are intent upon a 90-foot video screen. There's no porn, no bloodsport, no gladiatorial contest under the glare of hot and heavy wagering.

If there's an edge with blood on it, it's in the hands of two softspoken geeks from Comperio, an Oslo-based firm, who are walking us through a search tool that responds to music.

On the large screen is a databank of 10,000 songs, with columns labelled artist, genre, decade -- metatags provided by the customer who supplied the songs.

The last column is different. It's labelled "mood," and derives from Comperio's search wizard that purports to extract metadata not from labels, but from the music itself, assigning to each song a "vector" of the song's musical character.

According to the blurb:

Monetize the Long Tail
Unlike regular multimedia search, Comperio Music SearchTM unlocks the information inside the music. State-of-the-art content analysis algorithms pull the music apart and record over 100 unique features for every song, covering important musical facets like rhythm, harmony, timbre and instrumentation, intensity, structure and complexity. This is Comperio Music MinerTM.

The brief demo involved looking up the mood for "Party." The user was offered two moods: Furious and Relaxed. As proxies for the user (and perhaps as Norwegians in Orlando), they chose Relaxed. An Elvis Costello song played -- (sorry, memory fails as to which - anyone who was there recall?). Then the engine was told to search the database for similar relaxed party music.

The algorithms fetched the hardest working man in show business:

A second test began with ABBA's "Dancing Queen."


The search for similar songs brought back a long list, among which was "My Baby Left Me" by Creedence Clearwater.



More later.

Technorati Profile