Mapping the music genome

A new service can match your tastes with tracks you don't even know. Danny Bradbury reports

The US company Pandora Media (formerly Savage Beast Technologies) launched the project in 2000, mapping the underlying characteristics of hundreds of thousands of songs to create a music recommendation system that it hopes will find new links between songs.

Having spent the past few years selling the service to large music retail chains, Pandora launched an online consumer service at the end of August. It offers customers a more tailored experience than traditional music recommendation systems, says its founder Tim Westergren.

"Our idea is that by understanding the core sounds - the qualities that give each song its sound and style - we can recommend other music to you based not on purchasing history and what other people have bought, but on what the music actually sounds like," he says.

Many music recommendation systems try to automate the process by basing their techniques on sales. Sometimes, this can backfire. Browse Tracy Chapman's album Matters of the Heart on Amazon's website, for example, and it will tell you that people who looked at that album also bought a Bryan Adams record - and eight other Tracy Chapman albums.

Pandora's service seems more intelligent. Enter Tracy Chapman's song "Woman's Work" into the Pandora system, and the first three things it plays are "Emerald River Dance" (a home recording by Judee Sill), "Storm" by José Gonzalez and "Mercy Street" by Kate McGarry. Even though they're by different artists, the songs have a similar quality.

By operating at a song-by-song level, Westergren hopes to break down the barriers between genres, finding matches between songs by artists people would not normally associate. "The analysis precludes the genome from having any preconceived notion of where a particular artist belongs. So The Beatles and Johnny Cash would wind up on the same playlist if it makes musical sense at a 'song' level."

The key to Pandora's system is the use of humans in the equation. The company employs a team of analysts, musicians who have studied music formally for at least four years and regularly play professionally. They listen to songs and rate them using up to 400 individual parameters.

Users of the service give examples of some songs and artists they like, which are then matched with songs sharing their musical qualities. The service, delivered via a Macromedia Flash-based player, plays a stream of these songs, creating a personalised online radio station. Listeners can tell the software whether they like or dislike particular songs as they play, refining Pandora's playlist over time.

Listeners can even ask the station why it is playing a particular song. When it played the Judee Sill track, Pandora's player said: "Based on what you've told us so far, we're playing this because it features mild rhythmic syncopation, minor key tonality, acoustic sonority, a good dose of acoustic guitar pickin' and acoustic rhythm guitars."

Traditional recommendation systems tend to advise people to listen to the same albums many others listen to. Westergren hopes to exploit whatWired magazine calls the "long tail effect". The idea is that, while a small number of products make up a large quantity of sales, there are many products in relatively low demand that don't sell well on their own, but which together can outsell the more popular products.

This gives an advantage to e-commerce retailers, who can manage larger product inventories. They can carry individual products that rarely sell, offering more choice to the customer - but only if the customer can find them.

That is where smart recommendation systems come in. "That's what the genome intends to do - to be that record-store clerk, in the small shop who knows you as a listener and has an encyclopedic knowledge of music," Westergren says. "That's how you solve the long tail."

No wonder, then, that Pandora has signed deals with Amazon and iTunes - long tail-friendly vendors - enabling listeners to select the song Pandora is playing and buy directly from them. In this sense, it's serving as an enhanced filtering system for online retailers. Commissions from these sales will form part of Pandora's revenues, but Westergren is still expecting the majority to come from subscriptions. Listeners pay $3 per month (about £1.65).

Nolan Gasser, a lecturer at Stanford University's music school, helped to define the parameters. "You look at melody and harmony and form and rhythm and texture and instrumentation to see how all of these fit together to make a composite piece of music," he says. Song profiles also include an analysis of voices, examining qualities such as gender, whether it is a high or low voice, and whether it is smooth or gravelly.

But the very thing that makes Pandora attractive may also be its weakness - analysts can only plough through so many songs per week (more than a thousand, Westergren says), and it shows. Search for Inspiral Carpets - hardly obscure - and the service comes up empty-handed.

There is also no classical music yet. Try entering Philip Glass's opera Einstein on the Beach as a favourite track, and the software automatically adds Einstein on the Beach by Counting Crows without giving you the chance to validate the entry. Such howlers can lead you to delete and start again; irritating if you have spent weeks building up your profile. Pandora will fix that soon, Westergren says .

It may not be perfect, but Pandora's music genome service still presents some interesting choices. It recommended music from my own collection that I hadn't told it about - and also challenged my preconceptions by recommending music by artists I didn't think I'd like because their image didn't appeal. I always thought I mainly liked quirky bands like Pavement and Ween. Who knew I'd end up enjoying tracks by Christina Aguilera?