News

It's funny how things sometimes have to hit you in the head before you actually notice them. When Jon Iverson and I were cruising CES 2007 in January, we were stopped in our tracks by Sooloos's media server technology. Sooloos got the whole no-compromise audiophile approach, but what really knocked us out was how thoroughly it utilized the metadata from AMG (All Media Guide): Feed the the Sooloos server a disc and it would access AMG's database and collect artist name, album title, track names, cover art, genre, mood, side-man information, and even "associated concepts."

When Rob Darling asked for the name of any musician we liked and instantly created a playlist of every album he had played on—not simply headlined—we were stunned. "Who needs a memory when you have AMG?"

"Precisely," Darling purred.

"But," I persisted, "What happens when you have a mix CD that has different tracks on it than any commercially available disc?"

"AMG has a technology called Lasso that doesn't just use track durations and sequencing to identify tracks. It uses acoustical analysis of digital recordings to identify unique acoustic fingerprints. I don't have the data at my fingertips, but I think AMG has a baseline of something like six million fingerprints."

Recent events, specifically the still-breaking Joyce Hatto scandal, have catapulted computer recognition of individual recorded performances into the headlines, so I thought I'd investigate AMG and see if it had other technologies I was unaware of.

Boy, do they.

Most music lovers probably know AMG as the publisher of the book series All Music Guide, which is now over 15 years old, with editions dedicated to popular, classical, jazz, hip-hop, blues, and every other genre with traction. The AMG books were compiled from capsule reviews written by professional critics and were accurate and as comprehensive as physical publishing could make them—which means they were always at least a year out of date.

Early on, AMG realized that it needed to move its database into the Internet age, and in 1995 it established a World Wide Web presence through its website www.allmusic.com. Actually, as useful as reviewers like me—and consumers everywhere—found allmusic, its real purpose was to demonstrate the breadth of information contained within its databanks for potential licensees.

AMG realized the importance of information early in the digital age, so it collected metadata, or "data about data," which not only included genres, credits, copyright registration, and catalog numbers, but also descriptive information such as styles, "tones," moods, and themes. It also collected (and still collects) information about concepts, such as relational content, linking to similar or influential artists or recordings.

I recently spoke to Greg Smith, vice president of technology; Andrew Stess, vice president of consumer electronics; and Allen Schrott, senior managing editor for classical about aspects of AMG's technology that might be important to audiophiles.

Allen Schrott started the ball rolling with a statement some vinylphiles might find provocative. "As people move to digital storage as a consumption model, we can supply them with the same information they used to rely on liner notes for."

Andrew Stess agreed, "It's not just about artist/title/track any more."

"Although," Greg Smith added, "the consumption model today is very strongly track-centric."

"That brings up a concern I have as a classical listener," I said. "If you have a large collection and you listen with the random function, you end up getting bleeding chunks—isolated movements rather than complete works. Is there any way of linking complete works through the metadata?"

"That's actually an essential part of our classical data structure that we're very proud of and very attentive to—something that we recognized years ago," said Schrott. "Every single track of every classical recording that we deal with is hard-linked through a unique ID back to the unique composition that it comes from. We actually have a separate database of musical works—separate from the album content—so if it's Beethoven's 5^th symphony, it always gets that unique ID on the track. In other words, all of the movements of the 5^th symphony get linked together, as well as with the other tracks on that recording. We can do other, more creative cross-linking, too, but we definitely have the ability to keep works complete within our metadata structure."

"So how do I link movements together?"

"That depends on the implementation of whoever is using our data," Schrott replied. "Not every music server has quite caught up to what the consumer wants to do yet, but we have the data marked so that when they do, the possibility is there."

"So when I buy a server, I have to find out if they're really using AMG's technologies the way I want to employ them?"

All three: "Pretty much."

"So what's a good test of how effectively somebody is using all of your data? Do I try to run down every track in my collection by an obscure musician?"

"We can go even deeper than that," said Schrott. "We can use the characteristics of the music, including the instrumentation, genre, or other qualifiers to create dynamic playlists attuned to your situation. That includes mood, by the way. Music is not just mathematics, it's situational and it's cultural, so striking the right mood is important to people—we call that situational listening."

"How do you devise a metric for that?"

"We have people that listen to each track and indicate parameters from our immensely detailed lists, but we don't stop there," said Schrott. "We also rely on our community of contributors, combining, if you will, the opinions of 'experts' with input from users. One of our core technologies is a data-mining technology in order to process all that information and turn it into something useful."

"By data mining," said Stess, "we mean comparing large numbers of submissions and coming up with estimates of individual input compared to that of the group's."

"That's where we have to be very careful," explained Schrott. "You don't want to average all the input, that just homogenizes the results—that skews things in the wrong direction. We want information to be specific, not an average."

"It's the unique stuff that people are really looking for," added Stess.

"We like to establish the credibility of our contributors," said Schrott. "With our recognition technology, we can see that a database contributor's collection is skewed toward jazz, so we might trust that person's opinion of jazz music more than their opinion of rap music, if they just had one rap track on their hard drive.

"We're all about content. We use technology to get the job done, but what we're selling is making it easier to enjoy the music. All of that should be fairly transparent to the end-user."

Since I always have to be a wise guy, I quipped, "You know what would be really transparent? A program that would sense what I'm in the mood for without my having to think about it!"

"We're working on that," Greg Smith laughed. " If you go to Tapestry on our website, and you'll see that we're almost there."

Oh brave new world, that has such wonders in it.