Here's the second in my series of interviews for the book, with many thanks to Zac Johnson for his time and insights. As these interviews are not the primary source of research for the book, I'm not aiming to be comprehensive, but I'm very open to talking to anyone who'd like volunteer their views and describe work relevant to the book's themes, especially if these are different to the ones I've documented. I guess what I'm saying is that if you read this and think, "Hey, I've got something to say about that, so David should speak to me", please don't wait to see if I contact you, because I may not — please just get in touch.
Zac Johnson has been working at All Media Guide (AMG) for six years. His focus is mostly on AMG's music site, AllMusic.com, though AMG also publishes the All Movie Guide and the All Game Guide, and Zac also has a view across the three sites. Recently his focus has been on what he calls "intelligent playlisting", through the Tapestry service (which I covered in part on my other blog).
Zac sees two main applications for intelligent playlisting. The first is music discovery: where you provide examples of the music you like and ask, "play me more stuff like this, but which I haven't heard before". The second is soundtracking, where you say, "I want to go driving/fishing/to the gym — give me a bunch of music that fits this activity (and which I may already know)".
The same technologies underlie these two applications. They are generally based on one of three ways of collecting data about music, which drive the intelligent playlisting.
- The community/collaborative filtering approach, epitomised by Last.fm (as well as Amazon and others), collects data on users' preferences and uses this to make recommendations for other music that individuals might like.
- Acoustic wave analysis, employed by companies like MusicIP, compares the physical properties of the sound of different pieces of music and maps them on scales like mellow versus aggressive, or slow versus fast tempo.
- The editorial-driven approach, exemplified by Pandora and AMG's Tapestry, involves trained people analysing and tagging each track.
The area that Zac thinks is promising is the combination of the first and third of these approaches. The limitation of collaborative filtering is that it tends to reinforce popular and established patterns of preference, at the expense of exposing new or less widely appreciated music. So, for example, if you say you like The Beatles, you may get recommended Bob Marley or James Brown, because they are all very popular artists and many people like all three. By factoring in the editorial approach, Zac believes, it's possible to open up what he calls the 'tastemaker' end of the spectrum: providing recommendations for new music which is just starting to gain a following.
I ask how well this editorial approach to intelligent playlisting extends to the other media that AMG covers. At first, Zac replies, they could see the point of a playlist of 20 songs, which you could listen to in about an hour, but were sceptical about the value of a playlist of 20 films. That's enough to programme a whole festival. But now AMG sees some value in such playlists for recommendation purposes. Film recommendation services are fairly thin on the ground — Zac references Netflix and movielens as examples.
We talk too about classical music, and I observe that the pop and rock music entries on AllMusic.com are tagged with 'mood' entries (things like 'spacey', 'autumnal', 'precious' and 'wry'), but the classical entries don't have moods and the descriptors could be said to be more conservative. I wonder if that is because the critical discourse around classical music is longer-established and doesn't generally condone such flowery adjectives. AMG has some more classical music developments going on behind the scenes, Zac says, while conceding that at present the descriptions are more objective.
Although my interests are more with the discovery applications of intelligent playlisting, I'm also slightly curious about the soundtracking applications. Playlisting with digital music has really opened up this area, which previously required the labour-intensive compilation of mix-tapes. But what level of granularity can you sensibly apply to the activities that music accompanies? Soundtracks for driving, working out, and housework I can imagine; but how much further can you go in identifying more detailed activities that require a different playlist?
Zac explains that the level of definition they apply to the music is very fine-grained: AMG can find you a playlist made entirely of music with an organ, for example, or they can pin it down to a Hammond organ, or a pipe organ, or one of several other varieties. But sometimes, yes, they have to simplify. He gives the example of some work that the AMG Tapestry team are doing with Delphi, the car radio/audio company. The Tapestry data and technology can provide very fine-grained definition: for example, country music is broken down in a parent-child relationship to 'Bakersfield country', 'traditionalist country', 'neo-traditionalist country' and so on. But if drivers were required to specify their desires at this level of detail, they would, as Zac puts it, "wrap themselves around a tree" before they got to hear any music. So Zac works with Delphi's Human Factors Manager to develop a user interface that is more manageable and safe, probably with no more than about 20 genres of music. Similarly, AMG are working with several companies who aim to develop a user interface that, with a single click, will generate a playlist of songs similar to the one playing at that moment.
Finally I ask, "What makes a really good intelligent playlist system?" The important issue, Zac replies, is to evaluate the outputs — the playlists generated by any system — in a really critical, sophisticated way, and ask how they could be improved. Zac suspects that many software-led businesses may get to the point where they produced reasonable, but not necessarily great, playlists. MusicStrands, he says, are producing some excellent technologies, but, to take the example of the movies-to-music recommendations I wrote about recently, the quality of some of the recommendations seems open to question. Zac gives an example: "If you do a MusicStrands Labs Movie search on the film Blade Runner, it suggests A Horse With No Name by America and a blues version of The Simpsons Theme by Danny Gatton. I see a real value in the concept, and the technology seems like a fun exploration, but (in the true spirit of a 'Labs' environment), the results that come back as recommendations aren't quite fully developed yet."
In a different way, Last.fm is doing the best intelligent playlisting among the services currently commercially available, but is still subject to the Beatles/Marley/Brown limitations mentioned above. Zac is confident that some of the Tapestry-based services that his team have developed can improve on the Last.fm experience.
Nothing, he says, can replace a really good human DJ — whether that's John Peel on the radio or Paul Oakenfield in a club — but you can't have these DJs arranging your listening 24/7. And then there's the anonymity of using a computer-based system, which enables you to explore your 'guilty pleasures' without scoffing at you.