dekstop: weblog

Tom Conrad has apparently just presented "Pandora" at Bar Camp. Scoble quotes an email from Tom with a short description of the service:

"Pandora is a "music discovery service" designed to help you find and enjoy music that you'll love. It works like this: you give us the name of an artist or song and we instantly create a "station" that plays songs that share musical characteristics with the artist/song you entered. From there you can fine-tune the station to your tastes by giving us feedback on the individual tracks we play. You can make up to 100 unique stations that play all kinds of music - Pop, Rock, Jazz, Electronica, Hip Hop, old, new, big names, and small acts -- over 300,000 songs from more than 10,000 artists. Pandora is entirely web-based; you won't need to install any software to start listening."

...in short, it's the alternative for Last.fm that I've been waiting for. Judging from this description it is targeted towards very similar usage patterns, but has a slightly different focus. And it arrives just in time.

Last.fm had some major usability problems that prevented me using it more regularly. The major drawback of Last.fm was the slow server that liked to choke randomly when loading pages -- no wonder they were having these performance issues, they were serving some pretty complex pages. This seems to be adressed by their recent relaunch which involved a major overhaul of the site, but which also introduced a requirement of questionable benefit: you now need to install a player software to use it. This alone means Last.fm suddenly stopped being interesting to me. I'd like to choose my own tools, thank you.

What's interesting is the description of Pandora's internal logic that is briefly described on their website under the title The Music Genome Project:

On January 6, 2000 a group of musicians and music-loving technologists came together with the idea of creating the most comprehensive analysis of music ever.

Together we set out to capture the essence of music at the most fundamental level. We ended up assembling literally hundreds of musical attributes or "genes" into a very large Music Genome. Taken together these genes capture the unique and magical musical identity of a song - everything from melody, harmony and rhythm, to instrumentation, orchestration, arrangement, lyrics, and of course the rich world of singing and vocal harmony. It's not about what a band looks like, or what genre they supposedly belong to, or about who buys their records - it's about what each individual song sounds like.

Which means instead of recording what people listen to and using the collected data to infer musical connections (as Last.fm does it), they manually describe entities of music using a meta-language. They attempt to encode a musical Genome.

In 2001 two friends and I were working on something similar, on a much smaller scale. Falko Schmid, Michael Gustmann and I did a small University procect called A.R.S.E - the Associative Review Search Engine, which implemented a simple approach to a music description and recommendation system. The name of the project should make it clear that this was just a fun side-project, but we believed that the subject matter was potentially interesting to a lot of people, and even briefly pondered making it a commercial product. The talks we had during that period were both exciting and entertaining -- thinking up new product features was a lot of fun ("Personalize my A.R.S.E.").

In the end we decided against commercializing it, and the reasons for that are relevant to the Pandora project, which is why I'm mentioning it here:

• This is a lot of work. The descriptions can not be generated as a side-effect of consuming music, which is how last.fm does it, and which makes their service so powerful.
• Your descriptions can't change with your audience. Unlike Last.fm's approach, Pandora's descriptions can't easily adapt to a different focus. A description of a song, once written, remains fixed.
• You need to teach people how to properly describe music. You need to be consistent in your descriptions to make them valuable, even when very different people describe very different kinds of music. And the tools used are probably non-trivial due to the number of parameters encoded.
• It is an open question if you can even find a stable method to objectively describe audio. This is a major problem that many people would be interested to work out -- I once had a chat with Stephan Schmitt of Native Instruments, who said that one of their unsolved problems is categorizing instrument presets made by very different people at different times; they usually have so many presets per instrument that browsing by name is a very painful process. (You can only have so many onomatopoetic variations of "bruaaang").

These points touch on the core problem Pandora faces: they need to describe a large set of data, which involves an immense effort, but (as far as I can tell) they can't simply use the power of a large community to generate this data. It doesn't currently look like they will be trying the Wikipedia approach, where large and complex sets of data are accumulated over a longer time period by having many people do small pieces of work, one at a time. In contrast, they have a smaller set of people who each do a lot of work. This makes it possible to create a consistent and high-quality set of data, but it definitly limits the application. The strength of Wikipedia is that anybody can come and write an article about Saul Kaiserman.

It's not clear yet if this even matters. Not having used Pandora (it's still in closed beta) I can only guess at how they implement some of the specifics, and it is possible that they have found a way so that you only need to describe a subset of all music to be able to navigate all of it.

On the other hand it is questionable if, even if you have a refined description scheme, this description has relevance. I don't like Bloc Party because they use guitars, or because they had a release in 2005, or because their song "Banquet" uses an 8-part song structure and starts with a verse. The fact that you like a piece of music does not necessarily coincide with any measurable set of data; it might just be the fact that the track was playing while you met some girl for the first time, or that it reminds you of summer of 1999 in Italy. You can't encode emotion, and surely not the individual emotion of every listener. Reception of music, even if it is one of the most natural things in the world, is also one of the most complex.

I nevertheless believe that Pandora has a potentially exciting product on their hands, and I'm looking forward to using it. I'm welcoming Last.fm's first serious competitor to the open marketplace, and concratulate them for their effort. It's going to be exciting to see how these two evolve -- using two very different approaches to model similar types of data.

See also: William Lazar's two-paragraph account of the Pandora presentation at Bar Camp; and Danah Boyd's piece on her first experiences with the service.