Think of your favourite song. Now ask yourself this: Why is it your favourite? What is it about that particular collection of notes and rhythms that resonates with you emotionally and physically? Now think of another song that you really, really like. Why? And what’s the relationship of that song to your all-time favourite? Finally, think about discovering new music. What attracts you to a song you’ve never heard before?
All this can be very hard to put into words and you may find yourself saying “I just know what I like, okay? And I also know what I don’t like.”
That’s not good enough for Stephen Phillips, one of the leaders in for a 2009 startup called We Are Hunted, which has since been acquired by Twitter. He’s fascinated by the things that drive our musical preferences, prejudices and biases. And after reading this article, so am I.
The music industry has grown through evolutions of media. From physical formats like LPs and CDs to digital downloads and streaming. For fans, this means cheaper music, bigger catalogs and a range of devices. Today music fans are listening to more music than ever before. This growth demands ever more intelligent software that delivers better music experiences.
Determining song similarity is the base function of intelligent music discovery. Given a song or artist as a starting point, a similarity function is used to generate song playlists, determine related artists, or make personal recommendations.
There are several techniques for determining song similarity:
Expert – A music expert reviews each song against a set of descriptive features. These features are used to group related songs together. This technique produces really high quality playlists, but is difficult and expensive to scale.
Social – Use listening history to group songs together. People who like The Killers, also like The Strokes. Collaborative filtering produces excellent playlists for popular music, but is poorer for new music as it lacks sufficient history.
Acoustic – Group songs by their acoustic properties like tempo, timbre, melody and more. Calculating these features is technically challenging, but works really well as a similarity measure. It can perform poorly in certain situations as it lacks context like artist personality and popularity.
Mixed – Use a combination of these techniques in a way that favors their strengths and minimizes their weaknesses. At We Are Hunted, we wrote software that searched song reviews on music blogs (experts) to find the most common descriptive phrases (social). This approach had the positives of both techniques, in a way that would scale, with the ability to capture new music as soon as it was released.
There’s plenty more. Continue reading.