By analyzing the acoustic properties of songs on Spotify, intern and PhD student Sander Dieleman hopes to advance the streaming service’s recommendation algorithms to aid users in discovering new and lesser known music. Rather than basing recommendations on the choices people with similar tastes make, they would be based on songs the user listens to. This method, which requires deep learning, would then mix more obscure but user relevant songs into the recommendations.

The University of Ghent (Belgium) student, now based in New York, wrote on his findings in a blog post.

He explains that most recommendation systems use collaborative filtering, a method that requires usage data. This means that it only works on popular songs, as less popular songs have less listener data.

Dieleman’s program learned the acoustic features of thirty-second samples from around 50 percent of the million most popular songs on Spotify. He then used the other half of the most popular songs to test the deep learning system.

“What’s really cool about the blog post is that Dieleman includes sample playlists based on songs that activated certain of the 256 low-level filters in the neural network in different ways (e.g. bass drums, vocal thirds or chords),” explains GigaOM. “He also includes high-level-feature playlists (which combine results for all the low-level features and, in some cases, pretty much grouped songs by genre) and soundalike playlists for specific songs.”