For a while, I've thought music recommendation systems were interesting. Some friends and I all felt that our Spotify Discover Weekly playlists were pretty bad, and so I was inspired to build a recommendation algorithm. Amit Gattadahali and I researched conventional music recommendation algorithms, and decided we would try something a little different by building one that used lyrics as the primary predictor.

This ballooned into a pretty big undertaking - involving several datasets and incorporating lots of nautral language processing and deep learning techniques.

We first used lyrics to predict 3 characteristics of a song: genre, topic, and language. We used a combination of the lyrics, these features, and several audio-related statistics to make the final recommendation.

Check out our paper and github repository for more details.