As you said, as far as I know, Spotify does its playlists mainly using machine-learning, audio files' tags and user ratings, and not that much via analyzing the

As you said, as far as I know, Spotify does its playlists mainly using machine-learning, audio files' tags and user ratings, and not that much via analyzing the actual content of songs. As it's an open-source project that is supposed to work even without internet access and we don't want to have huge databases of user recommandations, we chose to go for the audio analysis part. I'll enhance the documentation (as it may be a bit scarce right now) about the analysis process, but Bliss extracts features that are supposed to be disjoint to compute the coordinates: Tempo is how much « quick » the track is, attack, how much there is abrupt changes in the music, amplitude, how « loud » the song is overall, and the frequency analysis checks whether the track is globally high-pitched or deep.