Behind the scenes, music streaming service Deezer is having some fun experimenting with deep learning and detecting the ‘mood’ of a song – based on both lyrics and instrumental set-up.
It may not be yet available to help Deezer curate playlists and songs for users, but the service’s research into allowing technology to work out what makes a happy and sad song is nonetheless interesting.
The researchers over at Deezer are training AI to monitor both how lyrics and instruments craft a song. The team states that it has developed a deep learning system that processes the full song (not just a cheerful-seeming chorus) and measures the emotion of the song through analysing different data. The AI makes use of raw audio signals, linguistic context reconstruction models and a Million Song Dataset that aggregates Last.fm tags describing tunes (e.g. ‘sad’ or ‘upbeat’) to do this.
Researchers were able to extract certain words from the lyrics in the process of mapping this Million Song Dataset to its own library using metadata – creating a database of over 18.5k songs to experiment with.
Deezer is clear that this is very much a work in progress at the moment. Researchers say although it’s better than mood detection techniques in the past, there’s still work to be done improving the system. Next on the agenda is developing various training models and creating a system that can process songs unsupervised and critically, without tags.
Later down the line, the perfect song to capture you mood could be only a click away without even thinking about it – whether that’s a scary or exciting thought, is down to you.