Good news for anyone who likes crunching through an academic paper about neural networks on a Monday morning. A team of researchers at Deezer’s Paris headquarters have published a paper titled ‘Music mood detection based on audio and lyrics with deep neural net’, which may shed some light on how the streaming service may evolve in the future.
The researchers took a database of 18,000 tracks, then used a deep learning-based approach to try to analyse their valence (“from negative to positive mood”) and their arousal (“from calm to energetic mood”) from their audio and also from their lyrics, as well as a model using both.
The experiments seem to have gone well: the researchers report that their models “achieve better results than classical approaches on arousal detection, and both methods perform equally on valence detection”.
There’s an argument that Deezer is playing catch-up here, given that The Echo Nest was exploring this territory years ago, before being acquired by Spotify. Still, these kinds of models are important cogs in the modern music-streaming process, so Deezer will doubtless be able to make use of this research in its service.