New AI Better at Judging Music Quality, Could Improve Playlists
This could mean smarter music recommendations and better AI-composed tunes for your ears.
Researchers propose FUSEMOS, a dual-encoder fusion framework for perceptual evaluation of generated music in text-to-music systems. It integrates CLAP, which extracts audio and text features for predicting human-annotated mean opinion scores, with MERT, a self-supervised model pretrained on large-scale music audio, to better capture both high-level semantic alignment and detailed musical characteristics for more accurate overall musical impression and textual alignment estimation.
FUSEMOS also uses a novel ranking-aware composite loss to address limitations of conventional regression losses in measuring human-preferred relative ranking of music segments. On the MusicEval benchmark, it outperforms existing state-of-the-art methods in mean squared error and rank correlation metrics by a large margin.
- FUSEMOS combines two AI models to better capture both the overall mood and fine details of music.
- It uses a new training method that mimics how humans rank music, improving accuracy.
- In tests, it outperformed existing methods by a large margin, promising better music recommendations and AI-generated music.
Why It Matters
Better AI music evaluation could lead to more personalized playlists and more natural-sounding AI-composed music for everyone.