TuneJury: open metric for music generation preference alignment

Arxiv CS.SD (Sound/Audio ML)

TuneJury is an open, instance-level pairwise reward model for text-to-music that predicts preference from prompt and audio, trained on public human labels. A challenge entry using TuneJury as a reward signal improved FAD-CLAP and CLAP scores in the ATTM Grand Challenge.

text-to-music-generationAlignment

Arxiv CS.SD (Sound/Audio ML)·Yonghyun KimJunwon LeeHaiwen XiaYinghao MaJunghyun KooKoichi SaitoYuki MitsufujiChris Donahue

Jun 16, 4:00 AM