TuneJury: An Open Metric for Improving Music Generation Preference Alignment

TuneJury is a newly released open reward model designed for text-to-music generation, which predicts music preference scores based on text prompts and audio clips. It employs a pairwise comparison approach, utilizing a variety of human-preference labels for training, and demonstrates strong performance on both held-out test pairs and out-of-distribution benchmarks. The model introduces anchor calibration for improved data efficiency in aligning preferences across multiple applications, making it a valuable tool for practitioners in music generation who seek to enhance preference alignment in their systems.