Music Generation

Reinforcement Learning for Music Generation with Pairwise Reward Modeling
DuetJudge directly compares two music clips under the same caption, prioritizing text–music alignment before perceptual quality and reaching 71% agreement with human preferences. Its pairwise rewards improve Audex-2B, Audex-30B-A3B, and diffusion-based ETTA through reinforcement learning.
Zhisheng Zheng, Zhifeng Kong, Sang-Gil Lee, Jaehyeon Kim, Arushi Goel, David Harwath, Mohammad Shoeybi, Bryan Catanzaro, Wei Ping
Cite Demo Audio Samples
Reinforcement Learning for Music Generation with Pairwise Reward Modeling