Reproduce “Conversational Voice Aesthetic Model with Reinforcement Learning from Human Listeners” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xilin Jiang, Shun Zhang, Tejas Jayashankar, Yinghao Aaron Li, Osama Hanna
CVAM evaluates conversational voice qualities using preceding audio and a response, producing descriptions and labels. It is trained on model-written examples, then tuned with reinforcement learning against human vote distributions. On a CANDOR-derived test, forward-KL training reached 69.6% average accuracy and 50.1% macro-F1 across four perceptual label types, measuring agreement with listeners rather than objective truth.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.