arXiv 2511.07931
SpeechJudge: Towards Human-Level Judgment for Speech Naturalness
By Xueyao Zhang, Chaoren Wang, et al.
Published 2025-11-11
Wiki summary
Explore the paper's summary, context, and related research on Papiers.
Aligning large generative models with human feedback is a critical challenge. In speech synthesis, this is particularly pronounced due to the lack of a large-scale human preference dataset, which hinders the development of models that truly align with human perception. To address this, we introduce SpeechJudge, a comprehensive suite comprising a dataset, a benchmark, and a reward model centered on naturalness--one o…