# SpeechJudge: Towards Human-Level Judgment for Speech Naturalness

**Authors:** Xueyao Zhang, Chaoren Wang, Huan Liao, Ziniu Li, Yuancheng Wang, Li Wang, Dongya Jia, Yuanzhe Chen, Xiulin Li, Zhuo Chen, Zhizheng Wu  
**arXiv:** [2511.07931](https://arxiv.org/abs/2511.07931)  
**Date:** 2025-11-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2511.07931v2/

## Key points

- Aligning large generative models with human feedback is a critical challenge in speech synthesis due to the lack of a comprehensive human preference dataset.
- The SpeechJudge suite was introduced to address this issue, consisting of a dataset, benchmark, and reward model focused on naturalness – a key subjective metric in speech synthesis.
- SpeechJudge-Data is a substantial corpus comprising 99K speech pairs annotated for both intelligibility and naturalness preference, incorporating diverse zero-shot text-to-speech (TTS) models across various speech styles and languages.
- SpeechJudge-Eval serves as a rigorous benchmark for evaluating speech naturalness judgment and highlighted the shortcomings of existing metrics and AudioLLMs in this task.
- SpeechJudge-GRM, a generative reward model based on Qwen2.5-Omni-7B, demonstrated superior performance on the SpeechJudge-Eval benchmark through post-training processes involving Supervised Fine-Tuning (SFT) with Chain-of-Thought rationales and Reinforcement Learning (RL) with GRPO on challenging cases.
- SpeechJudge-GRM achieved an accuracy of 77.2% (and 79.4% after inference-time scaling @10), surpassing a classic Bradley-Terry reward model at 72.7%.
- This tool can enhance the alignment of speech generation models with human preferences during the post-training phase, providing valuable resources for researchers and developers in advancing the quality and naturalness of synthesized speech outputs effectively.

## Summary

, , , , 

In the field of speech synthesis, aligning large generative models with human feedback is a critical challenge. This is due to the lack of a comprehensive human preference dataset, which hinders the development of models that accurately reflect human perception. To address this issue, the SpeechJudge suite was introduced. It consists of a dataset, benchmark, and reward model focused on naturalness – a key subjective metric in speech synthesis. The foundation of SpeechJudge is SpeechJudge-Data, a substantial corpus comprising 99K speech pairs annotated for both intelligibility and naturalness preference. This dataset incorporates diverse zero-shot text-to-speech (TTS) models across various speech styles and languages. From this data, SpeechJudge-Eval was established as a rigorous benchmark for evaluating speech naturalness judgment. The evaluation highlighted the shortcomings of existing metrics and AudioLLMs in this task. Even leading models like Gemini-2.5-Flash fell short of achieving high agreement with human judgment. To bridge this performance gap, SpeechJudge-GRM – a generative reward model based on Qwen2.5-Omni-7B – was developed. Trained on SpeechJudge-Data through a two-stage post-training process involving Supervised Fine-Tuning (SFT) with Chain-of-Thought rationales and Reinforcement Learning (RL) with GRPO on challenging cases, SpeechJudge-GRM demonstrated superior performance on the SpeechJudge-Eval benchmark. It achieved an accuracy of 77.2% (and 79.4% after inference-time scaling @10), surpassing a classic Bradley-Terry reward model at 72.7%. Furthermore, SpeechJudge-GRM can serve as a valuable tool during the post-training phase of speech generation models to enhance their alignment with human preferences. The comprehensive nature of the SpeechJudge suite provides researchers and developers in the field of speech synthesis with valuable resources to advance the quality and naturalness of synthesized speech outputs effectively.