Research map / Large language models / Alignment and preferences
ParaS2S: Benchmarking and Aligning Spoken Language Models for Paralinguistic-aware Speech-to-Speech Interaction
Shu-wen Yang, Ming Tu, Andy T. Liu, Xinghua Qu, Hung-yi Lee, Lu Lu, Yuxuan Wang, Yonghui Wu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GRPO RL with the distilled reward model after SFT | Kimi-Audio SFT · ParaS2SBench (all) | ParaS2SBench score: 3.955 -> 4.382 | Improved |
| add SFT warmup and GRPO to the base Kimi-Audio model | Kimi-Audio · ParaS2SBench (all) | ParaS2SBench score: 2.350 -> 4.382 | Improved |
| add GRPO on top of SFT, judged by humans | Kimi-Audio SFT · Human evaluation (all) | mean opinion score: 3.996 -> 4.303 | Improved |
| use multi-stage judge with ground-truth labels instead of end-to-end audio LLM judge (ablation) | gpt-audio judge · Judge-human correlation | Pearson correlation (avg): 0.618 -> 0.776 | Improved |
| describe output style with a natural-language tone description (ablation) | multi-stage judge O1 · Judge-human correlation | Pearson correlation (avg): 0.757 -> 0.776 | Improved |
| add output emotion tags instead of tone descriptions (ablation) | multi-stage judge O2 · Judge-human correlation | Pearson correlation (avg): 0.776 -> 0.731 | Hurt |
| add both emotion tag and tone description for output (ablation) | multi-stage judge O2 · Judge-human correlation | Pearson correlation (avg): 0.776 -> 0.743 | Hurt |
| replace ground-truth input labels with captioner predictions (with PolyTone) | multi-stage judge O2 · Judge-human correlation | Pearson correlation (avg): 0.776 -> 0.723 | Hurt |
| add PolyTone training to the input style captioners (ablation) | multi-stage judge · Judge-human correlation | Pearson correlation (avg): 0.701 -> 0.723 | Improved |
| add PolyTone training to the captioners (ablation) | multi-stage judge · Judge-human correlation (emotion) | Pearson correlation: 0.645 -> 0.684 | Improved |
| add PolyTone training to the captioners (ablation) | multi-stage judge · Judge-human correlation (sarcasm) | Pearson correlation: 0.669 -> 0.703 | Improved |
| RL with the judge on SFT models trained with 10 to 100 hours of data (ablation) | Kimi-Audio SFT · SFT data regimes | ParaS2SBench score | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LLaMa-Omni 2 — LLaMA-Omni: Seamless Speech Interaction with Large Language Models (ICLR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Enhancing Speech Large Language Models through Reinforced Behavior Alignment (NeurIPS 2026 · similarity 0.89)
- Dynamic-SUPERB Phase-2: A Collaboratively Expanding Benchmark for Measuring the Capabilities of Spoken Language Models with 180 Tasks (ICLR 2025 · similarity 0.88)
- Aligning Spoken Dialogue Models from User Interactions (ICML 2025 · similarity 0.87)
- SD-Eval: A Benchmark Dataset for Spoken Dialogue Understanding Beyond Words (NeurIPS 2024 · similarity 0.87)
- MMSU: A Massive Multi-task Spoken Language Understanding and Reasoning Benchmark (ICLR 2026 · similarity 0.87)
- Paralinguistics-Aware Speech-Empowered Large Language Models for Natural Conversation (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
