Research map / Large language models / Alignment and preferences
Multi-Metric Preference Alignment for Generative Speech Restoration
Junan Zhang, Xueyao Zhang, Jing Yang, Yuancheng Wang, Fan Fan, Zhizheng 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 |
|---|---|---|---|
| apply multi-metric DPO on GenSR-Pref | AnyEnhance (MGM, 69k pairs) · Librivox-GSR | NISQA: 4.346 -> 4.865 | Improved |
| apply multi-metric DPO with only ~3k pairs | AR+Soundstorm · Librivox-GSR | NISQA: 4.535 -> 4.923 | Improved |
| apply multi-metric DPO with ~3k pairs | Flow-SR · Librivox-GSR | NISQA: 4.184 -> 4.825 | Improved |
| DPO alignment | AnyEnhance · Librivox-GSR | Speaker similarity: 0.955 -> 0.954 | No clear effect |
| DPO alignment generalizes to DNS speech enhancement | AnyEnhance · DNS With Reverb | OVRL: 3.204 -> 3.438 | Improved |
| use unanimous multi-metric preference pairs instead of single-metric pairs (ablation) | AR+Soundstorm DPO · Voicefixer-GSR | SIG/OVRL: SIG 3.564 vs 3.531 (NISQA-only), 3.537 (SIM-only) | Improved |
| single-metric NISQA preference pairs (ablation) | AR+Soundstorm DPO · Voicefixer-GSR | SIG: 3.550 -> 3.531 | Hurt |
| use DPO instead of SFT on ground truth or winner samples (ablation) | AnyEnhance (MGM) · Librivox-GSR training curves | all metrics | Improved |
| use ground-truth audio as fixed DPO winner (ablation) | AnyEnhance (MGM) · training curves | model stability | Hurt |
| use in-paradigm preference data instead of cross-paradigm (ablation) | AnyEnhance (MGM) · Librivox-GSR | OVRL: 3.232 (FM data) -> 3.409 (own data) | Improved |
| fine-tune Voicefixer on DPO-aligned model pseudo-labels | Voicefixer · real-world singing recordings | OVRL: 2.295 -> 2.756 | Improved |
Closest papers by meaning
- SpeechAlign: Aligning Speech Generation to Human Preferences (NeurIPS 2024 · similarity 0.89)
- Advancing Audio-Based Text Generation with Imbalance Preference Optimization (AAAI 2025 · similarity 0.88)
- Clean First, Align Later: Benchmarking Preference Data Cleaning for Reliable LLM Alignment (NeurIPS 2025 · similarity 0.87)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.87)
- RePO: Understanding Preference Learning Through ReLU-Based Optimization (NeurIPS 2025 · similarity 0.86)
- MM-RLHF: The Next Step Forward in Multimodal LLM Alignment (ICML 2025 · 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.
