Research map / Large language models / Language, safety and interpretability
Bridging the Stability-Expressivity Gap: Synthetic Data Scaling and Preference Alignment for Low-Resource Spoken Language Models
Yizhong Geng, Yanliang Li, Jinghan Yang, Tianhan Jiang, Boxun An, Ya Li, Xiaoyu Shen
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 expressivity objective (ablation) | DGSA · Thai alpha=80% | NMOS: 4.4 -> 3.7 without it (ΔNMOS -0.7) | Improved |
| add stability objective (ablation) | DGSA · Thai alpha=80% | WER (%): 46.5 without it vs 38.9 | Improved |
| use identity-consistent pairs instead of random pairs (ablation) | DGSA · Thai alpha=80% | WER (%): 42.8 without it vs 38.9 | Improved |
| add dynamic weight scaling (ablation) | DGSA · Thai alpha=80% | NMOS: 4.2 -> 4.4 (ΔNMOS -0.2 when removed) | Improved |
| add DPO loss (ablation) | TDSC · Lao | NMOS: 3.4 without it vs 3.9 (ΔNMOS -0.5) | Improved |
| add multi-temperature exploration (ablation) | TDSC · Lao | Hp (bits): 10.42 vs 10.05 with T=1.0 only | Improved |
| add length filter (ablation) | TDSC · Lao | WER (%): 33.5 without it vs 29.8 | Improved |
| add temperature curriculum (ablation) | TDSC · Lao | NMOS: 3.6 without it vs 3.9 (ΔNMOS -0.3) | Improved |
| add repetition filter (ablation) | TDSC · Lao | WER (%): 31.2 without it vs 29.8 | Improved |
| increase synthetic ratio from 3% to 50% | SFT · Thai | NMOS: 3.82 -> 4.51 | Improved |
| increase synthetic ratio from 50% to 100% | SFT · Thai | NMOS: 4.51 -> 3.08 | Hurt |
| increase synthetic ratio from 50% to 100% | SFT · Thai | Repetition rate (%): 2.16 -> 9.83 | Hurt |
Closest papers by meaning
- SpeechAlign: Aligning Speech Generation to Human Preferences (NeurIPS 2024 · similarity 0.88)
- RVCBench: Benchmarking Robustness of Voice Cloning Across Modern Audio Generation Models (NeurIPS 2026 · similarity 0.86)
- Speech Self-Supervised Learning Using Diffusion Model Synthetic Data (ICML 2024 · similarity 0.86)
- Aligning Spoken Dialogue Models from User Interactions (ICML 2025 · similarity 0.86)
- Continuous Autoregressive Modeling with Stochastic Monotonic Alignment for Speech Synthesis (ICLR 2025 · similarity 0.86)
- Scaling Speech-Text Pre-training with Synthetic Interleaved Data (ICLR 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.
