Long-form Audio Quality Description on QualiSpeech
51ROUGE-LLLM-judge dim.-wise
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| LLM-judge dim.-wiseReasoning strategy=LLM-judge dimension-wise rewards2026.03 | 51 | 0.83 | |
| Acc.+Sem. dim.-wiseReasoning strategy=Accuracy and Semantic Similarity dimension-wise rewards2026.03 | 47 | 0.8 | |
| SQ-LLMType=Baseline2026.03 | 45 | 0.82 | |
| QualiSpeech-FTType=Baseline2026.03 | 40 | 0.78 | |
| QualiSpeech-FT-SALMONNType=Baseline2026.03 | 38 | 0.75 | |
| x (Calibration-only)Strategy=Calibration-only ablation2026.03 | 12 | 0.19 |