Emotional Text-to-Speech on ESD plus (test)
4.15WEREmo-LiPO
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Emo-LiPOBackbone=CosyVoice-300M-Instruct, Lambda weighting (λ)=omitted2026.06 | 4.15 | 4.79 | 3.24 | 4.17 | 92.3 | 26.21 | 37.59 | |
| Emo-LiPOBackbone=CosyVoice-300M-Instruct2026.06 | 4.26 | 4.79 | 3.26 | 4.18 | 91.93 | 27.56 | 39.54 | |
| CosyVoiceBackbone=CosyVoice2026.06 | 4.47 | 4.71 | 3.16 | 4.3 | 81.87 | 25.1 | 29.9 | |
| Emo-DPO (E)Preference strategy=pairwise preferences across different emotion categories at the same intensity level2026.06 | 4.78 | 4.66 | 3.23 | 4.06 | 91.73 | 24.08 | 34.92 | |
| EmoVoiceMode=freestyle text prompting2026.06 | 5.4 | 4.79 | 3.29 | 4.27 | 89.84 | 20.56 | 28.51 | |
| Emo-DPO (I)Preference strategy=pairwise preferences across different intensity levels within the same emotion category2026.06 | 6.79 | 4.6 | 3.21 | 4 | 91.85 | 26.87 | 37.21 | |
| Emo-DPO (R)Preference strategy=randomly sampled pairwise preferences2026.06 | 12.78 | 4.37 | 3.04 | 3.9 | 91.52 | 24.77 | 33.46 |