Speech Synthesis on Cantonese elder-facing speech synthesis dataset
7.54SERCosyVoice2-Yue + GRPO w/ OPRL Stage 2
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| CosyVoice2-Yue + GRPO w/ OPRL Stage 2Backbone=CosyVoice2-Yue, Training=GRPO, OPRL=true, Stage=22026.06 | 7.54 | 3.86 | 54.62 | 1.17 | 2.06 | 4.8 | 0.85 | 1.06 | 5.36 | 19.62 | 0.77 | 3.78 | |
| CosyVoice2-Yue + GRPO w/ OPRL Stage 1Backbone=CosyVoice2-Yue, Training=GRPO, OPRL=true, Stage=12026.06 | 8.27 | 4.38 | 57.7 | 1.16 | 2.2 | 4.77 | 0.81 | 1.02 | 5.95 | 20.6 | 0.77 | 3.54 | |
| CosyVoice2-Yue + SFTBackbone=CosyVoice2-Yue, Training=SFT2026.06 | 9.93 | 4.91 | 53.33 | 1.27 | 2.24 | 4.87 | 0.83 | 1.02 | 5.42 | 19.62 | 0.77 | 3.55 | |
| CosyVoice2-Yue + GRPO w/o OPRLBackbone=CosyVoice2-Yue, Training=GRPO, OPRL=false2026.06 | 11.58 | 6.99 | 49.62 | 1.58 | 3.96 | 5.1 | 0.75 | 1.18 | 11.51 | 27.62 | 0.76 | 2.7 | |
| Ground Truth2026.06 | 13.79 | 9.08 | 100 | 0 | 0 | 0 | 1 | 1 | 5.43 | 19.27 | 1 | 3.45 | |
| CosyVoice2-Yue (base)Backbone=CosyVoice2-Yue2026.06 | 14.89 | 7.52 | 38.66 | 2.04 | 2.68 | 7.37 | 0.64 | 0.94 | 3.39 | 14.52 | 0.65 | 2.53 |