Text-to-Speech on Seed-TTS-Eval (test-hard)
5.09Character Error Rate (CER)CosyVoice 3.0-0.5B
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| CosyVoice 3.0-0.5BRL=LM-DiffRO2026.06 | 5.09 | 0.75 | 0.809 | — | — | |
| CosyVoice 3.0-1.5BRL=LM-DiffRO2026.06 | 5.66 | 0.75 | 0.81 | — | — | |
| CosyVoice 3.0-1.5BRL=without2026.06 | 5.83 | 0.758 | 0.816 | — | — | |
| CosyVoice 3.0-0.5B-2512 + LM w. RLRL=FM-GRPO, Step=95452026.06 | 5.89 | 0.79 | 0.841 | 3.971 | 3.544 | |
| CosyVoice 3.0-0.5B-2512 + LM w. RLRL=without, Step=02026.06 | 6.01 | 0.756 | 0.803 | 3.871 | 3.39 | |
| CosyVoice 3.0-0.5BRL=without2026.06 | 6.08 | 0.758 | 0.815 | — | — | |
| CosyVoice 2.0RL=without2026.06 | 6.83 | 0.724 | 0.776 | — | — | |
| CosyVoice 3.0-0.5B-2512RL=FM-GRPO, Step=95452026.06 | 7.08 | 0.792 | 0.844 | 3.976 | 3.559 | |
| CosyVoice 3.0-0.5B-2512RL=without, Step=02026.06 | 7.32 | 0.757 | 0.808 | 3.875 | 3.393 | |
| Seed-TTSRL=without2026.06 | 7.59 | 0.776 | — | — | — | |
| F5-TTSRL=FM-GRPO, Step=12892026.06 | 7.86 | 0.741 | 0.791 | 3.973 | 3.562 | |
| F5-TTSRL=without2026.06 | 8.67 | 0.713 | 0.762 | — | — | |
| F5R-TTSRL=FM-GRPO2026.06 | 8.79 | 0.718 | — | — | — | |
| F5-TTSRL=without, Step=02026.06 | 9 | 0.73 | 0.759 | 3.765 | 3.304 | |
| F5R-TTSRL=without2026.06 | 9.87 | 0.702 | — | — | — | |
| MaskGCTRL=without2026.06 | 10.27 | 0.748 | 0.72 | — | — | |
| CosyVoiceRL=without2026.06 | 11.75 | 0.709 | 0.755 | — | — |