Speech-to-Speech Translation on S2ST English-to-Chinese (test)
54.26BLEUGemini + Base
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Gemini + BaseS2T Backbone=Gemini 2.5 Pro, TTS Backbone=CosyVoice3 Base, System Architecture=Cascaded2026.06 | 54.26 | 3.6 | 15.7 | 16.6 | |
| Qwen2.5-OmniType=Multimodal LLM, Access=Open-source2026.06 | 51.35 | 3.53 | 25.7 | 21.1 | |
| GPT-4o-audioType=Multimodal LLM, Access=Closed-source2026.06 | 50.63 | 3.46 | 25.2 | 19.4 | |
| ProposedTTS Backbone=CosyVoice3, TTS fine-tuning=true, System Architecture=Complete pipeline2026.06 | 47.35 | 3.68 | 60.8 | 58.3 | |
| StressTransfer + BaseArchitecture=from [17], TTS Backbone=CosyVoice3 Base, TTS fine-tuning=false2026.06 | 46.34 | 3.54 | 24.6 | 22.3 |