Speech-to-speech translation on CVSS-C ES → EN (test)
88.54ASR-BLEUGround Truth
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Ground Truth2025.11 | 88.54 | — | |
| RosettaSpeech (Paralleled, Fine-tuned)Protocol=Non Zero-Shot, Training setup=Paralleled, Initialization=Fine-tuned, Model scope=Pair-specific models2025.11 | 33.05 | 35.23 | |
| RosettaSpeech (Unparalleled)Protocol=Zero-Shot, Training setup=Unparalleled2025.11 | 29.86 | 32.64 | |
| RosettaSpeech (Paralleled, Fine-tuned)Protocol=Non Zero-Shot, Training setup=Paralleled, Initialization=Fine-tuned, Model scope=Single model for all language pairs2025.11 | 29.35 | 30.92 | |
| StreamSpeechProtocol=Non Zero-Shot, Decoding strategy=Beam search (size 10)2025.11 | 27.25 | 30.35 | |
| StreamSpeechProtocol=Non Zero-Shot, Decoding strategy=Greedy search2025.11 | 26.16 | 28.97 | |
| UnitYProtocol=Non Zero-Shot, Decoding strategy=Beam search (size 10)2025.11 | 24.95 | — | |
| ComSpeechProtocol=Zero-Shot2025.11 | 24.8 | 26.51 | |
| UnitYProtocol=Non Zero-Shot, Decoding strategy=Greedy search2025.11 | 23.93 | — | |
| Translatotron 2Protocol=Non Zero-Shot, Decoding strategy=Beam search (size 10)2025.11 | 22.93 | — | |
| Translatotron 2Protocol=Non Zero-Shot, Decoding strategy=Greedy search2025.11 | 22.35 | 25.82 | |
| DASpeechProtocol=Non Zero-Shot2025.11 | 21.37 | — | |
| S2UTProtocol=Non Zero-Shot, Decoding strategy=Beam search (size 10)2025.11 | 18.53 | — | |
| S2UTProtocol=Non Zero-Shot, Decoding strategy=Greedy search2025.11 | 16.94 | — | |
| RosettaSpeech (Paralleled, Scratch)Protocol=Non Zero-Shot, Training setup=Paralleled, Initialization=Scratch2025.11 | 9.38 | 23.62 | |
| TranslatotronProtocol=Non Zero-Shot2025.11 | 8.72 | — | |
| Qwen3-0.6B2025.11 | — | 34.35 |