Speech-to-speech translation on CVSS-C + SpeechMatrix FR-EN (test)
22.72BLEUAudio-LLM filter
Evaluation results
| Method | Links | |
|---|---|---|
| Audio-LLM filterRetained Pairs=477,7732026.06 | 22.72 | |
| Stage II onlyRetained Pairs=577,7492026.06 | 22.49 | |
| LLM filtering (LLaMA-70B)Retained Pairs=577,749, Threshold=≥802026.06 | 22.42 | |
| LLM filtering (70B)Retained Pairs=477,7732026.06 | 22.32 | |
| BLEURT filteringRetained Pairs=477,7732026.06 | 22.09 | |
| BLEURT+LLM ablationRetained Pairs=469,7122026.06 | 22.08 | |
| BLEURT filteringRetained Pairs=434,254, Threshold=≥0.72026.06 | 21.94 | |
| Stage II onlyRetained Pairs=477,7732026.06 | 21.91 | |
| BLASER 2.0-QE ablationRetained Pairs=482,5912026.06 | 21.81 | |
| BLASER 2.0-QE filteringRetained Pairs=477,7732026.06 | 21.71 | |
| Audio Flamingo 3Retained Pairs=405,4682026.06 | 21.53 | |
| SNR filteringRetained Pairs=474,570, Threshold=≥30 dB2026.06 | 21.46 | |
| Keep 20% SpeechMatrixRetained Pairs=614,2652026.06 | 21.32 | |
| Random keepRetained Pairs=477,7732026.06 | 21.27 | |
| SNR filteringRetained Pairs=553,800, Threshold=≥25 dB2026.06 | 21.15 | |
| MOS filteringRetained Pairs=479,132, Threshold=≥2.02026.06 | 21.06 | |
| MOS filteringRetained Pairs=408,410, Threshold=≥2.22026.06 | 20.72 | |
| LLM filtering (LLaMA-8B)Retained Pairs=431,319, Threshold=≥802026.06 | 19.95 | |
| CVSS-C (reproduced)Retained Pairs=207,3652026.06 | 15.93 | |
| CVSS-C (FR-EN)Retained Pairs=207,3652026.06 | 15.44 |