Automatic Speech Recognition on Common Voice (ENG, DEU, ESP), CGN (VL), and GCND (DVL) OWSM v3.2 (test)
13.4Word Error Rate (ENG)Initial model
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Initial modelParams=–2026.06 | 13.4 | 11.3 | 11.3 | 37.9 | 86 | 31.98 | — | |
| Separate ModelParams=244.8M2026.06 | 13.4 | 11.3 | 11.3 | 15.2 | 29.7 | 16.17 | 0 | |
| CSSVDParams=8.9M2026.06 | 15.3 | 14.9 | 13.2 | 21.5 | 59.2 | 24.82 | -2.2 | |
| LoRA + FTAParams=9.3M2026.06 | 17.3 | 22 | 16.3 | 20.3 | 57.1 | 26.58 | -4.8 | |
| EWC-LoRAParams=9.3M2026.06 | 30.7 | 72.9 | 30 | 23.2 | 43.7 | 40.06 | -26.1 | |
| BiLoRAParams=9.3M2026.06 | 33.5 | 71.1 | 34.6 | 36.2 | 35.9 | 42.25 | -30.4 | |
| Full Fine-TuningParams=244.8M2026.06 | 45.9 | 91.5 | 43.4 | 34.2 | 29.7 | 48.95 | -41 | |
| LoRAParams=9.3M2026.06 | 90.5 | 100 | 98 | 53.5 | 29.9 | 74.7 | -72.5 | |
| SSVDParams=9.0M2026.06 | 97.6 | 100 | 93.7 | 49.9 | 32.7 | 74.76 | -71.7 |