Automatic Speech Recognition on Common Voice 16.1 (test)
7.8WER (de)w/ MoE, modality-aware (top-1 per modality)
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| w/ MoE, modality-aware (top-1 per modality)Backbone=Conformer, # Active Params=113M2026.02 | 7.8 | 16 | 7.8 | 12.3 | 9.1 | 10.6 | — | |
| w/ MoE (no modality grouping, top-2)Backbone=Conformer, # Active Params=113M2026.02 | 8.4 | 16.6 | 8.3 | 13.1 | 9.8 | 11.2 | — | |
| AED, 17-layer enc / 6-layer decBackbone=Conformer, # Active Params=139M2026.02 | 9.3 | 17.8 | 9.2 | 14.1 | 10.5 | 12.2 | — | |
| dec-only 17-layerBackbone=Conformer, # Active Params=113M2026.02 | 10.1 | 18.9 | 10 | 15 | 11.8 | 13.2 | — | |
| dec-only 17-layerBackbone=Transformer, # Active Params=64M2026.02 | 12.5 | 21.9 | 12 | 17.5 | 14.4 | 15.7 | — | |
| Final Aligner + InterCTCSystem=+ InterCTC2026.06 | — | — | — | — | — | — | 11.2 | |
| Final Aligner + InterCTC + InterAlignerSystem=+ InterAligner2026.06 | — | — | — | — | — | — | 10.9 | |
| Final Aligner onlySystem=Final Aligner only2026.06 | — | — | — | — | — | — | 12.4 |