Phoneme-to-Grapheme on CV-Lang10
8.04WER (en)Qwen3-4B-Base + DANP
Evaluation Results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-4B-Base + DANPBackbone=Qwen3-4B-Base, Robustness=DANP, top-K marginalization decoding=true, K=82026.03 | 8.04 | 6.29 | 10.31 | 6.74 | 5.14 | 5.33 | 3.91 | 14.56 | 7.84 | 16.41 | 8.45 | 8.11 | |
| Qwen3-4B-Base + S-SKM + oversamplingBackbone=Qwen3-4B-Base, Robustness=S-SKM, Sampling=oversampling, top-K marginalization decoding=true, K=82026.03 | 8.22 | 6.12 | 10.62 | 7.17 | 2.82 | 5.48 | 3.9 | 10.59 | 7.33 | 14.34 | 7.66 | 8.22 | |
| Qwen3-4B-Base + S-SKMBackbone=Qwen3-4B-Base, Robustness=S-SKM, top-K marginalization decoding=true, K=82026.03 | 8.24 | 6.15 | 10.53 | 7.18 | 7.07 | 6.03 | 4.95 | 15.88 | 9.78 | 20.65 | 9.64 | 8.41 | |
| Qwen3-4B-BaseBackbone=Qwen3-4B-Base, top-K marginalization decoding=true, K=82026.03 | 8.26 | 5.84 | 10.44 | 6.84 | 10.07 | 6.05 | 5.98 | 17.94 | 10.92 | 23.25 | 10.56 | 8.46 | |
| WFST baselinetop-K marginalization decoding=true, K=82026.03 | 8.8 | 7.02 | 14.02 | 8.16 | 0.94 | 6.22 | 1.46 | 5.06 | 7.05 | 6.92 | 6.56 | 9.2 |