Dysfluency Detection on SEP-28k-E (test)
35F1 (Block)WavLM + attn. pool
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| WavLM + attn. poolBackbone=WavLM, Pooling method=attn. pool2026.06 | 35 | 82 | 47 | 42 | 74 | 78 | |
| Whisper + max. poolBackbone=Whisper, Pooling method=max. pool2026.06 | 35 | 74 | 50 | 53 | 80 | 78 | |
| Whisper + attn. poolBackbone=Whisper, Pooling method=attn. pool2026.06 | 35 | 82 | 49 | 53 | 78 | 78 | |
| WavLM + max. poolBackbone=WavLM, Pooling method=max. pool2026.06 | 34 | 83 | 31 | 51 | 72 | 78 | |
| TRILANG-LLTraining data=TRILANG-LL, Backbone (W2V2 base-weights)=XLSR-532026.03 | 33 | 77 | 51 | 53 | 71 | — | |
| Haas et al.2026.06 | 33 | 77 | 51 | 53 | 71 | — | |
| Baseline (SEP-28k-E training)Training data=SEP-28k-E, Backbone (W2V2 base-weights)=W2V2-base2026.03 | 32 | 77 | 53 | 53 | 64 | — | |
| wav2vec 2.0 + attn. poolBackbone=wav2vec 2.0, Pooling method=attn. pool2026.06 | 31 | 71 | 39 | 44 | 41 | 69 | |
| Miyahara et al.2026.06 | 30 | 78 | 53 | 46 | 67 | 82 | |
| wav2vec 2.0 + max. poolBackbone=wav2vec 2.0, Pooling method=max. pool2026.06 | 30 | 40 | 25 | 46 | 34 | 70 | |
| TRILANGTraining data=TRILANG, Backbone (W2V2 base-weights)=XLSR-532026.03 | 26 | 77 | 52 | 47 | 61 | — | |
| AS-70 (XLSR-53)Training data=AS-70, Backbone (W2V2 base-weights)=XLSR-532026.03 | 22 | 46 | 34 | 27 | 28 | — |