Single-class unlearning on LibriSpeech 100h
97.57FAOriginal
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| OriginalBackbone=HuBERT-Large2026.04 | 97.57 | 95.63 | — | — | — | — | |
| OriginalBackbone=Wav2Vec2-Large2026.04 | 91.23 | 94.88 | — | — | — | — | |
| Gradient AscentBackbone=Wav2Vec2-Large2026.04 | 21.47 | 81.62 | 79.93 | 32.7 | 58.2 | 612.8 | |
| Gradient AscentBackbone=HuBERT-Large2026.04 | 19.41 | 83.72 | 82.55 | 28.9 | 57.8 | 598.4 | |
| Fisher ForgettingBackbone=Wav2Vec2-Large2026.04 | 13.92 | 87.34 | 84.51 | 19.3 | 56.7 | 341.9 | |
| Fisher ForgettingBackbone=HuBERT-Large2026.04 | 12.36 | 89.14 | 87.09 | 16.1 | 55.9 | 335.1 | |
| Quantum-Inspired Audio UnlearningBackbone=Wav2Vec2-Large2026.04 | 7.41 | 92.88 | 90.74 | 7.8 | 54.9 | 168.4 | |
| Quantum-Inspired Audio UnlearningBackbone=HuBERT-Large2026.04 | 6.42 | 94.12 | 92.87 | 6.3 | 54 | 159.8 | |
| Graph-Propagated Projection UnlearningBackbone=Wav2Vec2-Large2026.04 | 0.19 | 94.63 | 96.12 | 0.6 | 53.1 | 29.4 | |
| Graph-Propagated Projection UnlearningBackbone=HuBERT-Large2026.04 | 0.16 | 95.38 | 97.21 | 0.5 | 52.6 | 25.9 |