Research map / Label-efficient and robust learning / Knowledge distillation
Cross-Modal Distillation for Speaker Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use margin-based feature distillation instead of no distillation | ResNet34 speaker recognition · VoxCeleb1 original test | EER: 1.71% -> 1.51% | Helped |
| Use margin-based relation distillation instead of no distillation | ResNet34 speaker recognition · VoxCeleb1 original test | EER: 1.71% -> 1.59% | Helped |
| Use margin-based response distillation instead of no distillation | ResNet34 speaker recognition · VoxCeleb1 original test | EER: 1.71% -> 1.57% | Helped |
| Add quality-based adaptive weighting to feature-based MKD | ResNet34 with feature-based MKD · VoxCeleb1 original test | minDCF: 0.178 -> 0.165 | Helped |
| Add quality-based adaptive weighting to feature-based MKD | ResNet34 with feature-based MKD · VoxCeleb1 noisy test (∆db = 5) | EER: 5.51% -> 4.77% | Helped |
| Use VGSR cross-modal distillation | X-Vector · VoxCeleb1 original test | EER: 7.20% -> 6.29% | Helped |
| Use VGSR cross-modal distillation | ResNet34 · CN-Celeb test | EER: 14.49% -> 13.03% | Helped |
| Set teacher-feature mixing ratio α to 0.2 rather than the best reported α=0.6 (ablation) | VGSR with ResNet34 · VoxCeleb1 original test | EER: 1.51% -> 1.68% | Hurt |
| Set teacher-feature mixing ratio α to 0.4 rather than the best reported α=0.6 (ablation) | VGSR with ResNet34 · VoxCeleb1 original test | EER: 1.51% -> 1.67% | Hurt |
| Set teacher-feature mixing ratio α to 0.5 rather than α=0.6 (ablation) | VGSR with ResNet34 · VoxCeleb1 original test | EER: 1.51% -> 1.52% | No clear effect |
| Set teacher-feature mixing ratio α to 0.8 rather than the best reported α=0.6 (ablation) | VGSR with ResNet34 · VoxCeleb1 original test | EER: 1.51% -> 1.68% | Hurt |
| Set feature-distillation margin to m=0 rather than m=cos(30°) (ablation) | VGSR with ResNet34 · VoxCeleb1 original test | EER: 1.51% -> 1.66% | Hurt |
Closest papers by meaning
- C2KD: Bridging the Modality Gap for Cross-Modal Knowledge Distillation (CVPR 2024 · similarity 0.86)
- XKD: Cross-Modal Knowledge Distillation with Domain Alignment for Video Representation Learning (AAAI 2024 · similarity 0.85)
- Probabilistic Conformal Distillation for Enhancing Missing Modality Robustness (NeurIPS 2024 · similarity 0.83)
- Over-parameterized Student Model via Tensor Decomposition Boosted Knowledge Distillation (NeurIPS 2024 · similarity 0.83)
- Heuristic-free Knowledge Distillation for Streaming ASR via Multi-modal Training (AAAI 2025 · similarity 0.83)
- Int*-Match: Balancing Intra-Class Compactness and Inter-Class Discrepancy for Semi-Supervised Speaker Recognition (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.