Research map / Vision-language and multimodal / Cross-modal learning
Detect All-Type Deepfake Audio: Wavelet Prompt Tuning for Enhanced Auditory Perception
Yuankun Xie, Ruibo Fu, Xiaopeng Wang, Zhiyong Wang, Songjun Cao, Long Ma, Haonan Cheng, Long Ye
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add wavelet prompt tokens (4 WPT + 6 PT) to prompt tuning (ablation) | PT-XLSR-AASIST · speech-trained, avg EER over 4 types | EER (%): 30.79 -> 26.86 | Improved |
| use WPT instead of full fine-tuning with 458x fewer trainable parameters | XLSR-AASIST · speech-trained | EER (%): 27.68 -> 26.86 | Improved |
| co-train on all four audio types with WPT | XLSR-AASIST · all-type ADD | avg EER (%): 4.98 (FT) -> 3.58 (WPT) | Improved |
| apply WPT to other SSL front-ends | WavLM-AASIST, MERT-AASIST · all-type ADD co-trained | avg EER (%): 10.81 -> 7.66 (WavLM); 13.21 -> 9.70 (MERT) | Improved |
| increase prompt tokens from 10 to 100 or 200 (ablation) | PT-XLSR-AASIST · speech-trained | avg EER (%): 30.79 -> 31.28 (100), 33.36 (200) | Hurt |
| insert prompts in every layer instead of first layer only (ablation) | PT-XLSR-AASIST · speech-trained | avg EER (%): 32.54 -> 30.79 | Improved |
| keep prompt token in last layer instead of deleting it (ablation) | PT-XLSR-AASIST · speech-trained | avg EER (%): 33.07 -> 30.79 | Improved |
| use all 10 tokens as wavelet tokens (ablation) | WPT-XLSR-AASIST · speech-trained | avg EER (%): 26.86 -> 32.54 | Hurt |
| fine-tune SSL instead of freezing it on music-trained CMs (ablation) | XLSR/WavLM-AASIST · music-trained | EER (%): FT-XLSR avg 45.91 vs FR-XLSR avg 35.63 | Hurt |
| fine-tune SSL instead of freezing it on speech, sound, singing-trained CMs | XLSR-AASIST · in-domain EER | EER (%): 1.28 -> 0.38 (speech); 2.40 -> 0.21 (sound) | Improved |
| train on a single audio type | all CMs · cross-type ADD | EER (%): speech-trained AASIST: 1.48 in-domain vs 48.32 on sound | Hurt |
Closest papers by meaning
- Adaptive Compression and Targeted Perturbation: A Unified Framework for Generalized Audio Deepfake Detection (NeurIPS 2026 · similarity 0.88)
- I Can Hear You: Selective Robust Training for Deepfake Audio Detection (ICLR 2025 · similarity 0.88)
- What to Remember: Self-Adaptive Continual Learning for Audio Deepfake Detection (AAAI 2024 · similarity 0.87)
- Region-Based Optimization in Continual Learning for Audio Deepfake Detection (AAAI 2025 · similarity 0.86)
- Multi-modal Deepfake Detection via Multi-task Audio-Visual Prompt Learning (AAAI 2025 · similarity 0.85)
- Multi-View Collaborative Learning Network for Speech Deepfake Detection (AAAI 2025 · similarity 0.85)
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.
