WaveSP-Net: Learnable Wavelet-Domain Sparse Prompt Tuning
for Speech Deepfake Detection
Xi Xuan1, Xuechen Liu2, Wenxin Zhang3,5, Yi-Cheng Lin4, Xiaojian Lin6, Tomi Kinnunen1
1 University of Eastern Finland    2 National Institute of Informatics
3 University of Chinese Academy of Sciences    4 National Taiwan University
5 University of Toronto    6 Tsinghua University
DEMO Real vs Fake Audio Samples
Real Speech Sample REAL
🔹 Real0001.wav – Detection Result: Real
(WaveSP-Net inference fake probability: 0.0323)
Human voice, genuine recording.
Real waveform
Figure R1. Waveform (Real)
Real spectrogram
Figure R2. Spectrogram (Real)
Fake Speech Sample FAKE
🔸 Fake0001.wav – Detection Result: Fake
(WaveSP-Net inference fake probability: 0.8272)
AI-generated / synthesized voice.
Fake waveform
Figure F1. Waveform (Fake)
Fake spectrogram
Figure F2. Spectrogram (Fake)
CITATION Reference
@misc{xuan2025wavespnet, title={WaveSP-Net: Learnable Wavelet-Domain Sparse Prompt Tuning for Speech Deepfake Detection}, author={Xi Xuan and Xuechen Liu and Wenxin Zhang and Yi-Cheng Lin and Xiaojian Lin and Tomi Kinnunen}, year={2025}, eprint={2510.05305}, archivePrefix={arXiv}, primaryClass={eess.AS}, url={https://arxiv.org/abs/2510.05305}, }