Supervised Post-training of Speech Foundation Models for Robust Adaptation in Speech Deepfake Detection

2026-06-23

AI Summary

This work addresses the challenge that existing self-supervised speech foundation models struggle to robustly detect deepfake audio due to a mismatch between their pretraining objectives and the characteristics of synthetic artifacts. To overcome this limitation, the authors propose a hybrid frame-level post-training strategy that leverages frame-wise supervision signals—without relying on data augmentation—to guide the model in learning local time-frequency inconsistencies inherent in spoofed speech. By integrating hybrid frame perturbations with a self-supervised speech foundation model during targeted post-training, the method achieves a single-model equal error rate (EER) of 4.50% on ASVspoof5. Furthermore, on the ASVspoof2021 LA/DF tasks, it yields an exceptionally small EER gap of only 0.16% between the LA and DF subsets, significantly outperforming current state-of-the-art approaches.

Abstract

Large speech foundation models have shown strong potential for speech deepfake detection, but direct fine-tuning is limited by a mismatch between self-supervised pre-training objectives and spoof-specific artifacts. To address this, we propose a mix-frame post-training strategy to create localized spoof-oriented perturbations and use frame-level supervision to encourage the SSL model to learn local inconsistencies that are critical for robust spoof detection. On ASVspoof5, we achieve state-of-the-art EER 4.50% for a single model without data augmentation. On ASVspoof2021 LA/DF, it further achieves only 0.16\% absolute EER gap between LA and DF, indicating strong and balanced robustness across distinct distortion conditions. These results show that supervised post-training provides an effective and practical way to adapt speech foundation models for robust deepfake detection.

Problem

• speech deepfake detection
• foundation models
• supervised post-training
• spoof artifacts
• robust adaptation

Innovation

• supervised post-training
• speech foundation models
• deepfake detection
• frame-level supervision
• spoof-oriented perturbations