SpeeCheck: Self-Contained Speech Integrity Verification via Embedded Acoustic Fingerprints

ICLR 2026 Conference Submission15565 Authors

Keywords: speech integrity verification, tampering attacks, audio fingerprint, audio watermark, contrastive learning

Abstract: Advances in audio editing have made public speeches increasingly vulnerable to malicious tampering, raising concerns for social trust. Existing speech tampering detection methods remain insufficient: they often rely on external references or fail to balance sensitivity to attacks with robustness against benign operations like compression. To tackle these challenges, we propose SpeeCheck, the first self-contained speech integrity verification framework. SpeeCheck can (i) effectively detect tampering attacks, (ii) remain robust under benign operations, and (iii) enable direct verification without external references. Our approach begins with utilizing multiscale feature extraction to capture speech features across different temporal resolutions. Then, it employs contrastive learning to generate fingerprints that can detect modifications at varying granularities. These fingerprints are designed to be robust to benign operations, but exhibit significant changes when malicious tampering occurs. To enable self-contained verification, these fingerprints are embedded into the audio itself via a watermark. Finally, during verification, SpeeCheck retrieves the fingerprint from the audio and checks it with the embedded watermark to assess integrity. Extensive experiments demonstrate that SpeeCheck reliably detects tampering while maintaining robustness against common benign operations. Real-world evaluations further confirm its effectiveness in verifying speech integrity. The code and demo are available at https://speecheck.github.io/SpeeCheck/.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 15565