Language Orthogonalization for Zero-Shot Cross-Lingual Audio Deepfake Detection

Authors: Minu Kim, Ji Sub Um, Hoirin Kim

Published: 2026-09-15 00:23:42+00:00

Comment: Submitted to ICASSP 2027

AI Summary

This research addresses the challenge of zero-shot cross-lingual audio deepfake detection, where detectors need to generalize to languages not seen during training. The authors propose language orthogonalization, a target-free ridge map method that removes language-dependent variations from self-supervised speech model (S3M) representations, thereby isolating spoof cues. This approach consistently reduces Equal Error Rate (EER) across various unseen languages and S3M backbones, with greater gains for more linguistically distant transfers.

Abstract

Audio deepfake detectors need to transfer to languages absent from training, as multilingual speech synthesis outpaces labeled anti-spoofing resources. While detectors increasingly rely on self-supervised speech models (S3Ms), these backbones encode language-dependent structure that confounds spoof cues. We address this confound through language orthogonalization, a target-free ridge map that removes S3M variation projected onto continuous language-identification (LID) embeddings. Across six languages, six S3M backbones, and all Leave-N-Out settings, it consistently reduces EER across unseen languages. Cross-lingual EER correlates with LID-space distance, where orthogonalization yields larger gains for more distant transfers.

Key findings

Language orthogonalization consistently reduces EER across all tested S3M backbones and Leave-N-Out settings, with an average relative gain of 9–17%. The improvements are most significant for linguistically distant transfers, confirming that removing language confounds is particularly effective when the target language is very different from the training languages. The method also effectively closes cross-lingual domain gaps by aligning bonafide representations into a shared space, thereby improving bonafide-spoof separation.

Approach

The core approach involves 'language orthogonalization', which uses a target-free ridge map to remove language-predictable components from S3M representations. This map is learned by projecting S3M variations onto continuous language-identification (LID) embeddings using only bonafide speech from source languages. This process transforms the S3M representations, isolating spoof artifacts from language confounds, and the transformed representations are then used for deepfake detection.

Datasets

SEA-Spoof

Model(s)

MMS-300M, XLS-R-300M, wav2vec2-Large-LV60, HuBERT-Large, WavLM-Large, mHuBERT-147 (as S3M feature extractors), ECAPA-TDNN pre-trained on VoxLingua107 (for LID embeddings), Logistic Regression Classifier (for detection)

Author countries

USA, South Korea

← Previous