Hidden-Domain Routing for All-Type Audio Deepfake Detection

Authors: Yifan Gao, Yao Tian, Hongbin Suo, Haonan Lu

Published: 2026-08-01 07:32:54+00:00

Comment: Accepted by ACMMM 2026

AI Summary

This paper introduces a hidden-domain routing system for all-type audio deepfake detection (AT-ADD Track2), addressing the challenge of identifying deepfakes across diverse audio types (speech, environmental sound, singing voice, music) when the audio type is unknown at inference. The system first routes the input audio to a specific expert based on its predicted type and then interprets detector scores within that specialized branch. This approach significantly outperforms baselines, achieving a 96.10% Track2 Macro-F1 and ranking first in the official AT-ADD Track2 evaluation.

Abstract

All-type audio deepfake detection requires authenticity decisions across speech, environmental sound, singing voice, and music, while the audio type is unavailable at inference time. In AT-ADD Track2, this setting creates a hidden audio-domain condition: the binary real/fake label is shared across domains, but representation structure and detector-score behavior vary with audio type. We present a closed-condition routed system that first recovers the hidden audio domain and then interprets detector scores within the selected branch. The AudioType-BEATs-6s Router estimates audio type from a 6-second window; speech inputs are handled by the Speech-XLSR Expert, while sound, singing, and music rely on EAT-based general-audio experts with branch-local score interpretation. Development-set representation analysis, router-family comparisons, and component results show audio-domain separation and complementary detector strengths across audio types. On the official AT-ADD Track2 final evaluation, the system achieves 96.10% Track2 Macro-F1 and ranks first on the final leaderboard, with type-wise Macro-F1 scores of 88.07%, 98.18%, 99.07%, and 99.08% for speech, sound, singing, and music, respectively. These results support recovering the hidden audio domain before interpreting detector scores in all-type audio deepfake detection.

Key findings

The proposed hidden-domain routing system achieved a 96.10% Track2 Macro-F1, ranking first in the AT-ADD Track2 final evaluation. It demonstrated significant improvements over baselines, especially in non-speech domains, highlighting the importance of domain-conditioned detection. Speech detection remained the primary performance bottleneck, suggesting areas for future improvement.

Approach

The system employs an AudioType-BEATs-6s Router to estimate the audio type (speech, sound, singing, or music) from a 6-second window. Based on this predicted type, the system routes the audio to a specialized expert: Speech-XLSR Expert for speech, and EAT-based general-audio experts for sound, singing, and music. Each expert then applies branch-local score interpretation and decision rules to determine if the audio is real or fake.

Datasets

AT-ADD Track2 dataset (training and development splits)

Model(s)

AudioType-BEATs-6s Router (BEATs-based MLP), Speech-XLSR Expert (XLS-R frontend + AASIST backend), SoundMusic-EAT Expert (EAT-large frontend + AASIST backend), Singing-EAT Expert (EAT-large frontend + AASIST backend)

Author countries

China

← Previous