AT-ADD: All-Type Audio Deepfake Detection Challenge Summary

View on arXiv ← Back to list

Authors: Yuankun Xie, Haonan Cheng, Jiayi Zhou, Xiaoxuan Guo, Tao Wang, Changhao Zhang, Jian Liu, Weiqiang Wang, Ruibo Fu, Xiaopeng Wang, Hengyan Huang, Xiaoying Huang, Long Ye, Guangtao Zhai

Published: 2026-08-14 12:28:54+00:00

Comment: Accepted to ACM MM 2026

AI Summary

This paper summarizes the ACM Multimedia 2026 AT-ADD Grand Challenge, focusing on two tracks: robust speech deepfake detection under realistic acoustic and channel variations, and type-agnostic detection across various audio types. The challenge highlights the effectiveness of self-supervised audio representations, data augmentation, multi-crop inference, and structured fusion or routing in achieving high Macro-F1 scores, while also identifying ongoing challenges in generalization and robustness.

Abstract

This paper summarizes the ACM Multimedia 2026 AT-ADD Grand Challenge on all-type audio deepfake detection. AT-ADD contains two tracks: robust speech deepfake detection under realistic acoustic and channel variations, and type-agnostic detection over speech, environmental sound, singing voice, and music. We describe the challenge tasks, dataset and evaluation-set design, official leaderboard results, and common design patterns observed in participating systems. The best Track 1 system achieved 90.71% Macro-F1 on the final evaluation set, while the best Track 2 system achieved 96.10% Macro-F1. The final submissions show that strong systems commonly combine large-scale self-supervised audio representations, data augmentation, multi-crop inference, and structured fusion or routing. The results also reveal remaining challenges in generalization to unseen generators, robustness to realistic speech-domain distortions, and balanced performance across heterogeneous audio types.

Key findings

The top-performing systems achieved high Macro-F1 scores (90.71% for Track 1 and 96.10% for Track 2) by combining large-scale self-supervised audio representations, extensive data augmentation, multi-crop inference, and structured fusion or routing. However, significant challenges remain in generalizing to unseen generators, maintaining robustness to realistic speech-domain distortions, and achieving balanced performance across heterogeneous audio types.

Approach

The challenge uses two tracks: Track 1 focuses on robust speech deepfake detection, evaluating performance against unseen generators and realistic perturbations. Track 2 aims for all-type audio deepfake detection, requiring a binary real/fake decision for speech, environmental sound, singing voice, and music with unknown input types. Participating systems commonly leverage large-scale self-supervised audio representations, data augmentation, multi-crop inference, and structured fusion or routing.

Datasets

AT-ADD Track 1 Dataset (including AISHELL-3, LibriTTS-R, LJSpeech, Common Voice, 3D-speaker, EchoFake), AT-ADD Track 2 Dataset (including AudioCaps, OpenCpop, M4Singer, KiSing, MusicCaps, AVQA, CompA-R, VocalSound, TUT2016, FMA, FortisAVQA).

Model(s)

W2V-BERT 2.0, AASIST, AASIST3, XLSR (0.3B, 1B, 2B), TSSDNet/ATSSDNet, EnvRobust-XLSR-AASIST, BEATs, EAT-large, wav2vec2-XLSR, Whisper-large-v3.

Author countries

China

← Previous