AT-ADD: A Benchmark and Challenge for Robust and All-Type Audio Deepfake Detection

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: 2026-08-24 16:11:43+00:00

AI Summary

This paper introduces AT-ADD, a new large-scale benchmark and challenge for robust and all-type audio deepfake detection. It features two tracks: robust speech deepfake detection under diverse conditions (Track 1) and type-agnostic real/fake detection across speech, sound, singing, and music (Track 2). The research details dataset construction, evaluation protocols, and analyzes top-performing systems from the ACM Multimedia 2026 Grand Challenge, highlighting key strategies for generalization and remaining challenges.

Abstract

Recent audio generation models can synthesize high-fidelity speech, environmental sound, singing voice, and music, creating new risks for multimedia trust. Existing audio deepfake detection (ADD) benchmarks remain predominantly speech-centric and often underrepresent realistic channel variation and diverse audio types. This paper presents AT-ADD, a large-scale benchmark and challenge designed to evaluate both robust speech deepfake detection and all-type audio deepfake detection. Track 1 evaluates binary speech detection under unseen generators, diverse recording conditions, signal perturbations, and replay effects. Track 2 evaluates type-agnostic real/fake detection over speech, sound, singing, and music when the audio type is unknown at test time. We detail the dataset construction, evaluation protocol, and reproducible baselines, and analyze the final systems submitted to the ACM Multimedia 2026 Grand Challenge. The strongest official baseline obtains 76.73% and 79.47% Macro-F1 on the Track 1 and Track 2 evaluation sets, respectively, whereas the winning challenge systems reach 90.71% and 96.10%. Beyond aggregate rankings, sample-level analysis of the top five submissions examines generator- and type-level difficulty, cross-system error complementarity, and ranking stability. The results show that large-scale self-supervised representations, condition-aware augmentation, multi-crop inference, and structured fusion or routing are central to generalization, while generator-specific robustness and consistent performance across diverse audio types remain unresolved.

Key findings

Large-scale self-supervised representations, condition-aware augmentation, multi-crop inference, and structured fusion/routing are crucial for generalization in audio deepfake detection. However, generator-specific robustness and consistent performance across diverse audio types, especially speech under realistic conditions and certain non-speech types like sound, remain significant unresolved challenges, indicating the need for more advanced, adaptable detection methods.

Approach

The challenge uses a comprehensive dataset spanning diverse audio types, generators, and real-world conditions for training and evaluation. Detection models leverage self-supervised representations, condition-aware augmentation, multi-crop inference, and structured fusion or routing for improved generalization. The two tracks progressively evaluate robustness to channel variations and generalization across heterogeneous audio types.

Datasets

AT-ADD Track 1 (robust speech deepfake detection) and AT-ADD Track 2 (all-type audio deepfake detection). Track 1 includes real speech from AISHELL-3, LibriTTS-R, LJSpeech, Common Voice, 3D-Speaker, EchoFake, and internal recordings, and fake speech from 47 generators. Track 2 expands to environmental sound (AudioCaps, AVQA, CompA-R, VocalSound, TUT2016), singing voice (OpenCpop, M4Singer, KiSing), and music (MusicCaps, FMA, FortisAVQA), utilizing 68 generators.

Model(s)

Spec-ResNet, AASIST, FT-XLSR-AASIST, WPT-XLSR-AASIST, Qwen2.5-Omni-3B, Qwen2.5-Omni-7B. Top challenge systems included ensembles of W2V-BERT, XLSR, and EAT-based models, often incorporating specialized branches, multi-crop inference, and various augmentation techniques.

Author countries

China

← Previous