AI summaryⓘ
The authors created a big test called AT-ADD to see how well computer programs can detect fake sounds, including speech, singing, music, and other noises. They set up two tasks: one focused on spotting fake speech under tricky conditions, and the other tested detecting fakes in any audio type without knowing what it was beforehand. They shared how they built the test, ran it, and analyzed the results from many systems in a competition. The top systems did much better than the basic versions, but some challenges like handling all audio types equally well still remain. The authors also found that certain techniques, like using large pretrained models and smart ways of mixing data, help in making these fake sound detectors better.
audio deepfake detectionspeech synthesisself-supervised learningbenchmark datasetmacro-F1 scoreaudio augmentationmulti-crop inferencegenerator robustnesscross-system analysisACM Multimedia Grand Challenge
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
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.