AT-ADD: A Benchmark and Challenge for Robust and All-Type Audio Deepfake Detection
Authors: Yuankun Xie, Haonan Cheng, Jiayi Zhou, Xiaoxuan Guo, Tao Wang, Changhao Zhang, Jian Liu, Weiqiang Wang, +6 more
Organizations: Communication University of China, Beijing, China · Ant Group · Machine Intelligence, Ant Group, Shanghai, China · Institute of Automation, Chinese Academy of Sciences, Beijing, China · Beijing Institute of Technology, Beijing, China · Shanghai Jiao Tong University, Shanghai, China
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.
Figures & tables
Fig. 1: AT-ADD challenge overview.
Dataset / track
Year
Venue
Speakers
Language
Generators
Types
Clips
ASVspoof 2019 LA [ 76 ]
2019
INTERSPEECH
107
1
19
1
121,461
ADD 2022 LF [ 123 ]
2022
ICASSP
80+
1
–
1
165,607
ADD 2023 FG-D [ 124 ]
2023
IJCAI
–
1
22
1
175,601
ASVspoof 5 T1 [ 106 ]
2024
INTERSPEECH
1,922
1
32
1
1,004,081
SVDD 2024 Ctr [ 131 ]
2024
SLT
164
2
14
1
220,798
ESDD T1 [ 125 ]
2026
ICASSP
–
–
7
1
184,765
TABLE I: Comparison with representative audio deepfake detection challenge datasets.
Split
T1
T2
Speech
Sound
Singing
Music
Total
Train
49,575
49,575
39,840
36,000
21,366
146,781
Dev
49,734
49,734
19,929
16,000
5,406
91,069
Progress
29,269
28,813
5,729
4,872
6,461
45,875
Eval
146,346
144,078
28,593
24,332
32,370
229,373
TABLE II: AT-ADD statistics (number of clips) for Track 1 and Track 2.
ID
Source / Generator
Train
Dev
T1R1
Internal
2,000
2,000
T1R2
AISHELL-3 [ 97 ]
500
500
T1R3
LibriTTS-R [ 47 ]
2,500
2,500
T1R4
LJSpeech [ 40 ]
2,500
2,500
T1R5
Common Voice [ 3 ]
2,499
2,500
T1F1
ProDiff [ 38 ]
1,999
1,998
TABLE III: Real-source (T1R) and fake-source (T1F) conditions in the AT-ADD Track 1 train/dev subset. IDs continue across the Track 1 tables.
ID
Source / Generator
Progress
Eval
T1R1
Internal
974
5,000
T1R2
AISHELL-3 [ 97 ]
391
2,000
T1R3
LibriTTS-R [ 47 ]
832
3,999
T1R5
Common Voice [ 3 ]
1,022
5,000
T1R6
3D-Speaker [ 134 ]
385
1,999
T1R7
EchoFake [ 129 ]
396
2,000
TABLE IV: Real-source (T1R) and fake-source (T1F) conditions in the AT-ADD Track 1 progress/eval subset.
Type
ID
Source / Generator
Train
Dev
Sound
T2R8
AudioCaps [ 44 ]
9,854
4,940
T2F46
AudioLDM [ 61 ]
9,995
4,997
T2F47
AudioLDM2 [ 62 ]
10,000
5,000
T2F48
Audiogen [ 49 ]
9,991
4,992
–
Total
39,840
19,929
Singing
T2R9
OpenCpop [ 108 ]
3,414
–
TABLE V: Real-source (T2R) and fake-source (T2F) conditions in the AT-ADD Track 2 train/dev subset. IDs continue across the Track 2 tables.
Type
ID
Source / Generator
Progress
Eval
Sound
T2R8
AudioCaps [ 44 ]
999
4,938
T2R12
AVQA [ 120 ]
200
1,000
T2R13
CompA-R [ 25 ]
198
990
T2R14
VocalSound [ 28 ]
198
992
T2R15
TUT2016 [ 72 ]
200
1,000
T2F56
UniAudio2 [ 118 ]
942
4,711
TABLE VI: Real-source (T2R) and fake-source (T2F) conditions in the AT-ADD Track 2 progress/eval subset.
Type
Model
T1
T2
T2 Progress (by Type)
T2 Eval (by Type)
Dev
Progress
Eval
Dev
Progress
Eval
Speech
Sound
Singing
Music
Speech
Sound
Singing
Music
Conventional
Spec-ResNet
81.85
47.93
47.41
57.51
53.22
53.83
51.08
52.92
48.41
60.48
51.79
54.35
49.29
59.88
AASIST
94.16
60.78
60.39
93.63
62.38
62.21
63.69
56.88
64.08
64.87
63.58
56.62
63.81
64.85
SSL-based
FT-XLSR-AASIST
99.70
76.98
76.73
98.48
79.25
79.47
79.43
66.08
96.33
75.17
79.50
66.82
96.30
75.28
WPT-XLSR-AASIST
96.27
73.56
73.35
95.00
66.59
66.68
69.42
52.97
79.81
64.17
69.31
53.83
79.56
64.04
ALLM-based
Qwen2.5-Omni-3B
93.97
68.65
68.02
94.44
63.42
63.23
69.47
50.38
66.31
67.52
68.74
50.41
65.78
67.99
TABLE VII: Baseline performance (%) on the AT-ADD benchmarks. Best results are highlighted in bold.
Track
Rank
Team
Data Augmentation
System Name
Sub.
Progress
Eval
T1
1
WaveShield
Noise; Reverb; Codec; Perturb.; Replay; Seg.
w2vBERT2.0-AASIST
3
84.54
90.71
2
Fosafer
Noise; Reverb; Perturb.; Seg.
ssl_gfcc_multiscale_ensemble
4
85.31
86.67
3
sonomsl
Noise; Reverb; Codec; Perturb.; Replay
sonomsl
3
86.25
86.63
4
ThreeTO
Codec; RB; Perturb.; Replay
atadd_track1_xlsr_aasist_ensemble
3
83.61
83.79
5
NKU-HLT
Reverb; Codec
EnvRobust-XLSR-AASIST
1
84.20
83.68
T2
1
starfire
Codec; RB; Seg.
starfire_track2_audio_type_routed_ensemble
4
96.28
96.10
TABLE VIII: Final leaderboards for Track 1 robust speech deepfake detection and Track 2 all-type audio deepfake detection. Scores are Macro-F1 (%).
Fig. 2: Inference pipeline comparison of the top three systems in each AT-ADD track. Track 1 systems mainly differ in subsystem fusion, temporal scoring, and calibration, whereas Track 2 systems use either explicit audio-type routing, unified dual-representation fusion, or training-time specialists with a simplified unified inference path.
Fig. 3: Generator-wise fake recall (%) of the top five submissions. Cell colors span 50–100%, and values are rounded to integers for readability. Columns are sorted from difficult to easy within Track 1 and within each Track 2 audio type. Track 2 boundaries separate speech (T2F21–T2F45), sound (T2F56–T2F59), singing (T2F60–T2F64), and music (T2F65–T2F68). Generator names are given in Tables IV and VI .
Track
ID
Generator
Mean (range), %
T1
T1F24
BigVGAN
72.84 (62.81–85.93)
T1
T1F46
EchoFake replay
83.18 (74.65–93.91)
T1
T1F47
Internal replay
86.69 (73.04–99.13)
T2
T2F24
BigVGAN (speech)
58.33 (28.35–75.09)
T2
T2F30
StyleTTS 2 (speech)
88.49 (78.64–94.69)
T2
T2F59
MeanAudio (sound)
89.16 (64.12–100.00)
TABLE IX: Most difficult evaluation generators, ranked by mean fake recall across the top five systems. The range gives the minimum and maximum system recall.
Fig. 4: Source-wise recall (%) on real evaluation samples for (a) Track 1 and (b) Track 2. Cell colors span 50–100%. Only sources present in the evaluation sets are shown; thus, T1R4 is omitted from panel (a). Columns use the real-source IDs defined in Tables III – VI ; Track 2 boundaries separate speech, sound, singing, and music. Higher values indicate better bonafide preservation.
Fig. 5: Track 1 language-level robustness of the top five submissions. (a) Released train+dev support versus the mean language-level Macro-F1 across the five systems; marker size denotes evaluation support and the dashed line is a log-linear trend. (b) Macro-F1 averaged equally over languages within each resource group. Only languages with at least 20 real and 20 fake evaluation samples are included.
Fig. 6: Submission-level diagnostics. (a) Track 2 type-wise Macro-F1. (b) Pairwise Jaccard similarity of Track 1 error sets; the diagonal is omitted. (c) Official winner, hard-decision majority voting, and non-deployable oracle upper bounds. (d) Stratified-bootstrap 95% confidence intervals from 5,000 resamples.
Communication University of China & Ant Group Beijing, China · Machine Intelligence, Ant Group Shanghai, China · Institute of Automation, Chinese Academy of Sciences Beijing, China +2