We investigate whether AI-generated impulsive sounds can be distinguished from real ones through group delay analysis. Our central finding is that AI-generated impulsive sounds show near-identical onset-region group-delay distributions but exhibit measurably different group-delay behavior in the late decay region: decay-region KL divergence reaches 0.322 compared to near-zero onset divergence (0.022). Cross-band GD variability achieves single-feature AUC~=0.720, and a Random Forest (RF) over nine decay-region features reaches AUC=0.884 under sample-disjoint evaluation. A group delay map used as a standalone 2D input to CNN classifiers achieves 90--94% accuracy, demonstrating that group delay carries substantial discriminative information. Under generator hold-out, CNN and transformer classifiers show highly variable AUC (0.457--0.918). The group delay RF achieves the highest average hold-out accuracy among the evaluated methods (66.7%) and avoids extreme below-random collapse, although its average AUC (0.731) is lower than CNN avg (0.762) and AST (0.772). Parameter sensitivity analysis across 27 STFT configurations confirms that the RF AUC remains stable (0.700--0.847, std=~0.035). These results suggest that decay-region group delay can serve as a physically interpretable forensic cue that complements magnitude-based classifiers, while broader validation remains necessary.
Figures & tables
Fig. 1: Overall framework for decay-region group delay forensic analysis. The pipeline processes a 3-second audio clip through STFT and phase unwrapping to extract a mel-compressed group delay map Gˉ , which feeds three analysis branches: (A) decay-region scalar features with Random Forest classification, (B) GD-map-based CNN/AST detection, and (C) magnitude comparison. All branches are evaluated under sample-disjoint split, generator hold-out, real-source hold-out, and STFT sensitivity protocols.
Class
Source
Orig.
Final
×
Real
FSD50K
1,122
3,750
3.3
Real
SESA
585
3,750
6.4
Fake
ElevenLabs
1,119
2,500
2.2
Fake
Stable Audio
1,000
2,500
2.5
Fake
AudioLDM2
1,000
2,500
2.5
TABLE I: Dataset construction.
Condition
AUC
Acc (%)
FSD50K → SESA (hold-out)
0.849
76.47
SESA → FSD50K (hold-out)
0.857
77.00
Mixed → Mixed (reference)
0.878
79.33
TABLE II: Real-source hold-out: GD RF performance under real-source shift.
Fig. 2: Group delay CDF: onset region (left, KL = 0.022) shows near-identical real/fake distributions, while the decay region (right, KL = 0.322) reveals measurably different distributions. Shaded area indicates the gap between real and fake CDFs.
Region / Band
KL
Overall
0.2426
Low freq (0–500 Hz)
0.2339
Mid freq (500–2000 Hz)
0.2417
High freq (2000–8000 Hz)
0.2437
Onset (first 30%)
0.022
Decay (last 50%)
0.322
TABLE III: KL divergence: real vs. fake.
Feature
Cohen’s d
Cliff’s δ
AUC
Decay variance
0.588
0.261
0.651
Decay entropy
0.502
0.010
0.521
Decay abs. mean
0.578
0.204
0.628
Decay/onset var. ratio
0.036
0.096
0.573
Decay → onset KL
0.482
0.114
0.570
Decay → onset shift
0.084
0.102
0.530
TABLE IV: Decay-region scalar features ( N=15,000 , all p<0.001 ).
Fig. 3: Generator-wise group delay analysis ( ∗∗∗p<0.001 , ∗p<0.05 , ns = not significant vs. real). The most pronounced separation appears in Decay-to-Onset KL (top-right), where Stable Audio and AudioLDM2 differ significantly ( p<0.001 ) while ElevenLabs shows a weaker, non-significant trend. The compact fake distributions may partly reflect generator-specific homogeneity; these results are interpreted as evidence of a candidate forensic cue rather than a source-independent detector.
TABLE VII: Generator hold-out: Acc (%) and AUC. Bold : best Acc or best AUC per hold-out condition. CNN avg denotes the average over ResNet50, EfficientNet-B2, and CNN14; AST is reported separately as a transformer-based model.
Fig. 4: Phase-Flow CRNN temporal attention (note: y-axis scales differ across panels). Real samples (a,b) show relatively diffuse or transient attention patterns, whereas fake samples (c,d) exhibit concentrated attention in the late decay region. This pattern is consistent with the decay-region group delay findings.
We present ArtifactNet, a lightweight framework that detects AI-generated music by reframing the problem as forensic physics -- extracting and analyzing the physical artifacts that neural audio codecs inevitably imprint on generated audio. A bounded-mask UNet (ArtifactUNet, 3.6M parameters) extracts codec residuals from magnitude spectrograms, which are then decomposed via HPSS into 7-channel forensic features for classification by a compact CNN (0.4M parameters; 4.0M total). We introduce ArtifactBench, a multi-generator evaluation benchmark comprising 6,183 tracks (4,383 AI from 22 generators and 1,800 real from 6 diverse sources). Each track is tagged with bench_origin for fair zero-shot evaluation. On the unseen test partition (n=2,263), ArtifactNet achieves F1 = 0.9829 with FPR = 1.49%, compared to CLAM (F1 = 0.7576, FPR = 69.26%) and SpecTTTra (F1 = 0.7713, FPR = 19.43%) evaluated under identical conditions with published checkpoints. Codec-aware training (4-way WAV/MP3/AAC/Opus augmentation) further reduces cross-codec probability drift by 83% (Delta = 0.95 -> 0.16), resolving the primary codec-invariance failure mode. These results establish forensic physics -- direct extraction of codec-level artifacts -- as a more generalizable and parameter-efficient paradigm for AI music detection than representation learning, using 49x fewer parameters than CLAM and 4.8x fewer than SpecTTTra.
Heewon Oh
Intrect / MARTE Lab, Dongguk University, Seoul, South Korea
The proliferation of AI-generated audio (so-called "deepfake" audio) poses significant threats to information integrity, from voice cloning fraud to synthetic music copyright disputes. We present a temporal coherence analysis framework built upon Contrastive Language-Audio Pretraining (CLAP) embeddings that spans speech, instrumental music, and music with vocals. By computing pairwise cosine similarities between audio segment embeddings and extracting statistical features from the resulting distributions, we train lightweight ensemble classifiers that reliably distinguish authentic from synthetic audio. Our work provides an interpretable, computationally efficient alternative to common deep learning methods while still achieving competitive performance across speech and music domains. Further, we reveal two notable empirical findings about audio deepfakes: (1) a feature-label inversion phenomenon in which 21 of 29 statistical features reverse their discriminative direction between training and in-the-wild deployment, and (2) a speech--music direction reversal in which entropy discriminates in opposite directions for speech and music deepfakes.
We present a novel method for AI-generated music detection in scenarios where the models that generated the input samples are unknown to the detector (e.g., from a newly released service). Since 2023, there has been a multiplication of user-friendly AI-music generation services (e.g., Suno, Udio), along with regular updates and new features. There is thus a need to address synthetic content detection in an unsupervised way to adapt to this rapidly changing context. This angle has not been much studied in music yet. We propose to study two tasks. First, discriminating between real and synthetic music. This may be approached in a one-class manner, namely, using some baseline real music and trying to determine what falls outside. Second, zero-shot multi-class identification, which is more similar to an unsupervised clustering task on a mix of real and various AI-music generations, where the goal is to create coherent, high-purity clusters. We propose a combination of a previously proposed artifact-extraction method, on top of which we apply non-negative matrix factorization and simple classification and clustering methods. We achieve excellent performance on both tasks, showing that the proposed methods may be used to monitor large-scale catalogs that may receive AI-generated samples from various newly released generative models.