PaperlayerView exact revision on arXiv

Source-linked AI summary

Decay-Region Group Delay as a Forensic Cue for AI-Generated Impulsive Sounds

JaeHyeong Chang, Chengzhe Sun, Siwei Lyu

arXiv:2608.26346v1cs.SDcs.AI

TL;DR

The paper asks whether AI-generated impulsive sounds can be distinguished from real sounds using group delay, addressing limited forensic evidence for this domain. It analyzes decay-region group-delay cues with scalar and map-based classifiers, finding strong sample-disjoint discrimination but variable generator hold-out transfer. The results support decay-region group delay as a physically interpretable complementary forensic cue, while broader validation remains necessary.

  • Problem

    AI-generated environmental sound forensics remains comparatively underexplored despite increasingly realistic synthesized impulsive sounds.

  • Method

    The study extracts decay-region group-delay features and maps from STFT phase, evaluates Random Forest, CNN, and transformer detectors, and tests sample-disjoint, generator hold-out, and sensitivity conditions.

  • Results

    90–94% standalone accuracy is achieved by group delay maps in-distribution, while generator hold-out AUC varies from 0.457–0.918 and the group delay RF reaches 66.7% average accuracy.

  • Takeaways & Limitations

    Decay-region group delay provides a physically interpretable and complementary lens for AI-generated impulsive sound forensics.

  • Takeaways & Limitations

    Only three generators are evaluated, and generalization is not uniform, with the GD RF reaching AUC = 0.580 on AudioLDM2 hold-out.

Abstract

from arXiv · show

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.

I. INTRODUCTION

The paper addresses the underexplored forensic detection of AI-generated impulsive sounds by examining physically constrained decay-region group delay. It presents decay-region group delay as a complementary cue alongside magnitude-based approaches and evaluates generalization across detection settings.

  • AI-generated environmental sound forensics remains comparatively underexplored despite increasingly realistic synthetic impulsive sounds.
  • Decay-region group delay is proposed as an important forensic cue because physical propagation, reflection, and dissipation constrain real impulsive sounds.
  • The framework evaluates decay-region scalar features, GD-map classifiers, and magnitude comparisons under sample-disjoint, hold-out, and STFT-sensitivity protocols.
  • 90–94% standalone accuracy is achieved by group delay maps in the sample-disjoint setting.
  • CNN and transformer AUC ranges from 0.457–0.918 under generator hold-out, while the group delay RF avoids extreme below-random collapse and attains higher average accuracy.

B. Phase and Group Delay in Audio Analysis

Prior work shows that phase and group delay capture temporal structure complementary to magnitude, while this study extends those ideas to impulsive sound forensics using generated and real datasets.

  • Group delay is used in audio analysis because it captures temporal structure not visible in magnitude spectrograms and can complement magnitude features.
  • The study extends phase-based forensic analysis from speech deepfakes to the less-studied domain of impulsive sounds, focusing on decay-region group delay.
  • The dataset contains 15,000 balanced three-second clips, with real recordings from FSD50K and SESA and fake samples from three audio generators.
  • Augmentation is applied identically to real and fake classes, while sample-disjoint partitioning keeps variants of each source file within one split.
  • The authors frame the results as initial evidence for a physically interpretable forensic cue rather than corpus-independent benchmark performance.

A. Real-Source Bias Control

The real-source hold-out experiment tests whether group delay detection persists when the real recording source changes. Performance remains high but is lower than the mixed reference, and partial source separability remains measurable.

  • The real-source hold-out swaps FSD50K and SESA real samples while preserving sample-disjoint fake partitions.
  • AUC = 0.849–0.857 under real-source shift versus AUC = 0.878 for the mixed reference.
  • AUC = 0.737 for classifying FSD50K versus SESA real samples confirms partial source separability.
  • Group delay is defined as the negative frequency derivative of phase and is interpreted in the decay region through physically constrained acoustic propagation.

B. STFT and Phase

The preprocessing converts three-second waveforms into STFT phase representations, unwraps phase across frequency, and approximates group delay with finite differences subject to clipping for numerical stability.

  • The STFT processes a 3-second waveform sampled at 16 kHz using a Hann window.
  • The configuration uses N = 1024 and hop H = 256 for the short-time Fourier transform.
  • Group delay is discretized using finite differences of the frequency-unwrapped phase.
  • Values are clipped at τmax = 500 samples, approximately 31 ms at 16 kHz, for numerical stability.

D. Mel-compressed Group Delay Map

The mel-compressed group delay map reduces frequency resolution to 128 mel bands and produces a 128 × 188 representation for a three-second clip. Image-based classifiers receive a zero-padded 128 × 1024 version, while scalar features use the original representation.

  • 128 mel bands compress the group-delay frequency axis.
  • The normalized group-delay map is represented using values in [−1, 1].
  • A three-second, 16 kHz clip with hop H = 256 yields a 128 × 188 group-delay map.
  • The map is zero-padded to 128 × 1024 only for image-based classifiers, whereas scalar features use the original 128 × 188 representation.

E. Decay-Region Features

The paper defines decay-region features over the late half of each clip, where impulsive energy is reduced, and uses cross-band variability and temporal KL divergence to characterize group-delay behavior.

  • Frames m ≥ floor(0.5T), with T = 188, define the decay region as the clip’s late temporal half.
  • Cross-band GD variability measures the standard deviation of per-band mean group delay in the decay region.
  • AI-generated samples show higher cross-band GD variability than real impulsive sounds, indicating less stable phase-derivative behavior across frequency bands.
  • The Decay-to-Onset KL divergence quantifies temporal distributional shift between onset and decay group-delay distributions.
  • Decay-region CDF divergence is KL = 0.322, versus KL = 0.022 for the onset region.

V. GROUP DELAY FORENSIC ANALYSIS

Decay-region group-delay statistics provide stronger real–fake separation than onset statistics, while combined features achieve strong sample-disjoint classification but require careful interpretation of KL measures.

  • KL = 0.022 in the onset region indicates near-identical real and fake distributions, while KL = 0.322 in decay reveals larger divergence.
  • AUC = 0.720 is achieved by cross-band GD variability, the strongest single scalar feature in the decay-region analysis.
  • AUC = 0.884 is achieved by a Random Forest over nine decay-region features under sample-disjoint evaluation.
  • Decay-to-Onset KL has strong distribution-level separation but standalone AUC = 0.570 because real samples have high variance.
  • Generator-wise analysis identifies the strongest Decay-to-Onset KL separation for Stable Audio and AudioLDM2, while ElevenLabs shows a weaker non-significant trend.

C. Consistency Across Generators

Decay-region group-delay behavior shows systematic but uneven generator-wise differences, and Random Forest performance remains relatively stable across the tested STFT configurations.

  • Consistency Across Generators: The strongest generator-wise visual separation appears in Decay-to-Onset KL, with Stable Audio and AudioLDM2 near zero and real samples highly variable.
  • Consistency Across Generators: ElevenLabs shows a weaker but directionally consistent Decay-to-Onset KL trend that is not significant on this feature alone.
  • Consistency Across Generators: Cross-band GD variability and decay variance differ significantly across tested generators but remain partially overlapping.
  • Consistency Across Generators: The evaluation covers 27 STFT configurations spanning three FFT sizes, three hop sizes, and three clip settings.
  • Consistency Across Generators: RF AUC ranges from 0.700–0.847 with std = 0.035 across all 27 configurations.
  • Consistency Across Generators: For N ≥1024, cross-band GD variability stabilizes at AUC 0.686–0.734, making N = 1024 and hop = 256 representative of the stable region.

VI. DETECTION EXPERIMENTS

The detection experiments compare magnitude, group-delay, and combined representations under sample-disjoint and generator-hold-out evaluation. Group-delay maps are strongly discriminative in-distribution, while cross-generator performance is variable and the GD RF offers greater stability than some neural classifiers.

  • Feature Representation Comparison: 90–94% standalone accuracy is achieved by GD-only maps, demonstrating substantial discriminative information in group delay.The comparison includes Mag, GD, and Mag+GD inputs across ResNet50, EfficientNet-B2, CNN14, and AST.
  • Feature Representation Comparison: 0.4 percentage points is the ResNet50 improvement from adding GD to magnitude, while higher-capacity models show a ceiling effect.
  • Generator Hold-out Evaluation: 0.457–0.918 is the AUC range for CNN and transformer classifiers under generator hold-out, indicating highly variable cross-generator performance.CNN14 ranges from AUC = 0.918 for Stable Audio hold-out to AUC = 0.457 for AudioLDM2 hold-out.
  • Generator Hold-out Evaluation: 66.7% is the GD RF’s highest average accuracy, despite its lower average AUC of 0.731 versus 0.762 for CNN avg and 0.772 for AST.The GD RF avoids the extreme below-random collapse observed for CNN14 on AudioLDM2 hold-out.
  • Generator Hold-out Evaluation: AUC = 0.580 on AudioLDM2 hold-out shows that GD cues do not transfer uniformly across generator architectures.AudioLDM2 is reported as the most challenging hold-out condition for all methods.

C. Supplementary: Phase-Flow CRNN Attention

The supplementary Phase-Flow CRNN analysis examines where temporal attention concentrates in impulsive sounds. Its attention pattern provides qualitative support for the paper’s decay-region group-delay finding, while broader results emphasize limited generator generalization and dataset constraints.

  • C. Supplementary: Phase-Flow CRNN Attention: 92.33% in-distribution accuracy is achieved by the 2.24M-parameter Phase-Flow CRNN.The model processes temporal phase-flow with a BiGRU and temporal attention.
  • C. Supplementary: Phase-Flow CRNN Attention: Late decay regions attract concentrated attention in fake samples, whereas real samples show relatively diffuse or transient attention patterns.The figure caption notes that y-axis scales differ across panels.
  • VII. CONCLUSION: 0.457–0.918 is the AUC range observed for CNN and transformer classifiers under generator hold-out, while the group delay RF achieves the highest average accuracy and avoids extreme below-random collapse.The conclusion characterizes phase-based representations as a physically interpretable and complementary forensic lens.
  • VII. CONCLUSION: 0.035 is the RF AUC standard deviation across 27 STFT configurations, supporting robustness in the reported parameter sensitivity analysis.
  • VIII. LIMITATIONS AND FUTURE WORK: A 6.4× SESA augmentation ratio, three-generator evaluation, and AUC = 0.580 on AudioLDM2 hold-out constrain the scope of generalization claims.The paper also notes that the STFT sensitivity analysis uses a smaller subset and that adversarial robustness remains untested.
Loading 2608.26346v1…