PADP: Perceptual Audio Data Perturbation for Probing Perception Awareness in Audio Quality Models

1. INTRODUCTION

• The paper asks whether objective audio quality metrics assess quality consistently with human perception, despite using explicit psychoacoustic principles or learning them from data.A robust metric should remain insensitive to signal modifications that listeners perceive as irrelevant.Source: §1 · p. 1
• PADP introduces four transforms—APX, SCD, DEL, and DIF—as robustness probes for audio quality models.The transforms introduce distortions intended to be perceptually irrelevant to listeners.Source: §1 · p. 1
• Subjective listening experiments confirm that the transforms substantially alter waveforms while preserving perceived sound quality and content across diverse signal types.The experiments provide perceptual validation for using the transformations as stress tests.Source: §1 · p. 1
• The transforms challenge state-of-the-art audio quality systems and foundation models, exposing discrepancies between model responses and human judgments for listener-irrelevant perturbations.The proposed probes target cases where models respond to signal changes that listeners do not perceive as relevant.Source: §1 · p. 1

2. METHOD

• QMF analysis decomposes broadband audio into equally spaced subbands that can be grouped to approximate perceptual frequency regions, with near-alias-free reconstruction after modification.A fixed binary mapping associates perceptual bands with QMF subbands, using progressively more subbands at higher frequencies.Source: §2 · p. 1

2.1. All-Pass Band Phase Shift Combination (APX)

• APX applies shared phase shifts to groups of perceptual bands while preserving spectral magnitudes, exploiting limited sensitivity to some phase changes within auditory bands.An offset excludes the lowest bands, and a stride sets the groupings for shared phase shifts.Source: §2.1 · p. 2

2.2. Frequency Selective Delay (DEL)

• DEL applies a fixed frequency-selective delay pattern, assigning larger delays to lower perceptual bands in accordance with frequency-dependent temporal resolution.The band delays are mapped to QMF subbands and applied to their coefficient sequences before time-domain reconstruction.Source: §2.2 · p. 2

2.3. Scaled Coding Distortion (SCD)

• SCD interpolates between the original signal and a codec reconstruction, with λ controlling perturbation strength and sufficiently small values verified as transparent by listening tests.The reconstruction is assumed perceptually indistinguishable from the original when the bitrate is sufficiently high.Source: §2.3 · p. 2

2.4. Decorrelator / Diffusion (DIF)

• DIF mixes direct and decorrelated components to control spatial diffusion while largely preserving perceived timbre.Its target inter-channel coherence sets the diffusion level, and the decorrelator uses lattice all-pass filters in the QMF domain.Source: §2.4 · pp. 2–3

3.1. Validation of Inaudibility of PADP

• The listening tests use 12 clean ODAQ signals spanning average and critical items, with specified settings for APX, DIF, and mp3-based SCD perturbations.The tests include hidden references and a low-pass filtered anchor; all investigations use monophonic signals.Source: §3.1 · p. 3
• Two MUSHRA tests separate SCD from APX, DEL, and DIF to limit conditions per session and reduce listener fatigue while improving sensitivity to subtle differences.After screening, nine listeners were retained for SCD and eight for APX/DEL/DIF.Source: §3.1 · p. 3

3.2. Robustness of Audio Models under PADPs

• The robustness evaluation covers four perception-motivated models, two learning-based models, and four audio foundation models, comparing outputs with average subjective ratings.For models without conventional full-reference quality scores, the evaluation uses Euclidean distances between reference and transformed-signal embeddings.Source: §3.2 · p. 3

4.1. Listening Tests

• APX, DIF1 (and potentially DIF2), and SCD with λ≤0.6 are transparent across signal types; DEL and SCD at λ=0.9 occasionally exceed the threshold.SCD with λ≥1.2 and DIF3–5 are slightly audible while preserving signal content.Source: §4.1 · p. 3, Figure 1 · p. 4

4.2. Robustness of Audio Quality Metrics

• APX, DEL, and DIF5 substantially alter waveforms, yet are less audible and disturbing than LP35, showing waveform similarity has limited correspondence with perception.SCD introduces smaller waveform changes.Source: §4.2 · p. 3, Figure 2 · p. 4
• DeePAQ has the highest absolute SRCC (0.74) and ViSQOL v3 the lowest (0.38); foundation models show limited accuracy under PADPs.The remaining models range from 0.60 to 0.70 in SRCC.Source: §4.2 · p. 4

5. CONCLUSION

• APX, DIF1–2, and SCD with λ≤0.6 remain transparent across all signal types.The passage also begins to identify settings for less stringent applications, but does not provide the complete statement.Source: §5 · p. 4

Quality Metrics

• SCOREQ is −0.60 and DeePAQ is −0.74 in the supplied metric listing.Source: Quality Metrics · p. 4

• Table 1 reports Spearman correlations between audio quality estimates and subjective ratings.Source: Table 1 · p. 4
• Figure 3 compares DeePAQ and ViSQOL predictions with subjective ratings across signals and PADP transformation conditions.Source: Figure 3 · p. 4
• Applying PADP reveals susceptibility in objective audio quality metrics and foundation models, and suggests PADP as a potential data augmentation technique.Source: Models · p. 4