Representation-Based Data Quality Audits for Audio

Abstract

Data quality issues such as off-topic samples, near duplicates, and label errors often limit the performance of audio-based systems. This paper addresses these issues by adapting SelfClean, a representation-to-rank data auditing framework, from the image to the audio domain. The approach leverages self-supervised audio representations to identify common data quality issues and produce ranked review lists within a single unified process. We benchmark the method on ESC-50, GTZAN, and a proprietary industrial dataset, using both synthetic and naturally occurring corruptions. Results show state-of-the-art ranking performance, often outperforming issue-specific baselines, and enabling significant annotation savings by efficiently guiding human review.

Results

Synthetic evaluation on ESC-50 under three corruption rates (α ∈ {0.05, 0.1, 0.2}). Below are the paper tables reporting ranking performance (AUROC, AP) for off-topic (OT), near-duplicate (ND), and label-error (LE) detection.

Table 1 — Pre-trained representations on ESC-50 (synthetic noise)

Performance of different audio embeddings for detecting OT / ND / LE across contamination rates.

Issue	Model	α=0.05 AUROC	α=0.05 AP	α=0.1 AUROC	α=0.1 AP	α=0.2 AUROC	α=0.2 AP
OT	CLMR	0.506	0.050	0.502	0.098	0.497	0.196
OT	CAV-MAE	0.309	0.049	0.260	0.075	0.273	0.161
OT	M2D	0.689	0.074	0.510	0.095	0.373	0.159
OT	EAT	0.591	0.070	0.596	0.138	0.544	0.222
OT	BEATs	0.766	0.253	0.745	0.316	0.673	0.341
OT	CLMR (SSL)	0.222	0.031	0.175	0.058	0.163	0.118
OT	BEATs (LoRA)	0.724	0.202	0.743	0.330	0.653	0.313
ND	CLMR	0.740	0.001	0.747	0.001	0.744	0.001
ND	CAV-MAE	0.744	0.032	0.724	0.017	0.730	0.018
ND	M2D	0.992	0.606	0.993	0.587	0.993	0.617
ND	EAT	0.930	0.482	0.922	0.468	0.931	0.476
ND	BEATs	0.972	0.606	0.978	0.595	0.978	0.625
ND	CLMR (SSL)	0.911	0.400	0.888	0.393	0.898	0.384
ND	BEATs (LoRA)	0.970	0.608	0.975	0.588	0.977	0.619
LE	CLMR	0.477	0.049	0.484	0.094	0.492	0.197
LE	CAV-MAE	0.721	0.222	0.693	0.299	0.658	0.387
LE	M2D	0.998	0.970	0.995	0.950	0.986	0.943
LE	EAT	0.969	0.668	0.969	0.759	0.954	0.793
LE	BEATs	0.996	0.927	0.992	0.908	0.980	0.903
LE	CLMR (SSL)	0.957	0.586	0.959	0.723	0.942	0.792
LE	BEATs (LoRA)	0.997	0.932	0.992	0.915	0.978	0.903

Table 2 — SelfClean vs issue-specific baselines (ESC-50 synthetic)

SelfClean (with BEATs embeddings) compared against Isolation Forest (OT), Dejavu fingerprinting (ND), and Confident Learning (LE).

Issue	Model	α=0.05 AUROC	α=0.05 AP	α=0.1 AUROC	α=0.1 AP	α=0.2 AUROC	α=0.2 AP
OT	IForest	0.791	0.212	0.676	0.177	0.406	0.188
OT	SelfClean	0.766	0.253	0.745	0.316	0.673	0.341
ND	Dejavu	0.862	0.017	0.835	0.033	0.845	0.068
ND	SelfClean	0.972	0.606	0.978	0.595	0.978	0.625
LE	CLearning	0.994	0.884	0.994	0.951	0.993	0.973
LE	SelfClean	0.996	0.927	0.992	0.908	0.980	0.903

Tip: on mobile, swipe horizontally to view full tables.

BibTeX

Copy-paste citation

@article{gonzalezjimenez2025representation,
  title   = {Representation-Based Data Quality Audits for Audio},
  author  = {Gonzalez-Jimenez, Alvaro and Gr{\"o}ger, Fabian and Wermelinger, Linda and B{\"u}rli, Andrin and Kastanis, Iason and Lionetti, Simone and Pouly, Marc},
  journal = {IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP)},
  year    = {2026}
}