Domain-Agnostic Incremental Learning for Sound Classification. A DCASE 2026 Challenge task

Annamaria Mesaros; Jesper Rindom Jensen; Manjunath Mulimani; Manu Harju; Riccardo Casciotti

The DCASE 2026 challenge defines a task where systems learn ten sound classes sequentially across three acoustic domains without access to prior data.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-06-28 12:57 UTC pith:AP577FBZ

load-bearing objection This is a DCASE challenge announcement that defines a domain-incremental sound classification task and supplies a weak baseline, not a research paper with new methods or findings.

arxiv 2606.02173 v2 pith:AP577FBZ submitted 2026-06-01 eess.AS

Domain-Agnostic Incremental Learning for Sound Classification. A DCASE 2026 Challenge task

Riccardo Casciotti , Manjunath Mulimani , Manu Harju , Jesper Rindom Jensen , Annamaria Mesaros This is my paper

classification eess.AS

keywords incremental learningdomain incremental learningsound classificationDCASE challengeaudio classificationcontinual learningacoustic domainsdomain agnostic

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper presents the Domain-Agnostic Incremental Learning for Sound Classification task for the DCASE 2026 Challenge. It requires participants to train a single system that learns the same ten sound classes in three different acoustic domains one after another, with no access to data from earlier domains during later training stages. Systems are ranked solely by their average accuracy across all three domains after completing the sequence. The provided baseline reaches 52.5 percent accuracy on the final two domains, primarily because it fails to correctly identify the domain of test samples. The task therefore tests whether audio classifiers can handle domain shifts incrementally while retaining class knowledge.

Core claim

The paper formalizes domain-incremental learning for sound classification as training on ten classes in three sequential domains where each incremental step supplies data only from the current domain, with submitted systems evaluated by their overall average accuracy over the three domains.

What carries the argument

The no-rehearsal domain-incremental protocol that ranks systems by final average accuracy across domains.

Load-bearing premise

The task setup assumes that sequential learning across domains can be evaluated fairly using only the final average accuracy metric without access to previous task data.

What would settle it

Running the baseline on the test set while also measuring its domain-prediction accuracy separately; if domain errors account for most of the accuracy drop, the claim that domain inference is the main bottleneck would be confirmed.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

Submitted systems must classify sounds correctly in the current domain while retaining performance on earlier domains without storing old data.
Domain identification becomes an implicit requirement for high average accuracy.
The average-accuracy ranking metric penalizes any domain-specific degradation.
The challenge isolates the effect of domain shift from class-incremental forgetting.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Methods that explicitly detect or normalize for domain before classification could outperform the baseline substantially.
The task setup could be extended to measure per-domain accuracy to diagnose whether errors concentrate in particular domains.
Real-world deployment would require the system to handle unknown future domains without task boundaries being announced in advance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

This is a DCASE challenge announcement that defines a domain-incremental sound classification task and supplies a weak baseline, not a research paper with new methods or findings.

read the letter

This document sets up the Domain-Agnostic Incremental Learning task for DCASE 2026. Participants must learn ten sound classes across three acoustic domains in sequence, with no access to earlier data, and systems are ranked by average accuracy over all domains. The supplied baseline reaches 52.5% on the final two domains, with most errors traced to wrong domain inference at test time.

The setup itself is new in the DCASE context and matches a practical problem: audio systems that encounter shifting recording conditions over time. The rules are stated plainly, the metric is straightforward, and the baseline gives participants a concrete starting point.

The main limitation is that the paper is a task definition rather than a technical contribution. It offers little analysis of why the baseline fails on domain inference, no comparisons to existing continual-learning approaches, and no ablation on the evaluation protocol. The claim that average accuracy is the right ranking criterion is asserted without supporting checks on per-domain retention or catastrophic forgetting. Because the full text is mostly organizational, the 52.5% figure stands as an observation about one starter system rather than a verified result.

People running continual-learning experiments in audio or planning to enter the challenge will find the task description useful for benchmarking. A referee could usefully verify the data splits, domain definitions, and reproducibility of the baseline, but the paper does not require heavy technical review. I would send it to peer review so the community can engage with the task, while keeping expectations low on scientific novelty.

Referee Report

0 major / 2 minor

Summary. The manuscript defines the Domain-Agnostic Incremental Learning for Sound Classification task for the DCASE 2026 Challenge. Participants must train a single system on ten sound classes across three acoustic domains in sequence, with no access to data from prior domains at each incremental step. Submitted systems are ranked by overall average accuracy across the three domains. The provided baseline achieves 52.5% accuracy on the final two domains, attributed primarily to domain mis-inference on test samples.

Significance. The task formalizes a domain-incremental continual learning benchmark for audio classification, which could serve as a standardized evaluation platform if the data splits, domain definitions, and evaluation protocol are made fully reproducible. The baseline result illustrates practical difficulties but does not constitute a novel scientific claim. Impact will depend on community participation rather than internal theoretical or empirical advances.

minor comments (2)

[Abstract] Abstract: the 52.5% baseline figure and its attributed cause (domain inference error) are stated without accompanying details on the evaluation protocol, data splits, or domain definitions; a short pointer to the relevant task-description section would improve self-containment.
The manuscript should explicitly list the ten sound classes and characterize the three domains (e.g., recording conditions or acoustic environments) to allow readers to assess the domain-shift severity without external material.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for recommending acceptance of the manuscript. The provided summary accurately describes the Domain-Agnostic Incremental Learning task, the sequential training protocol across domains, and the baseline performance.

Circularity Check

0 steps flagged

No significant circularity

full rationale

This is a DCASE 2026 challenge task definition paper. It contains no derivations, equations, predictions, or fitted parameters. The sole numerical claim (52.5% baseline accuracy) is presented as an empirical observation on a supplied starter system, not as a result derived from the paper's own inputs or self-citations. No load-bearing steps reduce to self-definition or prior author work.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No mathematical content, free parameters, axioms, or invented entities are present because the document is a challenge task announcement.

pith-pipeline@v0.9.1-grok · 5697 in / 1063 out tokens · 39238 ms · 2026-06-28T12:57:43.187627+00:00 · methodology

0 comments

read the original abstract

This paper presents the Domain-Agnostic Incremental Learning for Audio Classification Task of the DCASE 2026 Challenge. Incremental learning refers to sequentially learning new tasks with the same system while maintaining its knowledge and performance on the previously learned task. Domain-incremental learning for sound classification refers to learning the same sound classes but in different acoustic domains, and was formalized as a data challenge for the first time in DCASE 2026. Participants will train a system to learn ten sound classes in three different domains, with learning at each incremental task not having access to previous task data. Submitted systems will be ranked by the overall average accuracy calculated over the three domains. During the development stage, the provided baseline system obtains a modest performance of 52.5\% accuracy over the last two domains, mostly due to erroneous inference of the domain for the test sample.

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Domain-incremental audio classification using domain-specific experts and prototype classifier
eess.AS 2026-06 unverdicted novelty 3.0

Domain-incremental audio classification via frozen domain-specific experts plus prototype classifier on concatenated features yields 78.15% micro / 77.03% macro accuracy on the DCASE 2026 Task 7 development set.

Reference graph

Works this paper leans on

16 extracted references · cited by 1 Pith paper

[1]

Domain-incremental learni ng for audio classiﬁcation,

M. Mulimani and A. Mesaros, “Domain-incremental learni ng for audio classiﬁcation,” in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2025, pp. 1–5

2025
[2]

Unsuper- vised adversarial domain adaptation for acoustic scene cla ssiﬁcation,

S. Gharib, K. Drossos, E. Cakir, D. Serdyuk, and T. Virtan en, “Unsuper- vised adversarial domain adaptation for acoustic scene cla ssiﬁcation,” in W orkshop on Detection and Classiﬁcation of Acoustic Scenes and Events (DCASE), 2018, pp. 138–142

2018
[3]

Overcoming catastrophic forgetting in neural networks,

J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. V eness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, et al. , “Overcoming catastrophic forgetting in neural networks,” in Proceedings of the National Academy of Sciences , vol. 114, no. 13, 2017, pp. 3521– 3526

2017
[4]

icarl: Incremental classiﬁer and representation learning,

S.-A. Rebufﬁ, A. Kolesnikov, G. Sperl, and C. H. Lampert, “icarl: Incremental classiﬁer and representation learning,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 2001–2010

2017
[5]

Episodic memory in lifelong language learning,

C. d. M. d’Autume, S. Ruder, and L. Kong, “Episodic memory in lifelong language learning,” in Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL) , 2019, pp. 5367–5379

2019
[6]

A closer look at class-incre mental learning for multi-label audio classiﬁcation,

M. Mulimani and A. Mesaros, “A closer look at class-incre mental learning for multi-label audio classiﬁcation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 33, pp. 1293–1306, 2023

2023
[7]

Continual learning for on-device environment al sound classiﬁcation,

Y . Xiao, X. Liu, J. King, A. Singh, E. S. Chng, M. D. Plumble y, and W. Wang, “Continual learning for on-device environment al sound classiﬁcation,” in Proceedings of the 7th Detection and Classiﬁcation of Acoustic Scenes and Events 2022 W orkshop (DCASE2022) , Nancy, France, November 2022

2022
[8]

Learning without forgetting,

Z. Li and D. Hoiem, “Learning without forgetting,” IEEE Transactions on pattern analysis and machine intelligence , vol. 40, no. 12, pp. 2935– 2947, 2017

2017
[9]

Con- tinual learning of new sound classes using generative repla y,

Z. Wang, C. Subakan, E. Tzinis, P . Smaragdis, and L. Charl in, “Con- tinual learning of new sound classes using generative repla y,” in IEEE W orkshop on Applications of Signal Processing to Audio and A coustics (WASPAA). IEEE, 2019, pp. 308–312

2019
[10]

Addressing imbalanced domain-incremental learning through dual-balance collab orative experts,

L. Li, D.-W. Zhou, H.-J. Y e, and D.-C. Zhan, “Addressing imbalanced domain-incremental learning through dual-balance collab orative experts,” in F orty-second International Conference on Machine Learnin g, 2025

2025
[11]

Online domain-incrementa l learning for acoustic scene classiﬁcation across locations,

M. Mulimani and A. Mesaros, “Online domain-incrementa l learning for acoustic scene classiﬁcation across locations,” in Proceedings of the IEEE W orkshop on Applications of Signal Processing to Au dio and Acoustics (WASPAA), 2025

2025
[12]

Cla ss-incremental learning for sound event localization and detection,

R. Pandey, M. Mulimani, A. Politis, and A. Mesaros, “Cla ss-incremental learning for sound event localization and detection,” in 2025 IEEE International Conference on Acoustics, Speech, and Signal Processing W orkshops (ICASSPW). IEEE, 2025, pp. 1–5

2025
[13]

Three scenarios for con tinual learning,

G. M. V an de V en and A. S. Tolias, “Three scenarios for con tinual learning,” in Continual Learning W orkshop NeurIPS , 2018

2018
[14]

Acoustic scene classiﬁcat ion across mul- tiple devices through incremental learning of device-spec iﬁc domains,

M. Mulimani and A. Mesaros, “Acoustic scene classiﬁcat ion across mul- tiple devices through incremental learning of device-spec iﬁc domains,” in Proceedings of the DCASE W orkshop , 2024, pp. 81–85

2024
[15]

Audio set: An ontology and h uman- labeled dataset for audio events,

J. F. Gemmeke, D. P . Ellis, D. Freedman, A. Jansen, W. Law rence, R. C. Moore, M. Plakal, and M. Ritter, “Audio set: An ontology and h uman- labeled dataset for audio events,” in IEEE international conference on acoustics, speech and signal processing (ICASSP) . IEEE, 2017, pp. 776–780

2017
[16]

PANNs: Large-scale pretrained audio neural networks for a udio pattern recognition,

Q. Kong, Y . Cao, T. Iqbal, Y . Wang, W. Wang, and M. D. Plumb ley, “PANNs: Large-scale pretrained audio neural networks for a udio pattern recognition,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 28, pp. 2880–2894, 2020

2020

[1] [1]

Domain-incremental learni ng for audio classiﬁcation,

M. Mulimani and A. Mesaros, “Domain-incremental learni ng for audio classiﬁcation,” in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2025, pp. 1–5

2025

[2] [2]

Unsuper- vised adversarial domain adaptation for acoustic scene cla ssiﬁcation,

S. Gharib, K. Drossos, E. Cakir, D. Serdyuk, and T. Virtan en, “Unsuper- vised adversarial domain adaptation for acoustic scene cla ssiﬁcation,” in W orkshop on Detection and Classiﬁcation of Acoustic Scenes and Events (DCASE), 2018, pp. 138–142

2018

[3] [3]

Overcoming catastrophic forgetting in neural networks,

J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. V eness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, et al. , “Overcoming catastrophic forgetting in neural networks,” in Proceedings of the National Academy of Sciences , vol. 114, no. 13, 2017, pp. 3521– 3526

2017

[4] [4]

icarl: Incremental classiﬁer and representation learning,

S.-A. Rebufﬁ, A. Kolesnikov, G. Sperl, and C. H. Lampert, “icarl: Incremental classiﬁer and representation learning,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 2001–2010

2017

[5] [5]

Episodic memory in lifelong language learning,

C. d. M. d’Autume, S. Ruder, and L. Kong, “Episodic memory in lifelong language learning,” in Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL) , 2019, pp. 5367–5379

2019

[6] [6]

A closer look at class-incre mental learning for multi-label audio classiﬁcation,

M. Mulimani and A. Mesaros, “A closer look at class-incre mental learning for multi-label audio classiﬁcation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 33, pp. 1293–1306, 2023

2023

[7] [7]

Continual learning for on-device environment al sound classiﬁcation,

Y . Xiao, X. Liu, J. King, A. Singh, E. S. Chng, M. D. Plumble y, and W. Wang, “Continual learning for on-device environment al sound classiﬁcation,” in Proceedings of the 7th Detection and Classiﬁcation of Acoustic Scenes and Events 2022 W orkshop (DCASE2022) , Nancy, France, November 2022

2022

[8] [8]

Learning without forgetting,

Z. Li and D. Hoiem, “Learning without forgetting,” IEEE Transactions on pattern analysis and machine intelligence , vol. 40, no. 12, pp. 2935– 2947, 2017

2017

[9] [9]

Con- tinual learning of new sound classes using generative repla y,

Z. Wang, C. Subakan, E. Tzinis, P . Smaragdis, and L. Charl in, “Con- tinual learning of new sound classes using generative repla y,” in IEEE W orkshop on Applications of Signal Processing to Audio and A coustics (WASPAA). IEEE, 2019, pp. 308–312

2019

[10] [10]

Addressing imbalanced domain-incremental learning through dual-balance collab orative experts,

L. Li, D.-W. Zhou, H.-J. Y e, and D.-C. Zhan, “Addressing imbalanced domain-incremental learning through dual-balance collab orative experts,” in F orty-second International Conference on Machine Learnin g, 2025

2025

[11] [11]

Online domain-incrementa l learning for acoustic scene classiﬁcation across locations,

M. Mulimani and A. Mesaros, “Online domain-incrementa l learning for acoustic scene classiﬁcation across locations,” in Proceedings of the IEEE W orkshop on Applications of Signal Processing to Au dio and Acoustics (WASPAA), 2025

2025

[12] [12]

Cla ss-incremental learning for sound event localization and detection,

R. Pandey, M. Mulimani, A. Politis, and A. Mesaros, “Cla ss-incremental learning for sound event localization and detection,” in 2025 IEEE International Conference on Acoustics, Speech, and Signal Processing W orkshops (ICASSPW). IEEE, 2025, pp. 1–5

2025

[13] [13]

Three scenarios for con tinual learning,

G. M. V an de V en and A. S. Tolias, “Three scenarios for con tinual learning,” in Continual Learning W orkshop NeurIPS , 2018

2018

[14] [14]

Acoustic scene classiﬁcat ion across mul- tiple devices through incremental learning of device-spec iﬁc domains,

M. Mulimani and A. Mesaros, “Acoustic scene classiﬁcat ion across mul- tiple devices through incremental learning of device-spec iﬁc domains,” in Proceedings of the DCASE W orkshop , 2024, pp. 81–85

2024

[15] [15]

Audio set: An ontology and h uman- labeled dataset for audio events,

J. F. Gemmeke, D. P . Ellis, D. Freedman, A. Jansen, W. Law rence, R. C. Moore, M. Plakal, and M. Ritter, “Audio set: An ontology and h uman- labeled dataset for audio events,” in IEEE international conference on acoustics, speech and signal processing (ICASSP) . IEEE, 2017, pp. 776–780

2017

[16] [16]

PANNs: Large-scale pretrained audio neural networks for a udio pattern recognition,

Q. Kong, Y . Cao, T. Iqbal, Y . Wang, W. Wang, and M. D. Plumb ley, “PANNs: Large-scale pretrained audio neural networks for a udio pattern recognition,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 28, pp. 2880–2894, 2020

2020