Performance improvement of spatial semantic segmentation with enriched audio features and agent-based error correction for DCASE 2025 Challenge Task 4

Do-Hyeon Lim; Hong Kook Kim; Hyeongcheol Geum; Jeong Eun Lim; Jongyeon Park; Joonhee Lee

REVIEW 3 major objections 8 minor 21 references

Performance improvement of spatial semantic segmentation with enriched audio features and agent-based error correction for DCASE 2025 Challenge Task 4

T0 review · 3 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that adding spectral roll-off and chroma features to an audio-tagging model, cleaning the training data, and applying an agent-based label-correction post-processor improve the challenge's primary source-separation…

desk verdict A solid DCASE challenge report with a plausible 14.7% validation-set CA-SDRi gain, but the absence of test-set results and error bars means the headline number should be read as promising, not confirmed. read the letter →

arxiv 2506.21174 v1 pith:A2MIDWOR submitted 2025-06-26 eess.AS cs.LG

Jongyeon Park , Joonhee Lee , Do-Hyeon Lim , Hong Kook Kim , Hyeongcheol Geum , Jeong Eun Lim This is my paper · ORCID

classification eess.AScs.LG

keywords audiotaggingsourceseparationspatialsemanticsegmentationCA-SDRispectralroll-offchromafeatureagent-basedlabelcorrectiondatasetrefinement

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that spatial semantic segmentation of sound scenes — jointly detecting and separating overlapping sound events in multichannel audio — can be improved without changing the separation network by enriching the tagger's input and cleaning its outputs. The authors add spectral roll-off and chroma features to the mel-based embedding, refine the training mixtures by removing short, ambiguous, or out-of-class samples, and add an agent that re-classifies each separated source and deletes inconsistent labels. On the reported evaluation, the combined system raises the challenge's primary metric, class-aware signal-to-distortion ratio improvement (CA-SDRi), from 11.088 dB to 12.726 dB, a relative gain of 14.7% over the official baseline. The result matters because it shows that task-specific feature design and false-positive control, not just larger separation models, are what drive this competition metric.

What carries the argument

The mechanism that carries the argument is a dual-path feature-fusion tagging head: the pretrained M2D embedding, a masked-modeling audio representation, is concatenated with a 256-dimensional embedding of an auxiliary feature — spectral roll-off projected by multilayer perceptrons, or chroma projected by two convolutional layers — and the combined vector goes to a linear classifier. Around this sits an agent-based label-correction loop: after separation, each estimated source is re-tagged; a label is deleted when the second pass disagrees, and whenever more than three events score above threshold the system re-ranks labels and re-runs separation with the new set. The report also introduces two model-selection metrics, macro-averaged accuracy and false-positive-penalized accuracy $\mathrm{TP}/(\mathrm{TP}+\mathrm{FN}+\mathrm{FP})$, chosen because set-based tagging accuracy correlates poorly with CA-SDRi.

What would settle it

Rerun the full training recipe (four models, refined data, ensemble, agent) with five random seeds and report the spread of validation CA-SDRi; if the seed-to-seed spread exceeds the 1.638 dB gap over the baseline, the claimed 14.7% improvement is within run-to-run noise. Directly, if the official test-set evaluation of the submitted system puts CA-SDRi below 12.0 dB, the claim as stated does not transfer.

Watch

Extended reading notes

Core claim

The central claim is that a system built from the official baseline — a masked-modeling-duo audio-tagging backbone plus a ResUNetK separation network — can be pushed to substantially better CA-SDRi by three complementary interventions. First, the tagging model receives two auxiliary feature streams alongside the M2D embedding: spectral roll-off, the frequency below which a fixed share of spectral energy lies, which highlights impulsive high-frequency events, and chroma, energy across the 12 pitch classes, which separates tonal events such as Doorbell and MusicalKeyboard even when their mel-spectrograms overlap. Second, the training set is curated by removing samples shorter than 1.5 s and manually auditing class consistency, then adding external data for underperforming low-resource classes; this alone yields a 10.9% relative CA-SDRi gain over the baseline. Third, an agent-based label-correction post-processor re-classifies each separated source signal, removes tags whose second-pass label disagrees, and, when extra labels exceed threshold, recomputes top-three labels and re-runs separation; this reduces false positives and adds about 0.2 dB CA-SDRi despite a slight recall loss. A weighted ensemble of four configurations (roll-off plus chroma, chroma, roll-off, retrained baseline) reaches 12.726 dB with the agent versus 11.088 dB for the official baseline checkpoint.

Load-bearing premise

The 14.7% improvement rests on a single run of validation-set numbers obtained after manually cleaning data and tuning ensemble weights on that same validation set; if those choices do not generalize to a fresh test set or other random seeds, the gain can shrink or disappear.

Editorial extensions

If this is right

If the reported validation gain transfers to the test set, a 14.7% relative CA-SDRi improvement over the baseline is available from feature enrichment and post-processing alone, without a new separation architecture.
Because data refinement alone produces a 10.9% relative gain, the quality and intra-class consistency of the synthetic source pool is a first-order lever for this task, not a secondary detail.
The agent's gain despite lower recall demonstrates that CA-SDRi rewards precision more than set-based tagging accuracy; model selection should therefore rank models by false-positive-penalized accuracy.
The four-member weighted ensemble, not any single enriched model, gives the best CA-SDRi (12.726 dB), so the feature streams contribute complementary information rather than being interchangeable.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper: because Table 2 reports a single validation run with manually curated data and validation-tuned ensemble weights, the 14.7% figure should be read as an upper estimate until the official test-set score or repeated-seed runs appear.
A testable extension is to apply the agent-based correction to a held-out subset and measure the false-positive/recall trade-off per class; the report's aggregate numbers leave open whether the agent's gains concentrate in the tonal classes it was designed to disambiguate.
If chroma features are indeed the key to separating Doorbell and MusicalKeyboard, a synthetic experiment with pure tones and controlled pitch overlap should reproduce the separation gain without any dataset curation, which would isolate the feature mechanism from the data-refinement effect.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

You should know this is a straightforward DCASE 2025 Task 4 technical report, and it does what it says. The authors take the official baseline, add spectral roll-off and chroma features to the tagging model, manually prune and augment the training data, and then apply a simple agent-based label-correction post-processor. The ablation table (Table 2) shows each ingredient moving CA-SDRi in the expected direction, ending at 12.726 dB versus 11.088 dB for the baseline checkpoint, which is the reported 14.7% relative improvement. The numbers are internally consistent and the mechanisms are plausible.

The real credit: the paper is honest about being an engineering system. It separates the contributions cleanly, gives a per-class breakdown of the data audit (Table 1), and notes the trade-off between false-positive reduction and recall. The agent-based correction is a nice practical trick, even if it is just re-classifying the separated source with the same tagging model and dropping mismatched labels. The feature combination is not new in the abstract, but the specific application here is a legitimate empirical contribution for the challenge.

The soft spots are exactly what the stress-test flags. Every design decision—the manual data curation, the agent threshold, the ensemble weights—was tuned on the same validation set used for the final numbers. There is one run, no error bars, and no test-set evaluation. So the 14.7% figure could easily shrink on the official test set. That is a real limitation, but it is a standard limitation of challenge technical reports, not a sign of fraud. The authors do not hide it; they just do not emphasize it. The agent's use of the same model for both labels is a minor circularity, but the feature and data-refinement gains are independent of that, so it does not threaten the main result.

Who is this for? Anyone working on DCASE Task 4 or on practical post-processing for audio tagging plus separation will find the ablations useful. It is not a methodological breakthrough, and I would not cite it for novelty, but I would point a student to it as a clean example of how to do a challenge system write-up with honest ablations.

My recommendation: this deserves a serious referee for a workshop or challenge proceedings. The numbers should be checked for internal consistency and the authors should be asked to report test-set results if available, along with error bars or repeated runs. A journal submission would need more rigorous validation, but as a technical report, it is a competent piece of work.

Referee Report

3 major / 8 minor

Summary. This technical report describes three modifications to the DCASE 2025 Challenge Task 4 baseline system for spatial semantic segmentation of sound scenes: (i) enriching the audio-tagging model's input with spectral roll-off and chroma features, (ii) an agent-based label-correction post-processor that re-labels estimated sources and removes inconsistent predictions, and (iii) a training-data refinement procedure that removes short or perceptually heterogeneous samples and adds external samples for two low-resource classes. The authors report that their final weighted ensemble achieves a CA-SDRi of 12.726 dB on the development validation set, versus 11.088 dB for the official baseline checkpoint, a relative improvement of 14.7%. Table 2 provides ablations showing incremental gains from each component.

Significance. If the reported improvement generalizes to the official test set, the combination of enriched features and data curation would be a practically valuable recipe for the S5 task, and the paper's analysis of the weak correlation between tagging accuracy and CA-SDRi is a useful contribution. The paper is transparent in its ablation structure and reports the main metric alongside three tagging metrics. However, the evidence is currently limited to a single run on the validation set that was also used for model selection, so the headline claim is not yet established as a generalizable result.

major comments (3)

[Section 4.2 / Table 2] The central claim of up to 14.7% relative CA-SDRi improvement is computed on the same development validation set that was used to tune the ensemble weights (Section 4.1.1), the agent threshold (Section 3.3), and the manual data-curation decisions (Section 2.2). No official test-set results, cross-validation, or repeated runs are reported. This makes it impossible to rule out that part of the gain reflects selection on the validation set rather than a generalizable improvement. The authors should report evaluation on the official test split if available, or at minimum provide repeated-run statistics with standard deviations and an explicit statement of which design choices were fixed before the reported numbers were obtained.
[Section 4.1.1] The ensemble weights (0.35 for roll-off+chroma, 0.30 for chroma, 0.20 for roll-off, 0.15 for baseline-retrained) are selected based on validation-set performance, and the same validation results are then presented as the final outcome. This selection procedure can inflate the apparent improvement. As a robustness check, the authors should report the performance of an equal-weight ensemble on the same validation set, or use a nested validation scheme, so that the benefit of the weighting procedure itself can be assessed.
[Section 2.2] The data refinement step relies on a manual audit with a subjective criterion—removing 'perceptually heterogeneous samples'—and the paper does not provide a reproducible protocol or a list of the removed samples. While Table 1 gives counts per class, the actual removal decisions are not reproducible, and the magnitude of the reported data-refinement gain (10.9% relative CA-SDRi improvement) could be sensitive to the specific subjective choices. The authors should provide a detailed protocol, release the sample identifiers, or demonstrate robustness of the refinement gain to different manual choices.

minor comments (8)

[Section 3.3] There is a typo 'Lable-1' that should be 'Label-1'.
[References] Reference [5] contains the typo 'chorma' instead of 'chroma'.
[Figure 1 caption] The caption uses 'Door bell' with a space, while the class name elsewhere is 'Doorbell'; please make this consistent.
[Section 3.4.3] The FP-penalized accuracy in Eq. (1) is exactly the Jaccard index for the predicted and ground-truth label sets; citing this connection would help readers situate the metric.
[Abstract] The phrase 'relatively improve' is awkward; consider writing 'improve ... by up to 14.7% relative to the baseline'.
[Section 4.1.1] The sentence 'the spectral roll-off or M2D embedding was processed by two MLPs' is ambiguous because both features are processed in parallel paths; please reword to clarify that each feature is processed by its own two-MLP path.
[Table 2] The row labeled 'Baseline_retrained' has 'Only AT' under the source-separation model column, but this notation is not explained in the text; please clarify what 'Only AT' means (e.g., only the audio-tagging model is retrained, while the separation model remains the baseline checkpoint).
[Section 3.3] The agent uses the same audio-tagging model for Label-1 and Label-2, making the correction a self-consistency check rather than an independent verification. A brief discussion of this limitation and its potential failure modes would strengthen the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: reported gains are measured against an external baseline using an externally defined metric, and the proposed components are empirical modifications rather than quantities defined in terms of the outcome they claim to improve.

full rationale

The paper's central claim is an empirical performance comparison on an externally defined metric, CA-SDRi, against the official DCASE 2025 Challenge Task 4 baseline checkpoint. CA-SDRi is defined by the task organizers, not by this paper's formulas, and the baseline numbers come from the released checkpoint rather than from a quantity constructed by the authors. Each proposed component (spectral roll-off and chroma features, dataset refinement, agent-based label correction) is an architectural, data-processing, or post-processing change evaluated by retraining and measuring the same external metric; no component is defined in terms of the CA-SDRi gain it is claimed to produce. The ensemble weights are tuned on the validation split and Table 2 reports results on that same split, which is a generalization and selection concern, not a definitional circularity: the measured CA-SDRi values are still independent outputs of the evaluation pipeline. The FP-penalized accuracy in Eq. (1) is mathematically the Jaccard index under a new name, but it is used only for model selection and does not enter the CA-SDRi calculation, so it is a naming issue rather than a circular derivation. The agent reuses the same audio-tagging model for Label-1 and Label-2, which may propagate errors, but this is a design limitation, not circularity in the derivation of the reported improvement. There are no load-bearing self-citations, and no fitted parameter is relabeled as the headline result. The derivation chain is therefore self-contained with respect to circularity; the principal outstanding risk is external validity of the validation-only evaluation, which is a correctness concern rather than circular reasoning.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

No new physical entities are introduced. The main unknowns are hand-set hyperparameters, subjective data curation choices, and reliance on external pretrained models and datasets.

free parameters (3)

Ensemble weights = 0.35, 0.30, 0.20, 0.15
Assigned based on validation set performance (Section 4.1.1). These weights tune the final tagging prediction and are not determined a priori.
Short-sample duration cutoff = 1.5 seconds
Samples shorter than 1.5 s were excluded from training in Section 2.2 based on observed training degradation; no formal optimization or external validation is reported.
Agent label threshold = not reported
Section 3.3 uses a pre-defined threshold for including extra labels above top-3; the value is not specified, so the correction agent's operating point is not reproducible.

assumptions (3)

domain assumption The provided M2D pretrained backbone and ResUNetK baseline checkpoint are valid feature extractors and separation models for this task.
The system builds entirely on these released models (Sections 3 and 4.1), and reported gains are measured against the baseline checkpoint.
ad hoc to paper Manual perceptual auditing of source samples improves intraclass consistency without introducing unintended bias.
Section 2.2 removes samples by subjective judgment ('perceptually heterogeneous'), an assumption that cannot be verified from the report.
domain assumption Spectral roll-off and chroma features provide information complementary to mel-spectral embeddings for these event classes.
This motivates the architecture in Sections 3.1 and 3.2; it is supported only by single validation results, not by a formal analysis.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Performance improvement of spatial semantic segmentation with enriched audio features and agent-based error correction for DCASE 2025 Challenge Task 4." pith.science (2026). https://pith.science/paper/A2MIDWOR

@misc{pith2026250621174,
  author       = {Pith},
  title        = {Pith review of: Performance improvement of spatial semantic segmentation with enriched audio features and agent-based error correction for DCASE 2025 Challenge Task 4},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A2MIDWOR}},
  note         = {Machine review of arXiv:2506.21174}
}

read the original abstract

This technical report presents submission systems for Task 4 of the DCASE 2025 Challenge. This model incorporates additional audio features (spectral roll-off and chroma features) into the embedding feature extracted from the mel-spectral feature to im-prove the classification capabilities of an audio-tagging model in the spatial semantic segmentation of sound scenes (S5) system. This approach is motivated by the fact that mixed audio often contains subtle cues that are difficult to capture with mel-spectrograms alone. Thus, these additional features offer alterna-tive perspectives for the model. Second, an agent-based label correction system is applied to the outputs processed by the S5 system. This system reduces false positives, improving the final class-aware signal-to-distortion ratio improvement (CA-SDRi) metric. Finally, we refine the training dataset to enhance the classi-fication accuracy of low-performing classes by removing irrele-vant samples and incorporating external data. That is, audio mix-tures are generated from a limited number of data points; thus, even a small number of out-of-class data points could degrade model performance. The experiments demonstrate that the submit-ted systems employing these approaches relatively improve CA-SDRi by up to 14.7% compared to the baseline of DCASE 2025 Challenge Task 4.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 18 canonical work pages

[1]

This complex task requires a system to identify active sound classes (audio tagging) and to isolate their corresponding anechoic source signals accurately

INTRODUCTION The objective of spatial semantic segmentation of sound scenes (S5) is to jointly detect and separate multiple sound events from multichannel mixes [1][2]. This complex task requires a system to identify active sound classes (audio tagging) and to isolate their corresponding anechoic source signals accurately. The dual nature of this task pre...

work page 2025
[2]

This section outlines the com- position of the official dataset and the specific data curation and augmentation steps to improve the performance of the model

DATASET This work for the DCASE 2025 Challenge Task 4 is based on the provided development dataset, synthesized from various sources using the SpatialScaper toolkit [6]. This section outlines the com- position of the official dataset and the specific data curation and augmentation steps to improve the performance of the model. 2.1. Official dataset compos...

work page 2025
[3]

PROPOSED METHOD In this work, the proposed model is based on the baseline model

work page
[4]

Model training The audio-tagging and source-separation models were trained in- dependently

EXPERIMENTAL RESULTS 4.1. Model training The audio-tagging and source-separation models were trained in- dependently. All experiments were conducted on a single Nvidia RTX A6000 graphics processing unit. 4.1.1. Audio-tagging module The audio-tagging model paired a pretrained M2D [13] backbone with a custom, d ual-path classification head that processed th...

work page 2025
[5]

The proposed strategy combined additional audio feature input (spectral roll-off and chroma), a dataset refinement process, and an agent-based er- ror correction system

CONCLUSION This technical report presen ted the submission systems for the DCASE 2025 Challenge Task 4, focusing on a multifaceted ap- proach to improv e spatial semantic segmentation. The proposed strategy combined additional audio feature input (spectral roll-off and chroma), a dataset refinement process, and an agent-based er- ror correction system. Th...

work page 2025
[6]

SpatialScaper: A library to simulate and augment soundscapes for sound event localization and detection in re- alistic rooms,

I. R. Roman, C. Ick, S. Ding, A. S. Roman, B. McFee, and J. P. Bello, “SpatialScaper: A library to simulate and augment soundscapes for sound event localization and detection in re- alistic rooms,” 2022 IEEE ICASSP , pp. 1221–1225, Mar. 2024

work page 2022
[7]

FSD50K: An open dataset of human -labeled sound events,

E. Fonseca, X . Favory, J. Pons, F. Font, and X. Serra, “FSD50K: An open dataset of human -labeled sound events,” IEEE/ACM Trans. Audio Speech Lang. Process. , vol. 30, pp. 829–852, Jan. 2022

work page 2022
[8]

Description and Discussion on DCASE 2025 Challenge Task 4: Spatial Semantic Segmentation of Sound Scenes,

M. Yasuda et al., “Description and Discussion on DCASE 2025 Challenge Task 4: Spatial Semantic Segmentation of Sound Scenes,” arXiv.org, 2025. https://arxiv.org/abs/2506.10676v1 (accessed Jun. 14, 2025)

arXiv 2025

Show all 21 references

[9]

Baseline systems and evaluation metrics for spatial semantic segmentation of sound scenes,

B. T. Nguyen, M. Yasuda, D. Takeuchi, D. Niizumi, Y. Ohishi, and N. Harada, “Baseline systems and evaluation metrics for spatial semantic segmentation of sound scenes,” arXiv.org, 2025. https://arxiv.org/abs/2503.22088 (accessed Jun. 11, 2025)

2025 arXiv
[10]

of DCASE 2025 Challenge Task 4, Masked modeling duo audio-tagging (M2D-AT) and ResUNetK for audio-tagging and source separation, respectively. Our approach attempts to enhance audio-tagging performance by incorporating various audio fea- tures and applying an agent-based label...

2025
[11]

All audio mixes of 10 seconds long each were generated at a sampling rate of 32 kHz

datasets. All audio mixes of 10 seconds long each were generated at a sampling rate of 32 kHz. The number of active sound events in each mix ranged from one to three, with the signal-to-noise ratio of each event varying between 5 and 20 dB. 2.2. Data refinement and augmentatio...
[12]

Classification in the presence of label noise: A survey,

B. Frénay and M. Verleysen, “Classification in the presence of label noise: A survey,” IEEE Transactions on Neural Net- works, vol. 25, no. 5, pp. 845–869, May 2014

2014
[13]

Acoustic classification and seg- mentation using modified spectral roll-off and variance-based features

M. Kos, Z. Kacic, D. Vlaj, "Acoustic classification and seg- mentation using modified spectral roll-off and variance-based features", Digital Signal Processing, vol. 23, no.2, pp. 659 - 674, 2013

2013
[14]

Making chorma features more robust to timber changes

M. Muller, S. Ewert, S. Kreuzer, "Making chorma features more robust to timber changes", 2009 IEEE International Conference on Acoustics, Speech and Signal Processing , pp.1877-1880, 2009

2009
[15]

EARS: An anechoic fullband speech dataset benchmarked for speech enhancement and dereverbera- tion,

J. Richter et al., “EARS: An anechoic fullband speech dataset benchmarked for speech enhancement and dereverbera- tion,” Interspeech 2022, pp. 4873–4877, Sep. 2024

2022
[16]

Echo-aware adaptation of sound event localization and detection in unknown envi- ronments,

M. Yasuda, Y. Ohishi, and S. Saito, “Echo-aware adaptation of sound event localization and detection in unknown envi- ronments,” 2022 IEEE ICASSP, pp. 226–230, May 2022

2022
[17]

ESC: Dataset for environmental sound classi- fication,

K. J. Piczak, “ESC: Dataset for environmental sound classi- fication,” Proc. 23rd ACM Int. Conf. Multimedia, Oct. 2015

2015
[18]

13, 2025)

https://github.com/nfurnon/disco (accessed Jun. 13, 2025)

2025
[19]

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

J. F. Gemmeke et al., “Audio set: An ontology and human- labeled dataset for audio events,” IEEE Xplore, Mar. 01, 2017. https://ieeexplore.ieee.org/document/7952261 (accessed Oct. 19, 2022)

2017
[20]

Masked modeling duo: Towards a universal audio pre-training fram ework,

D. Niizumi, D. Takeuchi, Y . Ohishi, N. Harada, and K . Kashino, “Masked modeling duo: Towards a universal audio pre-training fram ework,” IEEE/ACM Trans. Audio Speech Lang. Process., pp. 1–15, Jan. 2024

2024
[21]

13, 2025)

https://github.com/nttcslab/dcase2025_task4_baseline (ac- cessed Jun. 13, 2025)

2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

This complex task requires a system to identify active sound classes (audio tagging) and to isolate their corresponding anechoic source signals accurately

INTRODUCTION The objective of spatial semantic segmentation of sound scenes (S5) is to jointly detect and separate multiple sound events from multichannel mixes [1][2]. This complex task requires a system to identify active sound classes (audio tagging) and to isolate their corresponding anechoic source signals accurately. The dual nature of this task pre...

work page 2025

[2] [2]

This section outlines the com- position of the official dataset and the specific data curation and augmentation steps to improve the performance of the model

DATASET This work for the DCASE 2025 Challenge Task 4 is based on the provided development dataset, synthesized from various sources using the SpatialScaper toolkit [6]. This section outlines the com- position of the official dataset and the specific data curation and augmentation steps to improve the performance of the model. 2.1. Official dataset compos...

work page 2025

[3] [3]

PROPOSED METHOD In this work, the proposed model is based on the baseline model

work page

[4] [4]

Model training The audio-tagging and source-separation models were trained in- dependently

EXPERIMENTAL RESULTS 4.1. Model training The audio-tagging and source-separation models were trained in- dependently. All experiments were conducted on a single Nvidia RTX A6000 graphics processing unit. 4.1.1. Audio-tagging module The audio-tagging model paired a pretrained M2D [13] backbone with a custom, d ual-path classification head that processed th...

work page 2025

[5] [5]

The proposed strategy combined additional audio feature input (spectral roll-off and chroma), a dataset refinement process, and an agent-based er- ror correction system

CONCLUSION This technical report presen ted the submission systems for the DCASE 2025 Challenge Task 4, focusing on a multifaceted ap- proach to improv e spatial semantic segmentation. The proposed strategy combined additional audio feature input (spectral roll-off and chroma), a dataset refinement process, and an agent-based er- ror correction system. Th...

work page 2025

[6] [6]

SpatialScaper: A library to simulate and augment soundscapes for sound event localization and detection in re- alistic rooms,

I. R. Roman, C. Ick, S. Ding, A. S. Roman, B. McFee, and J. P. Bello, “SpatialScaper: A library to simulate and augment soundscapes for sound event localization and detection in re- alistic rooms,” 2022 IEEE ICASSP , pp. 1221–1225, Mar. 2024

work page 2022

[7] [7]

FSD50K: An open dataset of human -labeled sound events,

E. Fonseca, X . Favory, J. Pons, F. Font, and X. Serra, “FSD50K: An open dataset of human -labeled sound events,” IEEE/ACM Trans. Audio Speech Lang. Process. , vol. 30, pp. 829–852, Jan. 2022

work page 2022

[8] [8]

Description and Discussion on DCASE 2025 Challenge Task 4: Spatial Semantic Segmentation of Sound Scenes,

M. Yasuda et al., “Description and Discussion on DCASE 2025 Challenge Task 4: Spatial Semantic Segmentation of Sound Scenes,” arXiv.org, 2025. https://arxiv.org/abs/2506.10676v1 (accessed Jun. 14, 2025)

arXiv 2025