On Temporal Guidance and Iterative Refinement in Audio Source Separation

Florian Schmid; Gerhard Widmer; Jonathan Greif; Paul Primus; Tobias Morocutti

REVIEW 3 major objections 5 minor 21 references

On Temporal Guidance and Iterative Refinement in Audio Source Separation

T0 review · 3 major / 5 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 Temporal event guidance and iterative refinement lift audio source separation from an 11.03 to a 13.98 CA-SDRi.

desk verdict Solid, honest DCASE system description with useful ablations, but the iterative-refinement evidence is single-run and confounded by DPRNN, so the flagship claim needs support. read the letter →

arxiv 2507.17297 v1 pith:HJJK3ZKO submitted 2025-07-23 cs.SD cs.AIcs.LGeess.AS

Tobias Morocutti , Jonathan Greif , Paul Primus , Florian Schmid , Gerhard Widmer This is my paper · ORCID

classification cs.SDcs.AIcs.LGeess.AS

keywords audiosourceseparationsoundeventdetectiontemporalconditioningTime-FiLMiterativerefinementDCASE2025Task4spatialsemanticsegmentationCA-SDRi

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 the conventional two-stage recipe for spatial semantic segmentation of sound scenes (clip-level tagging followed by class-conditioned separation) leaves useful temporal information unused, and that replacing the tagger with a frame-level sound event detector (SED) improves both stages. On the DCASE 2025 Task 4 development test set, it reports that feeding the separator time-varying class probabilities and hidden features from a trainable SED model raises class-aware SDR improvement (CA-SDRi) from 11.03 dB (baseline) to 13.42 dB, and that recursively feeding the separator's own output back through the model pushes the score to 13.98 dB. The same SED model also improves clip-level detection accuracy from 59.8% to 67.8%. A sympathetic reader would care because temporal guidance and self-refinement are generic levers that may transfer to other label-conditioned separation tasks without changing the underlying separator.

What carries the argument

The argument runs on four mechanisms. Time-FiLM generalizes FiLM by producing a time-varying sequence of channel-wise scale and shift parameters from frame-level SED probabilities. Embedding Injection aggregates hidden representations from all blocks of the Stage 2 SED model with learned weights and adds them to the ResUNet's latent features. The Dual-Path RNN, two stacked bidirectional GRUs sweeping time and frequency, is intended to capture long-range dependencies. Iterative refinement concatenates the separator's previous single-channel estimate with the original mixture channels and feeds the stack back into the model, with gradients detached between iterations so training memory stays bounded.

What would settle it

If a held-out test set using a different set of room impulse responses and a wider range of event counts and SNRs failed to reproduce the ordering—for instance, if the full temporal-guidance plus iterative-refinement system no longer beat the 11.03 dB baseline, or if DPRNN outperformed the non-DPRNN variant on that test set—the paper's central claim about these improvements would be falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that frame-level sound event detection is not just a better tagger but a better conditioning signal for separation. The authors fine-tune a masked-modeling audio Transformer (M2D) for SED, use attention pooling over its frame predictions as the Stage 1 tagger, and then insert a trainable copy of the same SED model into a ResUNet separator: its frame-level probabilities modulate features in time (Time-FiLM) and its hidden embeddings are injected into the ResUNet bottleneck. Each component contributes, but the trainable Stage 2 SED model gives the largest gain, and the full AudioSep-SED configuration beats the DCASE baseline (11.03) with both the baseline tagger (12.71) and their own tagger (13.42), and with oracle targets (15.29). Adding a Dual-Path RNN helps on validation but not on the test split; iterative refinement, by contrast, improves test performance, with a model trained for up to three iterations reaching 13.98 after ten inference iterations.

Load-bearing premise

The synthetic development mixtures, built with only five training room impulse responses, three validation RIRs, and narrow per-event SNR ranges, faithfully predict performance on the official test set; the DPRNN's validation-to-test reversal shows this transfer can fail.

Editorial extensions

If this is right

Frame-level SED labels, not just clip tags, can improve event detection accuracy by eight points on this task, suggesting that temporal labels are worth collecting even when only clip-level decisions are needed.
Temporal conditioning from a trainable SED model improves separation regardless of whether Stage 1 predictions come from the baseline tagger, the authors' tagger, or oracle targets, so the gain is not an artifact of better tags alone.
Retraining the separator on its own outputs for a few iterations yields further improvements; the paper's best system reaches 13.98 with ten inference iterations even though training used at most three.
The DPRNN's validation-to-test discrepancy shows that components selected purely on the synthetic development set can hurt on the official test mixtures, so validation score alone is not a safe guide within this challenge.

Reading between the lines

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

If the validation-to-test mismatch shown by the DPRNN is general, other development-set choices (SNR ranges, number of room impulse responses, event counts) may also transfer poorly; a cheap test would be to synthesize validation mixtures from held-out RIRs and re-rank configurations.
Time-FiLM and embedding injection are not specific to the DCASE task, so they could be dropped into language-conditioned or text-prompt separators; because the gains appear conditional on trainable guidance, the cost is training the tagger jointly with the separator.
The odd-even oscillation in iterative refinement recalls an oscillating feedback loop; a learned stopping rule or averaging across consecutive iterations might stabilize the gains.

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

Two things you should know. This is a clean, openly disclosed DCASE 2025 Task 4 system description: fine-tune M2D for sound event detection, use a trainable copy to inject frame-level guidance (Time-FiLM + Embedding Injection) into a ResUNet separator, add an optional DPRNN, and apply iterative refinement. The engineering result is real: CA-SDRi goes from 11.03 (baseline) to 13.42 with their M2D tagger, they finished second in the challenge, and code plus checkpoints are released. The ablation in Table 1 is genuinely well done, with oracle columns and per-component removals, averaged over two runs. The temporal guidance contribution is the most convincing part.

The soft spot is iterative refinement. Figure 2 compares refinement models only to AudioSep-SED+DPRNN, which scores 13.31 on test, while AudioSep-SED without DPRNN scores 13.42. So the claimed 13.98 after refinement may be mostly recovering the DPRNN-induced test degradation. They never run refinement on top of the better base, so a general refinement benefit is not established. The stress-test note is on target. Also, every refinement condition is a single run with no error bars, and the paper itself warns about random variability, yet the abstract says “significant improvements.” That is stronger than the evidence. The DPRNN saga is a known confound: it helps validation (13.79 to 14.35) but hurts test (13.42 to 13.31), and they keep it because they selected on validation. That is standard for a challenge submission, but it weakens the scientific claim about which components actually help.

Minor: the synthetic dev set uses only 5 training and 3 validation RIRs with narrow SNR ranges, so transfer to the official test conditions is a genuine limitation, though the paper does not overclaim generalizability.

Who benefits: someone working on DCASE S5 or on label-conditioned audio separation with temporal conditioning. It deserves a serious referee, but the revision should add repeated runs or confidence intervals for refinement, compare refinement with and without DPRNN, and temper the abstract. The core temporal-guidance result is solid and the code release is a plus.

Referee Report

3 major / 5 minor

Summary. The manuscript describes a two-stage system for DCASE 2025 Task 4 (S5): a fine-tuned M2D-based SED model for clip-level detection, and an AudioSep-initialized ResUNet separator augmented with Time-FiLM, embedding injection, a trainable Stage 2 SED, an optional Dual-Path RNN, and iterative refinement. The main evidence is Table 1 (two-run means) showing that the temporal-guidance components improve CA-SDRi over the DCASE baseline, and Figure 2 (single-run curves) showing iterative refinement on top of AudioSep-SED + DPRNN. The central claim is that temporal SED guidance and iterative refinement improve both event detection and source separation.

Significance. If the results hold, this is a useful, reproducible ablation study of a strong challenge system: it decomposes the contributions of Time-FiLM, embedding injection, trainable Stage 2 SED, and DPRNN, uses oracle conditions to isolate the separator's behavior, and releases code and checkpoints. The temporal-guidance claim is well supported by Table 1, which reports two-run means with ablations and oracle variants. However, as presented, the iterative-refinement claim is not supported at the same level: the only refinement comparisons are against a base configuration that is worse on the test split than the same model without DPRNN, and all refinement curves in Figure 2 are single runs without error bars. The paper is therefore significant as a system description, but the headline claims about iterative refinement need additional experiments or substantially softened wording.

major comments (3)

[§6, Figure 2 and Table 1] The iterative-refinement comparison is confounded by the choice of base model. Table 1 shows that AudioSep-SED without DPRNN reaches 13.42 ± 0.11 on the development test split, while AudioSep-SED + DPRNN reaches only 13.31 ± 0.07. Figure 2 reports refinement only on the DPRNN variant, so the best refinement result (13.98) is compared with 13.31 instead of with the stronger 13.42 baseline. The apparent gain may largely be recovery of the test-set degradation caused by DPRNN. To support the iterative-refinement claim, the authors should train and evaluate iterative refinement on AudioSep-SED without DPRNN and compare the refinement curves against that stronger base.
[§6, Figure 2] Each iterative-refinement configuration (Iter 2, 3, and 4) is a single run with no error bars, and the paper itself cautions that the results "may be influenced by random variability." The abstract's "significant improvements" and the text's "clearly outperforms" are stronger than this evidence warrants. The odd-even oscillation between consecutive inference iterations is consistent with high variance rather than a stable refinement effect. Please report multiple seeds or at least error bars, and temper the wording accordingly.
[§5.5 and §6, model selection] The DPRNN was included because it improved validation CA-SDRi from 13.79 to 14.35, yet it degraded test performance (13.42 to 13.31). This demonstrates that validation-to-test transfer is unreliable in this setup, and Figure 2 is evaluated only on the test split. It is not clear whether the choice of Iter 3, or the number of inference iterations, was made on validation or test; if any selection used the test split, the reported 13.98 is optimistic. The paper should either report the corresponding validation curves for Iter {2,3,4} or explicitly state the selection protocol.

minor comments (5)

[§5.1] In the dataset paragraph, "class distribution inbalance" should be "class distribution imbalance."
[§5.4] In the source separation model paragraph, "with an batch size of 8" should be "with a batch size of 8."
[§6, first paragraph] The 8-point improvement in Stage 1 (59.8% to 67.8%) is reported as "test set accuracy," but the exact metric (e.g., clip-level F1 or accuracy) is not defined; please specify it.
[References [4] and [21]] References [4] and [21] point to the same paper; please cite it once or clearly distinguish the two versions.
[§7, Conclusion] The conclusion that iterative refinement "progressively enhanced the separation quality" overstates the single-run evidence in Figure 2, which shows inconsistent and occasionally negative changes between consecutive inference iterations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all central claims are empirical comparisons against external DCASE baselines and oracle conditions.

full rationale

This paper is an empirical engineering report. The claims are measured CA-SDRi and tagging-accuracy numbers on the externally defined DCASE 2025 Task 4 development test split, compared against the challenge baseline, ablation variants, oracle targets, and the official validation/test split. The only self-citations are [6], a pointer to the authors' challenge technical report describing the same system, and [7], an externally published ICASSP 2025 result on AudioSet Strong used to initialize the SED backbone; neither functions as an unverified premise that predetermines the DCASE outcome. The Stage 2 SED model is initialized from the Stage 1 model but then trained jointly with the separator, and the separator predictions are evaluated against ground-truth sources, so no fitted parameter is renamed as a prediction. Iterative refinement is a recurrent inference scheme whose target is the true separated source, not the model's own previous output. The paper itself flags the single-run nature of the iterative-refinement experiments and the DPRNN validation/test mismatch; those are statistical-validity caveats, not circularity. No equation in the paper defines the reported output in terms of its own fit, and no load-bearing argument reduces to a self-citation chain.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities or mathematical objects. It relies on a standard linear additive mixture model, an approximation for direct-path signals, a synthetic data simulation assumption, and an architectural assumption about the usefulness of SED hidden features. Key hyperparameters were chosen by hand or selected by validation.

free parameters (4)

Loss weighting lambda = 0.5
Fixed balance between strong and weak BCE losses in Equation (6). Chosen by hand and used for all experiments, so all downstream results depend on it.
Maximum training iterations for iterative refinement = 2, 3, or 4; Iter3 selected
Configurations compared in Figure 2. Iter3 was chosen as the best balance between complexity and quality. Each is a single run, so the selection is partly sensitive to random variability.
Inference iteration count = up to 10; best around 3 to 10
Performance varies with the number of refinement steps. The paper selects the count after observing convergence, which is a post hoc choice that affects the reported final score.
Separator hop size = 160 samples instead of 320 in AudioSep pretraining
Section 5.4 states 'we found a hop size of 160 samples to perform better'. This manual design decision changes the time-frequency resolution and alignment of the separator.

assumptions (4)

domain assumption Each microphone signal is a linear convolution of dry sources with room impulse responses plus additive noise, as in Equation (1).
Standard S5 mixture model stated in Section 2. If the physical mixture is not linear and additive, the formulated objective and the separation approach do not hold.
domain assumption Direct-path signals for evaluation can be approximated by windowing the full RIR around the first significant energy peak.
Section 2 introduces this approximation to relate recorded reverberant mixtures to the anechoic target used in SDR computation. The evaluation metric depends directly on this approximation.
domain assumption Synthetic SpatialScaper mixtures with fixed SNR ranges and limited RIR variety are representative of real evaluation conditions.
Section 5.1 describes all training and validation data as synthesized on the fly with 5 training RIRs and 3 validation RIRs. The DPRNN result in Section 6 shows this assumption can fail when transferring from validation to test.
ad hoc to paper Stage 2 SED hidden features carry temporally aligned semantics useful for separation.
This motivates Time-FiLM and Embedding Injection in Sections 4.1 and 4.2. It is supported only by the paper's ablations, not by an external independent result.

how reviews work

0 comments

Cite this review

Pith. "Pith review of On Temporal Guidance and Iterative Refinement in Audio Source Separation." pith.science (2026). https://pith.science/paper/HJJK3ZKO

@misc{pith2026250717297,
  author       = {Pith},
  title        = {Pith review of: On Temporal Guidance and Iterative Refinement in Audio Source Separation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HJJK3ZKO}},
  note         = {Machine review of arXiv:2507.17297}
}

read the original abstract

Spatial semantic segmentation of sound scenes (S5) involves the accurate identification of active sound classes and the precise separation of their sources from complex acoustic mixtures. Conventional systems rely on a two-stage pipeline - audio tagging followed by label-conditioned source separation - but are often constrained by the absence of fine-grained temporal information critical for effective separation. In this work, we address this limitation by introducing a novel approach for S5 that enhances the synergy between the event detection and source separation stages. Our key contributions are threefold. First, we fine-tune a pre-trained Transformer to detect active sound classes. Second, we utilize a separate instance of this fine-tuned Transformer to perform sound event detection (SED), providing the separation module with detailed, time-varying guidance. Third, we implement an iterative refinement mechanism that progressively enhances separation quality by recursively reusing the separator's output from previous iterations. These advancements lead to significant improvements in both audio tagging and source separation performance, as demonstrated by our system's second-place finish in Task 4 of the DCASE Challenge 2025. Our implementation and model checkpoints are available in our GitHub repository: https://github.com/theMoro/dcase25task4 .

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 18 canonical work pages

[1]

Description and discussion on dcase 2025 challenge task 4: Spatial semantic segmentation of sound scenes,

M. Yasuda, B. T. Nguyen, N. Harada, R. Serizel, M. Mishra, M. Delcroix, S. Araki, D. Takeuchi, D. Niizumi, Y . Ohishi et al., “Description and discussion on dcase 2025 challenge task 4: Spatial semantic segmentation of sound scenes,” arXiv preprint arXiv:2506.10676 , 2025

arXiv 2025
[2]

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 preprint arXiv:2503.22088 , 2025

work page Pith review arXiv 2025
[3]

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

D. Niizumi, D. Takeuchi, Y . Ohishi, N. Harada, and K. Kashino, “Masked modeling duo: Towards a universal audio pre-training framework,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , 2024

work page 2024
[5]

Film: Visual reasoning with a general conditioning layer,

E. Perez, F. Strub, H. de Vries, V . Dumoulin, and A. C. Courville, “Film: Visual reasoning with a general conditioning layer,” in Proc. of the AAAI conference on artificial intelligence , 2018

work page 2018
[6]

Transformer- aided audio source separation with temporal guidance and iterative refinement,

T. Morocutti, F. Schmid, J. Greif, P. Primus, and G. Widmer, “Transformer- aided audio source separation with temporal guidance and iterative refinement,” DCASE2025 Challenge, Tech. Rep., 2025

work page 2025
[7]

Effective pre-training of audio transformers for sound event detection,

F. Schmid, T. Morocutti, F. Foscarin, J. Schl ¨uter, P. Primus, and G. Widmer, “Effective pre-training of audio transformers for sound event detection,” in Proc. of the International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2025

work page 2025
[8]

M2D-CLAP: masked modeling duo meets CLAP for learning general-purpose audio-language representation,

D. Niizumi, D. Takeuchi, Y . Ohishi, N. Harada, M. Yasuda, S. Tsubaki, and K. Imoto, “M2D-CLAP: masked modeling duo meets CLAP for learning general-purpose audio-language representation,” in Proc. of the Interspeech Conference, 2024

work page 2024
[9]

Dual-path RNN: efficient long sequence modeling for time-domain single-channel speech separation,

Y . Luo, Z. Chen, and T. Yoshioka, “Dual-path RNN: efficient long sequence modeling for time-domain single-channel speech separation,” in Proc. of the International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020

work page 2020

Show all 21 references

[10]

Flowsep: Language-queried sound separation with rectified flow matching,

Y . Yuan, X. Liu, H. Liu, M. D. Plumbley, and W. Wang, “Flowsep: Language-queried sound separation with rectified flow matching,” in Proc. ICASSP, 2025

2025
[11]

Soundbeam meets m2d: Target sound extraction with audio foundation model,

C. Hernandez-Olivan, M. Delcroix, T. Ochiai, D. Niizumi, N. Tawara, T. Nakatani, and S. Araki, “Soundbeam meets m2d: Target sound extraction with audio foundation model,” in Proc. of the International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2025

2025
[12]

Learning phrase representations using RNN encoder-decoder for statistical machine translation,

K. Cho, B. van Merrienboer, C ¸. G ¨ulc ¸ehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio, “Learning phrase representations using RNN encoder-decoder for statistical machine translation,” in Proc. of the Conference on Empirical Methods in Natural Language Processing ...

2014
[13]

Spatial scaper: A library to simulate and augment soundscapes for sound event localization and detection in realistic rooms,

I. R. Rom ´an, C. Ick, S. Ding, A. S. Roman, B. McFee, and J. P. Bello, “Spatial scaper: A library to simulate and augment soundscapes for sound event localization and detection in realistic rooms,” in Proc. of the International Conference on Acoustics, Speech and Signal Proce...

2024
[14]

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 Transactions on Audio, Speech and Language Processing , 2022

2022
[15]

EARS: an anechoic fullband speech dataset bench- marked for speech enhancement and dereverberation,

J. Richter, Y . Wu, S. Krenn, S. Welker, B. Lay, S. Watanabe, A. Richard, and T. Gerkmann, “EARS: an anechoic fullband speech dataset bench- marked for speech enhancement and dereverberation,” in Proc. of the Interspeech Conference, 2024

2024
[16]

Semantic hearing: Programming acoustic scenes with binaural hearables,

B. Veluri, M. Itani, J. Chan, T. Yoshioka, and S. Gollakota, “Semantic hearing: Programming acoustic scenes with binaural hearables,” in Proc. of the ACM Symposium on User Interface Software and Technology, UIST, 2023

2023
[17]

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

M. Yasuda, Y . Ohishi, and S. Saito, “Echo-aware adaptation of sound event localization and detection in unknown environments,” in Proc. of the International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022

2022
[18]

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

J. F. Gemmeke, D. P. W. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter, “Audio set: An ontology and human-labeled dataset for audio events,” in Proc. of the International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2017

2017
[19]

The benefit of temporally-strong labels in audio event classification,

S. Hershey, D. P. W. Ellis, E. Fonseca, A. Jansen, C. Liu, R. C. Moore, and M. Plakal, “The benefit of temporally-strong labels in audio event classification,” in Proc. of the International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2021

2021
[20]

Adam: A method for stochastic optimization,

D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in Proc. ICLR, 2015

2015
[21]

Universal source separation with weakly labelled data,

Q. Kong, K. Chen, H. Liu, X. Du, T. Berg-Kirkpatrick, S. Dubnov, and M. D. Plumbley, “Universal source separation with weakly labelled data,” arXiv preprint arXiv:2305.07447 , 2023

2023 arXiv
[22]

Separate anything you describe,

X. Liu, Q. Kong, Y . Zhao, H. Liu, Y . Yuan, Y . Liu, R. Xia, Y . Wang, M. D. Plumbley, and W. Wang, “Separate anything you describe,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , 2024

2024

Pith tools

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

[1] [1]

Description and discussion on dcase 2025 challenge task 4: Spatial semantic segmentation of sound scenes,

M. Yasuda, B. T. Nguyen, N. Harada, R. Serizel, M. Mishra, M. Delcroix, S. Araki, D. Takeuchi, D. Niizumi, Y . Ohishi et al., “Description and discussion on dcase 2025 challenge task 4: Spatial semantic segmentation of sound scenes,” arXiv preprint arXiv:2506.10676 , 2025

arXiv 2025

[2] [2]

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 preprint arXiv:2503.22088 , 2025

work page Pith review arXiv 2025

[3] [3]

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

D. Niizumi, D. Takeuchi, Y . Ohishi, N. Harada, and K. Kashino, “Masked modeling duo: Towards a universal audio pre-training framework,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , 2024

work page 2024

[4] [5]

Film: Visual reasoning with a general conditioning layer,

E. Perez, F. Strub, H. de Vries, V . Dumoulin, and A. C. Courville, “Film: Visual reasoning with a general conditioning layer,” in Proc. of the AAAI conference on artificial intelligence , 2018

work page 2018

[5] [6]

Transformer- aided audio source separation with temporal guidance and iterative refinement,

T. Morocutti, F. Schmid, J. Greif, P. Primus, and G. Widmer, “Transformer- aided audio source separation with temporal guidance and iterative refinement,” DCASE2025 Challenge, Tech. Rep., 2025

work page 2025

[6] [7]

Effective pre-training of audio transformers for sound event detection,

F. Schmid, T. Morocutti, F. Foscarin, J. Schl ¨uter, P. Primus, and G. Widmer, “Effective pre-training of audio transformers for sound event detection,” in Proc. of the International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2025

work page 2025

[7] [8]

M2D-CLAP: masked modeling duo meets CLAP for learning general-purpose audio-language representation,

D. Niizumi, D. Takeuchi, Y . Ohishi, N. Harada, M. Yasuda, S. Tsubaki, and K. Imoto, “M2D-CLAP: masked modeling duo meets CLAP for learning general-purpose audio-language representation,” in Proc. of the Interspeech Conference, 2024

work page 2024

[8] [9]

Dual-path RNN: efficient long sequence modeling for time-domain single-channel speech separation,

Y . Luo, Z. Chen, and T. Yoshioka, “Dual-path RNN: efficient long sequence modeling for time-domain single-channel speech separation,” in Proc. of the International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020

work page 2020