Pith. sign in

REVIEW 3 major objections 5 minor 21 references

Unleashing the Power of Natural Audio Featuring Multiple Sound Sources

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read ClearSep converts naturally mixed audio into single-source training tracks, beating artificial-mixture training on sound separation benchmarks.

desk verdict ClearSep's data engine idea is worth a serious referee, but the remix filter is unvalidated, the benchmark gains are small, and one table contradicts itself. read the letter →

arxiv 2504.17782 v1 pith:UOS3IJN5 submitted 2025-04-24 cs.SD cs.LG

Xize Cheng , Slytherin Wang , Zehan Wang , Rongjie Huang , Tao Jin , Zhou Zhao This is my paper · ORCID
classification cs.SDcs.LG
keywords universalsoundseparationquery-basedsourcenaturalaudiodataengineremixconsistencyself-supervisedtrainingsilenceaugmentationSet
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

The paper claims that naturally mixed audio, which normally cannot be used for supervised sound separation because no clean reference tracks exist, can be turned into high-quality single-source training data by an iterative data engine. The engine separates each mixture into label-specific tracks, keeps only tracks whose remix closely reconstructs the original audio, and retrains the separation model on the accepted tracks. This loop is claimed to improve query-based separation beyond training on the same 5,800 hours of artificially mixed audio, with Table 1 reporting 11.23 dB SDRi on AudioSet versus 10.63 dB for the strongest baseline. The relevance is practical: real-world audio separation currently lacks ground truth, so a way to mine natural recordings for training would remove the main bottleneck to scaling.

What carries the argument

The load-bearing object is the iterative data engine with remix-based filtering. Re-SDR and Re-SISDR compare the original mixture $X$ with the remix $\bar X = \sum_i \hat X_i$ of the separated tracks, computing signal-to-distortion ratios exactly as in standard SDR and SISDR but with the remix standing in for the ground-truth source. These two metrics are simultaneously an unsupervised evaluation, a data filter that admits tracks only above threshold, and a diagnostic signal for the loop; mutually exclusive class labels act as negative queries so tracks are forced apart, and silence augmentation trains the model to output zero for labels absent from the mixture.

What would settle it

On a set of synthetic mixtures whose true sources are known, compute Re-SDR and Re-SISDR for the data engine's accepted tracks and compare them with true SDRi: if many accepted tracks score above the 15 dB threshold while their true separation SDRi is near zero, or if a one-track-copies-mixture solution passes the filter, the remix proxy is not measuring separation.

Watch

Extended reading notes

Core claim

The central discovery claim is that remix consistency is a usable, unsupervised signal for curating training data from real-world mixtures. ClearSep uses mutually exclusive AudioSet labels as positive and negative queries, separates each natural clip into tracks, and admits a track into the training set only when Re-SDR and Re-SISDR between the original mixture and the sum of the separated tracks exceed thresholds (15 dB for self-separate training, 10 dB for independent-track training). Iterating this data-engine/training loop raised AudioSet SDRi from 9.56 dB after the first pass to 10.45 dB after the third, and the final model reached 11.23 dB SDRi on AudioSet and 14.16 dB on ESC-50 with positive-plus-negative queries, surpassing CLAPSep trained on the same 5,800 hours. The paper also claims that silence augmentation makes absent-query outputs nearly silent, with Silence-SDR jumping to 107.09 dB for positive-plus-negative queries.

Load-bearing premise

The loop assumes that if the separated tracks add back up to the original mixture, then each track has isolated its own sound event; a degenerate output where one track carries the entire mixture and all others are silent would also remix perfectly, so the filter's quality signal is only as good as that assumption.

Editorial extensions

If this is right

  • Label-queried separation can be scaled directly from raw, naturally mixed audio: the data engine added thousands of hours of separated tracks across iterations without human annotation of clean sources.
  • Training on single-source separated tracks outperforms training on mixed tracks at equal scale, supporting the decision to decompose rather than keep mixtures.
  • Iterating the data-engine/training loop yields steady gains: AudioSet SDRi rises from 9.56 to 10.45 dB across three iterations.
  • Silence augmentation gives the model a usable “not present” behavior, which is what makes separated tracks clean enough to trust as training data.

Reading between the lines

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

  • A natural extension would be to validate the Re-SDR thresholds against known ground truth on synthetic mixtures; if the thresholds correlate strongly with true SDRi, the data engine could be trusted on unlabeled audio.
  • Because Re-SDR only checks that tracks sum to the mixture, a degenerate separation that puts the entire mix in one track and silence in the others would pass the filter; adding a per-track energy or sparsity prior, or a cross-track disjointness term, would close that loophole.
  • The same remix-consistency criterion could be applied beyond labels to caption-based queries, where the mutually exclusive negative-query structure is less explicit, by clustering captions or using an ontology to define exclusions.
  • If the loop can bootstrap from a weak initial separator, the approach points toward purely self-supervised dataset curation for general audio, not just the AudioSet label set.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. ClearSep extends query-based universal sound separation with an iterative data engine. Starting from a CLAPSep-style model, it uses AudioSet category labels as positive and negative text queries to separate naturally mixed clips into per-event tracks, filters the resulting tracks with remix-based Re-SDR and Re-SISDR thresholds, adds the accepted tracks to the training set under two strategies (Independent Track Training and Self-Separate Training), and retrains the model in a two-stage loop. It also proposes silence augmentation so that the model outputs silence when the query label is absent. Experiments report improvements over CLAPSep on AudioCaps, AudioSet, and ESC-50, large gains in Silent-SDR on AudioCaps, and better Re-SDR/Re-SISDR on natural AudioSet mixtures. The central claim is that naturally mixed audio can be converted into usable single-source training data without manual supervision, yielding state-of-the-art separation performance.

Significance. If the data-engine claim holds, this is a genuinely useful contribution: it offers an automated way to turn weakly labeled natural recordings into single-source training material and an unsupervised monitor for separation quality. The pipeline is concrete, the writing is clear, and the silence-augmentation result in Table 3 is a striking, falsifiable positive result. However, the reported benchmark deltas are small relative to the per-sample standard deviations, and both the data-engine filter and the real-world evaluation rest on a remix-consistency proxy that has not been validated against ground-truth isolation. The significance of the paper therefore depends on additional validation rather than being established by the current evidence.

major comments (3)
  1. [§3.3.1 and Appendix B.3, Eqs. (12)–(15)] Re-SDR and Re-SISDR certify remix consistency, not source isolation. Because each track is reconstructed by applying a mask to the mixture spectrogram (Eq. 4), any set of masks that sums to the all-ones mask will remix exactly to the input; one track carrying the full mixture and all other tracks silent achieves perfect Re-SDR and Re-SISDR. The admission thresholds in §4.1 and Appendix A.2 therefore do not establish that an accepted track contains only its target event, and Table 4 evaluates real-world performance with the same unvalidated proxy. The paper should validate Re-SDR/Re-SISDR against ground-truth SDR on labeled or synthetic mixtures, or add a target-specific isolation check that is not satisfied by the mask-sum degeneracy.
  2. [Tables 1 and 2] The headline improvements are small relative to the reported standard deviations: for example, 11.23±5.07 versus 10.63±5.17 dB SDRi on AudioSet and 14.16±6.29 versus 13.44±6.43 dB SDRi on ESC-50. No significance tests, paired statistics, or confidence intervals are reported. Because the same evaluation items are scored across models, paired tests (for example, Wilcoxon signed-rank or bootstrap on per-sample deltas) would settle whether the differences are reliable rather than within run-to-run variability.
  3. [Table 2] The iteration comparison changes both data composition and data volume (5,800 → 9,429 → 10,288 hours of tracks) and adds an extra fine-tuning stage at each iteration. The reported monotone improvement is therefore not a clean test of the data engine's quality filtering: it could be driven by larger data volume or longer training. A matched control that uses an equal number of unfiltered or randomly selected separated tracks is needed to attribute the gain to the proposed filtering mechanism.
minor comments (5)
  1. [§4.4 and Table 4] The sentence 'surpassing ClearSep (16.764 dB)' should read 'surpassing CLAPSep (16.764 dB)' to match the table row.
  2. [Appendix B.2] The definition of the silent reference is malformed: 'Xsilent =′' should use a proper mathematical notation for the zero signal.
  3. [Appendix B, Eqs. (6)–(15)] The equations mix |·|² and ‖·‖² notation for signal energy; please use consistent norm notation throughout.
  4. [Abstract and Contributions] Calling Re-SDR and Re-SISDR the 'first' unsupervised evaluation metrics for sound separation is overstated given remix-consistency ideas in earlier unsupervised separation literature; please qualify this claim.
  5. [Table 1 and Related Work] The state-of-the-art claim is supported only by a small baseline set; OmniSep and GASS are cited in the text but not compared in Table 1, so the SOTA claim should be narrowed or the missing comparisons added.

Circularity Check

2 steps flagged · score 6.0 of 10

Real-world evaluation (Table 4) uses the same remix-based metrics that serve as data-engine admission thresholds, and those metrics reduce to mask-sum reconstruction by construction; Table 1's standard SDRi results are independent, so circularity is partial.

  1. fitted input called prediction [Section 3.3.1 (Remix-based Filter), Section 4.1 (Implementation Details), Section 4.4 (Table 4), Appendix B.3 (Eq. 12-15)]
    "we propose a filtering mechanism that selectively incorporates high-quality separated tracks into the training dataset... Re-SDR and Re-SISDR are used to measure the signal-to-distortion ratio and the scale-invariant signal-to-distortion ratio, respectively, between the natural audio X and the re-mixed audio ̄X. ... for a sample to be included in SST, it must satisfy both Re-SDR> 15 dB and Re-SISDR> 15 dB. ..."

    The same unsupervised metrics define the data-engine admission thresholds (Re-SDR and Re-SISDR > 10/15 dB in Section 4.1) and are then used as the sole evidence of real-world separation quality in Table 4. The model is trained on tracks that already passed the Re-SDR filter, so its high Re-SDR/Re-SISDR on real-world data partly measures the curation criterion itself rather than an independently verified separation property. The paper does not validate Re-SDR against ground-truth SDR on labeled mixtures, so this evaluation loop is closed by construction rather than by external evidence.

  2. self definitional [Appendix B.3 (Eq. 12-13) combined with Eq. 4 (mask-based reconstruction)]
    "Given a mixed-source audio ̄X, our model separates it into multiple independent tracks{̂X1, ̂X2,..., ̂Xn}. These separated tracks are then remixed to obtain the estimated mixture: ̄X = Σi=1..n ̂Xi. ... Re-SDR(̄X,X) = 10 log10 ‖X‖^2 / ‖X−̄X‖^2."

    Re-SDR is defined purely as reconstruction fidelity of the summed separated tracks. Since each track is reconstructed as ISTFT(M_i ⊙ |̃X| e^{jΦ}) (Eq. 4), the remix equals ISTFT((Σ M_i) ⊙ |̃X| e^{jΦ}). Any mask family with Σ M_i = 1, including the degenerate case of one mask equal to the all-ones mask and all others zero, yields ̄X = X exactly and Re-SDR = Re-SISDR = +∞. Thus the 'mutually exclusive and collectively exhaustive' principle (Appendix B.3) reduces by construction to collective exhaustiveness alone; the metric cannot certify that individual tracks isolate distinct events, yet it is used as the data-engine filter and as the real-world evaluation metric.

full rationale

The central benchmark claim (Table 1) is grounded in standard SDRi/SISDRi against human-annotated or synthetic ground-truth mixtures, so ClearSep's improvement over CLAPSep on AudioSet and ESC-50 is not itself circular. The circularity is confined to the real-world evaluation chain: the remix-based metrics that define the data-engine admission thresholds are reused as the real-world evaluation metric, and those metrics are definitionally insensitive to whether individual tracks isolate distinct sources. This makes the real-world robustness claim (Table 4) partially a re-statement of the filter criterion. No load-bearing self-citation chain or imported uniqueness theorem appears in the paper; the authors' self-citations (e.g., Cheng et al. 2024) are not used to justify the core derivation. The score is 6 rather than higher because the headline SOTA numbers and the iterative-improvement table (Table 2) rely on independent ground-truth metrics, so the derivation is only partially circular.

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

The central claim depends on four tuned quantities (two remix thresholds, a silence rate, and a stopping iteration) and on three domain assumptions: labels are exhaustive, audio mixes additively, and remix consistency tracks true separation quality. No new physical entities are posited. The remix-consistency assumption is the least supported because a degenerate separation can satisfy it without isolating individual events.

free parameters (4)
  • Remix-based threshold for Independent Track Training (ITT) = Re-SDR > 10 dB and Re-SISDR > 10 dB
    Chosen in Table 6; lower threshold admits more tracks (298h) but reduces average quality; R2/R6 use this.
  • Remix-based threshold for Self-Separate Training (SST) = Re-SDR > 15 dB and Re-SISDR > 15 dB
    Chosen in Table 6; higher threshold admits fewer high-purity tracks (29h in R5/R6); R5 shows +0.06 dB over 10 dB SST threshold.
  • Silence augmentation rate alpha = 0.05
    Selected from Table 5 sweep; alpha=0.05 gives best SDRi 9.21 vs 9.07 at alpha=0.
  • Number of data engine iterations = 3
    Iteration 3 improves AudioSet SDRi from 9.56 to 10.45 dB; the paper does not test iteration 4 or report convergence, so stopping is arbitrary.
assumptions (4)
  • domain assumption AudioSet labels per clip are mutually exclusive and collectively exhaustive after ontology filtering, so tracks sum to the original mixture.
    Section A.2 filters parent labels to enforce mutual exclusivity, but does not verify exhaustiveness; unlabeled co-occurring events would violate remix consistency.
  • ad hoc to paper Remix consistency (Re-SDR/Re-SISDR) indicates true separation quality; low reconstruction error between remixed tracks and original implies correct per-track separation.
    Appendix B.3 Eq (12)-(15); a degenerate output where one track carries the full mixture and others are silent also yields high Re-SDR, so the criterion is necessary but not sufficient.
  • domain assumption Audio events mix additively and can be represented as independent tracks (linear instantaneous mixing).
    Used throughout Section 3.3 and Eq (12); reverberant or overlapping events may violate additivity.
  • domain assumption CLAP text embeddings and the CLAPSep architecture provide adequate query conditioning for labels.
    Architecture and frozen CLAP encoder from Ma et al. 2024 are assumed; no experiments ablate this reliance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unleashing the Power of Natural Audio Featuring Multiple Sound Sources." pith.science (2026). https://pith.science/paper/UOS3IJN5

@misc{pith2026250417782,
  author       = {Pith},
  title        = {Pith review of: Unleashing the Power of Natural Audio Featuring Multiple Sound Sources},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UOS3IJN5}},
  note         = {Machine review of arXiv:2504.17782}
}
read the original abstract

Universal sound separation aims to extract clean audio tracks corresponding to distinct events from mixed audio, which is critical for artificial auditory perception. However, current methods heavily rely on artificially mixed audio for training, which limits their ability to generalize to naturally mixed audio collected in real-world environments. To overcome this limitation, we propose ClearSep, an innovative framework that employs a data engine to decompose complex naturally mixed audio into multiple independent tracks, thereby allowing effective sound separation in real-world scenarios. We introduce two remix-based evaluation metrics to quantitatively assess separation quality and use these metrics as thresholds to iteratively apply the data engine alongside model training, progressively optimizing separation performance. In addition, we propose a series of training strategies tailored to these separated independent tracks to make the best use of them. Extensive experiments demonstrate that ClearSep achieves state-of-the-art performance across multiple sound separation tasks, highlighting its potential for advancing sound separation in natural audio scenarios. For more examples and detailed results, please visit our demo page at https://clearsep.github.io.

Figures

Figures reproduced from arXiv: 2504.17782 by the authors.

Figure 1
Figure 1. Scale comparison between individual clean tracks and total tracks across seven different audio categories in AudioSet. The number of tracks in an audio clip is determined based on the number of audio categories present in AudioSet (Gemmeke et al., 2017). Clean Tracks refer to single-source audio containing only a single audio event, while Total Tracks represent the number of independent tracks corresponding to indiv… view at source ↗
Figure 2
Figure 2. Illustration of ClearSep and Data Engine Pipeline. ClearSep alternates between data engine and model training to progres￾sively enhance sound separation performance and robustness. During the data engine phase, the model employs mutually exclusive class labels as queries to guide separation, ensuring that the separated tracks are independent. A quality filtering process then evaluates the separation results, and onl… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 13 canonical work pages

  1. [7]

    Kong, Q., Cao, Y ., Iqbal, T., Wang, Y ., Wang, W., and Plumbley, M. D. Panns: Large-scale pretrained audio neural networks for audio pattern recognition.IEEE/ACM Transactions on Audio, Speech, and Language Process- ing, 28:2880–2894, 2020a. Kong, Q., Wang, Y ., Song, X., Cao, Y ., Wang, W., and Plumbley, M. D. Source separation with weakly labelled data:...

    work page 2020

  2. [9]

    D., and Wang, W

    Liu, X., Liu, H., Kong, Q., Mei, X., Zhao, J., Huang, Q., Plumbley, M. D., and Wang, W. Separate what you de- scribe: Language-queried audio source separation. arXiv preprint arXiv:2203.15147,

    arXiv

  3. [12]

    P., Watan- abe, S., and Khudanpur, S

    Maciejewski, M., Sell, G., Garcia-Perera, L. P., Watan- abe, S., and Khudanpur, S. Building corpora for single- channel speech separation across multiple domains.arXiv preprint arXiv:1811.02641,

    arXiv

  4. [14]

    Source separation by steering pretrained music models

    Manilow, E., O’Reilly, P., Seetharaman, P., and Pardo, B. Source separation by steering pretrained music models. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 126–130. IEEE,

    work page 2022

  5. [15]

    Listen to What You Want: Neural Network-based Universal Sound Selector

    12 Unleashing the Power of Natural Audio Featuring Multiple Sound Sources Ochiai, T., Delcroix, M., Koizumi, Y ., Ito, H., Kinoshita, K., and Araki, S. Listen to what you want: Neural network-based universal sound selector. arXiv preprint arXiv:2006.05712,

    work page Pith review arXiv 2006

  6. [19]

    and Chen, J

    Wang, D. and Chen, J. Supervised speech separation based on deep learning: An overview. IEEE/ACM Trans- actions on Audio, Speech, and Language Processing , pp. 1702–1726, Oct 2018a. doi: 10.1109/taslp.2018. 2842159. URL http://dx.doi.org/10.1109/ taslp.2018.2842159. Wang, D. and Chen, J. Supervised speech separation based on deep learning: An overview. IE...

    arXiv 2018

  7. [20]

    Tf-gridnet: Making time-frequency domain models great again for monaural speaker separation

    Wang, Z.-Q., Cornell, S., Choi, S., Lee, Y ., Kim, B.-Y ., and Watanabe, S. Tf-gridnet: Making time-frequency domain models great again for monaural speaker separation. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5. IEEE,

    work page 2023

  8. [21]

    Permutation invariant training of deep models for speaker-independent multi-talker speech separation

    Yu, D., Kolbæk, M., Tan, Z.-H., and Jensen, J. Permutation invariant training of deep models for speaker-independent multi-talker speech separation. In 2017 IEEE Interna- tional Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), pp. 241–245. IEEE,

    work page 2017

Show all 21 references
  1. [2000]

    Non-negative matrix factor deconvolution; extraction of multiple sound sources from monophonic inputs

    Smaragdis, P. Non-negative matrix factor deconvolution; extraction of multiple sound sources from monophonic inputs. In Independent Component Analysis and Blind Signal Separation: Fifth International Conference, ICA 2004, Granada, Spain, September 22-24,

    2004

  2. [2006]

    Le Roux, J., Wisdom, S., Erdogan, H., and Hershey, J. R. Sdr–half-baked or well done? In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 626–630. IEEE,

    2019

  3. [2015]

    Gass: General- izing audio source separation with large-scale data

    Pons, J., Liu, X., Pascual, S., and Serr`a, J. Gass: General- izing audio source separation with large-scale data. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 546–550. IEEE,

    2024

  4. [2016]

    J., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Wang, L., and Chen, W

    Hu, E. J., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Wang, L., and Chen, W. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685,

    arXiv

  5. [2017]

    R., Chen, Z., Le Roux, J., and Watanabe, S

    Hershey, J. R., Chen, Z., Le Roux, J., and Watanabe, S. Deep clustering: Discriminative embeddings for segmentation and separation. In 2016 IEEE international conference on acoustics, speech and signal processing (ICASSP), pp. 31–35. IEEE,

    2016

  6. [2018]

    Cutting music source separation some slakh: A dataset to study the impact of training data quality and quantity

    Manilow, E., Wichern, G., Seetharaman, P., and Le Roux, J. Cutting music source separation some slakh: A dataset to study the impact of training data quality and quantity. In 2019 IEEE Workshop on Applications of Signal Process- ing to Audio and Acoustics (WASPAA), pp. 45–49. IEEE,

    2019

  7. [2019]

    F., Ellis, D

    Gemmeke, J. F., Ellis, D. P., Freedman, D., Jansen, A., Lawrence, W., Moore, R. C., Plakal, M., and Ritter, M. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP) , pp. 776...

    2017

  8. [2020]

    Rtfs-net: Recurrent time- frequency modelling for efficient audio-visual speech separation

    Pegg, S., Li, K., and Hu, X. Rtfs-net: Recurrent time- frequency modelling for efficient audio-visual speech separation. arXiv preprint arXiv:2309.17189,

    arXiv

  9. [2021]

    Kavalerov, I., Wisdom, S., Erdogan, H., Patton, B., Wilson, K., Le Roux, J., and Hershey, J. R. Universal sound separation. In 2019 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), pp. 175–179. IEEE,

    2019

  10. [2022]

    D., and Wang, W

    Liu, X., Kong, Q., Zhao, Y ., Liu, H., Yuan, Y ., Liu, Y ., Xia, R., Wang, Y ., Plumbley, M. D., and Wang, W. Separate anything you describe. arXiv preprint arXiv:2308.05037,

    arXiv

  11. [2023]

    Clapsep: Leveraging contrastive pre-trained models for multi-modal query-conditioned target sound extraction

    Ma, H., Peng, Z., Shao, M., Liu, J., Li, X., and Wu, X. Clapsep: Leveraging contrastive pre-trained models for multi-modal query-conditioned target sound extraction. arXiv preprint arXiv:2402.17455,

    arXiv

  12. [2024]

    Demucs: Deep extractor for music sources with extra unlabeled data remixed

    D´efossez, A., Usunier, N., Bottou, L., and Bach, F. Demucs: Deep extractor for music sources with extra unlabeled data remixed. arXiv preprint arXiv:1909.01174,

    1909 arXiv

  13. [2025]

    Om- nisep: Unified omni-modality sound separation with query-mixup

    Cheng, X., Zheng, S., Wang, Z., Fang, M., Zhang, Z., Huang, R., Ma, Z., Ji, S., Zuo, J., Jin, T., et al. Om- nisep: Unified omni-modality sound separation with query-mixup. arXiv preprint arXiv:2410.21269,

    arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.