Pith. sign in

REVIEW 4 major objections 5 minor 71 references

User-guided Generative Source Separation

T0 review · 4 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 This paper claims that a diffusion model can pull any instrument out of a music mixture when the user hums or plays the target melody, optionally adding a rough sketch on the mel-spectrogram to mark the target and background regions.

desk verdict A solid, honest diffusion-MSS paper whose headline user-guided claim still rests on simulated oracle inputs; worth a serious referee but not a full endorsement yet. read the letter →

arxiv 2507.01339 v1 pith:C4SJCRN3 submitted 2025-07-02 cs.SD cs.AIeess.AS

Yutong Wen , Minje Kim , Paris Smaragdis This is my paper
classification cs.SDcs.AIeess.AS
keywords musicsourceseparationdiffusionmodelsuser-guidedhumming-basedconditioningmel-spectrogrammasksinstrument-agnosticgenerativeaudiocomplexspectrogramgeneration
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

GuideSep claims that a diffusion model can be steered to extract almost any instrument from a music mixture by two user-friendly cues: a waveform of someone humming or playing the target melody, and a rough sketch on the mel-spectrogram indicating where the target and background lie. The model is trained on Slakh2100 mixtures with MoisesDB backgrounds and rendered MIDI mimicry with pitch, timing, octave, and monophonic augmentations. On the Slakh2100 test set, the paper reports that GuideSep with both conditions reaches an overall SDR of 10.46 dB, beating a mask-prediction baseline with the same twin U-Net architecture (8.74 dB), and that listeners rate it much closer to ground truth (82.82 vs 50.69). The paper also shows the model still separates with only one of the two conditions, making the interaction practical. If these results transfer to real user input, source separation becomes a controllable tool for arbitrary instruments rather than a fixed four-stem service.

What carries the argument

The load-bearing object is a conditional complex-spectrogram diffusion model trained with the EDM ('Elucidating Diffusion Models') recipe, which schedules noise levels and uses preconditioned denoising score matching. Two U-Nets are coupled: a score U-Net that generates the target complex spectrogram from noise, and a condition U-Net that processes the mixture's complex STFT, the mimicry's magnitude STFT, and the positive/negative mel-spectrogram masks after a one-hidden-layer projection aligns the mel-frequency axis to the STFT frequency axis; each condition layer is element-wise added into the corresponding score layer. The complex spectrogram is amplitude-compressed by $\tilde{c} = \beta |c|^\alpha e^{i\angle c}$ to fit the network's input range. Sampling uses a multistep DPM solver in eight steps, so the generative process is cheap at inference.

What would settle it

Collect genuine user hums of target melodies from real songs and genuine hand-drawn mel-spectrogram masks, run GuideSep on the corresponding mixtures, and compare SDR and MUSHRA scores with the same-architecture mask-prediction baseline under identical conditions; if the gap disappears or reverses, the paper's central claim fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that conditioning a complex-spectrogram diffusion model on user mimicry and mel-spectrogram masks yields instrument-agnostic separation beyond the four-stem setup. The full GuideSep model scores 10.46 dB SDR on average over ten Slakh2100 instrument groups, compared with 8.74 dB for the same-architecture mask-prediction baseline; ablations give 9.60 dB for mimicry only, 10.09 dB for the positive mask only, and 13.61 dB overall when the mimicry is a real hum from the HumTrans dataset mixed with random MoisesDB backgrounds. The subjective listening test reports 82.82 for GuideSep versus 50.69 for the baseline on separation quality, and 79.38 versus 68.88 on timbre similarity. The model tolerates incomplete conditions, and pseudo-masks formed by blurring the mimicry and mixture spectrograms help seven of ten instrument classes at no extra user cost.

Load-bearing premise

The load-bearing premise is that the simulated user inputs, pitch-bent, time-jittered, monophonic MIDI renderings for mimicry and Gaussian-blurred, randomly dropped ground-truth spectrograms for masks, faithfully represent how people actually hum and sketch; if real inputs are more imperfect, the reported SDR and listening scores may not transfer.

Editorial extensions

If this is right

  • GuideSep lets a user name the target by sound rather than by class label, so the same trained model can extract instruments it was never given a label for, as long as the user can hum or play the part.
  • Because either condition can be dropped, the system degrades gracefully: a user who cannot sketch can hum, and a user who cannot hum can draw a mask, with only a small SDR loss relative to the full-condition setup.
  • Pseudo-masks can be generated automatically from the mimicry and the mixture, so the mask cue can be added even when the user provides no sketch, improving results for seven of ten instrument classes.
  • The diffusion formulation produces perceptually cleaner extractions than the matching mask-prediction baseline, with a much larger MUSHRA gap than SDR gap, indicating that modeling the prior distribution of clean sources matters for perceived quality.

Reading between the lines

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

  • The paper never tests real users humming and sketching on the same real mixtures; a field study with actual user inputs and actual songs is the direct next experiment, and its outcome would determine whether the synthetic-input gains hold.
  • The condition set suggests an interactive refinement loop, hum to start, look at the extraction, then correct with a mask, that the paper does not evaluate; iterative user steering is a plausible product-level use.
  • Because the paper's monophonic mimicry is weakest for piano, guitar, strings, and synth, conditioning on polyphonic hummed or played phrases, or on automatic chord inference, is a concrete way to attack the reported failures on polyphonic targets.
  • The same twin-U-Net conditioning scheme could be lifted onto other generative backbones, or extended to re-synthesize the extracted source with altered timbre, though the paper does not claim either.
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.

Referee Report

4 major / 5 minor

Summary. The paper proposes GuideSep, a diffusion-based music source separation model that conditions on two user-provided signals: a waveform mimicry (e.g., humming) and mel-spectrogram masks. The model extends the EDMSound backbone to accept mixture, mimicry, and mask inputs through a twin U-Net architecture, and is trained on Slakh2100 plus MoisesDB with randomized mixing. The authors compare GuideSep against a mask-prediction baseline built on the same backbone, report SDR gains across ten instrument classes on the Slakh2100 test split, and present a MUSHRA listening test with 13 participants. The central claims are that GuideSep achieves high-quality instrument-agnostic separation and that generative modeling outperforms the predictive baseline under identical conditioning.

Significance. If the claims are supported, the paper makes a useful contribution to interactive music source separation by introducing a flexible conditioning scheme that goes beyond fixed class labels and sound queries. The comparison between a diffusion model and a same-architecture mask-prediction baseline is a valuable and relatively rare systematic study of generative versus predictive approaches in MSS. The authors ship code and a demo page, and the reported SDR and MUSHRA numbers are internally consistent and come with confidence intervals. However, the central "user-guided" claim is currently validated only with simulated user inputs that are derived from ground-truth MIDI and ground-truth spectrograms, and the real-user HumTrans test is not set up to measure the intended real-world difficulty. The stress-test concern about oracle-like user inputs lands: the paper itself acknowledges in Section 4.2 that the strong hum result may be due to random mixing, and in Section 2.1.1 that the monophonic condition is a training-data limitation. The method is promising, but the external validity of the headline numbers is not yet established.

major comments (4)
  1. [Sections 2.1.1, 2.1.2, and Table 1] The main evaluation uses oracle-derived user conditions. The mimicry condition is rendered from the ground-truth MIDI score of the target source with FluidSynth, and the mel-masks are Gaussian-blurred versions of the ground-truth target and residual spectrograms. Because of this, the SDR numbers in Table 1 measure performance under user inputs that are semantically aligned with the target, not under realistic user inputs that may be off-octave, fragmentary, or misleading. This is load-bearing because Table 1 shows that masks are a large part of the full-condition gain (positive-mask-only 10.09 dB and mimicry-only 9.60 dB versus full 10.46 dB), so the optimistic mask simulation directly inflates the headline full-condition result. The authors should either run a real-user study or systematically degrade the simulated conditions in ways that model human errors, and report how SDR degrades.
  2. [Section 4.2] The HumTrans experiment does not validate the user-guided claim in a realistic setting. The target source is synthesized from the MIDI aligned to the humming, so the target and the condition share the same note sequence by construction, and the backgrounds are randomly mixed from MoisesDB with no musical context. The paper reports only a single overall SDR mean of 13.61 dB without a confidence interval, without the mask-prediction baseline, and without comparison to the Slakh test-set results. The authors even note that the strong performance 'could also be attributed to the random mixing used during evaluation.' To support the claim that the mimicry condition generalizes to real hums, the experiment should include user-recorded hums against realistic mixtures, report per-class or per-condition breakdowns, and include the same baseline for comparison.
  3. [Section 3.2 and Table 1] The mask-prediction baseline is not fully matched to the diffusion model. Section 3.2 states that the baseline has 80.3 million parameters while GuideSep has 93.3 million, and the baseline takes a magnitude spectrogram of the mixture as its score-U-Net input whereas GuideSep takes complex-valued noise. The 13-million-parameter difference and the different input representation mean that the SDR gap in Table 1 (10.46 versus 8.74 dB overall) cannot be attributed solely to the predictive-versus-generative modeling choice. The authors should train a mask-prediction baseline with a comparable parameter count and the same conditioning inputs, or otherwise quantify how much of the gap is due to capacity and input representation.
  4. [Section 4] The claim that GuideSep works 'beyond the VBDO setup' is only demonstrated on the Slakh2100 test split, whose targets are all synthesized from MIDI and whose backgrounds are drawn from the same training distribution. There is no evaluation on real-world multitrack recordings such as the MoisesDB test songs or other datasets. Since the method is proposed for real-world use, an evaluation on real recordings with realistic user inputs is necessary to support the general claim; otherwise, the 'instrument-agnostic' and 'versatile' statements remain training-distribution-specific.
minor comments (5)
  1. [Table 1] The final row labeled 'Frequency (%)' is not defined in the caption or in the text; please clarify whether it is the percentage of test examples per instrument class and how it was computed.
  2. [Section 3.1.1] There is a duplicated word in 'following our training procedure procedure described in Section 3.1,' and the sentence beginning 'All The input audio' has an inconsistent capitalization.
  3. [Section 4.1] The MUSHRA test reports 13 participants and 95% confidence intervals, but it does not state whether participants were screened for hearing ability or whether the GuideSep-baseline differences are statistically significant beyond the reported intervals; a paired significance test would strengthen the perceptual claim.
  4. [Section 3.1.1] The paper trains and evaluates on 4.1-second clips at 16 kHz, but does not state how the model would be applied to full-length songs; please clarify whether inference is performed on fixed-length chunks and, if so, how boundaries are handled.
  5. [Section 4] The demo page is mentioned as evidence for 'real-world conditions,' but no systematic evaluation from the demo page is included in the paper; please either describe the demo results in the text or remove the reliance on the page for that claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: GuideSep's reported gains come from a held-out empirical evaluation, not from a quantity fitted or defined in terms of the target result.

full rationale

GuideSep is a conditional diffusion model trained with denoising score matching on random mixtures of Slakh2100 and MoisesDB and evaluated on the Slakh2100 test split. The mimicry and mask conditions are synthesized from ground-truth MIDI and spectrograms, but the model never sees the ground-truth source as a direct condition; the reported SDR is computed on a held-out set against the ground-truth source. The central comparison with the mask-prediction baseline is a genuine head-to-head: the baseline uses the same twin U-Net backbone, training data, and training budget, and the differences in Table 1 are empirical outcomes, not identities. The paper relies on EDMSound [29] as the diffusion backbone, and that citation includes one of the present authors, but this is a normal method citation to an externally available system; the result does not depend on an unverified uniqueness theorem or on an ansatz imported solely from the authors' prior work. The Humming-only evaluation in Section 4.2 is acknowledged to be non-comparable because targets are MIDI-synthesized and backgrounds are randomly mixed, which is a limitation on generalization to real user inputs, not circularity. No equation in the paper reduces the predicted source to the conditioning input by construction, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

All numerical choices in this paper are training or simulation hyperparameters, not physical constants. The central result rests on the assumption that simulated user inputs (FluidSynth mimicry and blurred ground-truth masks) capture real user behavior, which is not independently validated. No new particles, forces, or conceptual entities are introduced.

free parameters (5)
  • Condition dropout rates = mimicry 0.3, mel-mask 0.7, mixture 0.1
    Tuned on the validation split in Section 3.1.2; the high mel-mask dropout intentionally shifts the model toward mimicry learning and affects the full-condition results.
  • Mimicry augmentation probabilities = pitch-bend 0.5, timing shift 0.4, octave shift 0.5
    Hand-specified in Section 2.1.1 to simulate off-pitch notes, imperfect timing, and note-range limits; no real-user study validates these distributions.
  • Synthetic mel-mask Gaussian blur sigma = 4 to 6; pseudo-mask sigma 5
    Chosen by hand in Sections 2.1.2 and 4.2 to approximate user sketching imperfections; the paper notes that a different sigma could change bass results.
  • Random mixing SNR range = -5 dB to +5 dB
    Hand-specified in Section 3.1.1 for training mixes; this choice controls task difficulty and may contribute to the high humming-only score.
  • Amplitude compression alpha and scaling beta = not reported
    Taken from prior work [42,51] for complex-spectrogram scaling in Section 2.2; values are not stated in this paper but the transform is load-bearing for the diffusion input.
assumptions (6)
  • standard math The EDM framework and denoising score matching are valid for complex-spectrogram diffusion.
    Section 2.2 invokes the Karras et al. EDM formulation and score matching without re-derivation.
  • domain assumption FluidSynth MIDI rendering plus the listed augmentations approximates real user mimicry.
    Section 2.1.1 uses this to synthesize all training mimicry; real humming is tested only once and against random background mixes.
  • domain assumption Gaussian-blurred ground-truth masks with 40% patch dropout approximate real user sketches.
    Section 2.1.2 generates synthetic masks this way; the user interface is described but not evaluated with real users.
  • domain assumption Random mixes of Slakh targets with MoisesDB backgrounds transfer to natural music.
    Section 3.1.1 relies on random mixing and cites [62] for its benefits, while conceding that musical context may be lost.
  • domain assumption A monophonic mimicry condition is sufficient for extracting polyphonic targets when masks are available.
    Section 2.1.1 restricts mimicry to monophonic lines; Section 4 acknowledges that piano, guitar, strings, and synth remain hard.
  • domain assumption Excluding drums is an acceptable interpretation of instrument-agnostic separation.
    Section 4 excludes drum tracks because the MIDI-based synthesis method does not apply, narrowing the scope of the claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of User-guided Generative Source Separation." pith.science (2026). https://pith.science/paper/C4SJCRN3

@misc{pith2026250701339,
  author       = {Pith},
  title        = {Pith review of: User-guided Generative Source Separation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C4SJCRN3}},
  note         = {Machine review of arXiv:2507.01339}
}
read the original abstract

Music source separation (MSS) aims to extract individual instrument sources from their mixture. While most existing methods focus on the widely adopted four-stem separation setup (vocals, bass, drums, and other instruments), this approach lacks the flexibility needed for real-world applications. To address this, we propose GuideSep, a diffusion-based MSS model capable of instrument-agnostic separation beyond the four-stem setup. GuideSep is conditioned on multiple inputs: a waveform mimicry condition, which can be easily provided by humming or playing the target melody, and mel-spectrogram domain masks, which offer additional guidance for separation. Unlike prior approaches that relied on fixed class labels or sound queries, our conditioning scheme, coupled with the generative approach, provides greater flexibility and applicability. Additionally, we design a mask-prediction baseline using the same model architecture to systematically compare predictive and generative approaches. Our objective and subjective evaluations demonstrate that GuideSep achieves high-quality separation while enabling more versatile instrument extraction, highlighting the potential of user participation in the diffusion-based generative process for MSS. Our code and demo page are available at https://yutongwen.github.io/GuideSep/

Figures

Figures reproduced from arXiv: 2507.01339 by the authors.

Figure 3
Figure 3. In the first part, where participants rated the sep [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

71 extracted references · 59 canonical work pages

  1. [1]

    User-Guided Gen- erative Source Separation

    INTRODUCTION Music source separation (MSS) aims to separate a mix- ture audio into its constituent sources, typically defined by the instrument. Since the 2015 Signal Separation Eval- uation Campaign (SiSEC) [1], the MSS community has largely focused on supervised models to separate songs into four stems: vocals, bass, drums, and others that in- cludes al...

    work page 2015

  2. [2]

    qNpYpA3PbTUP4hgRSvXYnbxuOM4=

    THE PROPOSED GUIDESEP METHOD GuideSep is a diffusion model conditioned by user input. Our approach leverages users’ input describing a source, i.e., the raw waveform of user mimicry to a target source as well as a rough mask in the mel-spectrogram domain. 1 Some literature refers to predictive models as discriminative or de- terministic. Lemercier et al. ...

    work page

  3. [3]

    We propose GuideSep, one of the first diffusion- based models designed to address music source sep- aration beyond the VBDO setup and we release the codebase 2) We introduce versatile, instrument-agnostic conditions—waveform mimicry conditions and mel- spectrogram domain masks—that are more practical for real-world applications 3) We design a mask-predict...

    work page

  4. [4]

    synthesized

    EV ALUA TION AND DISCUSSION We evaluate our model on the official test split of the Slakh2100 dataset. The mimicry condition signals are synthesized as described in Section 2.1, using randomly selected virtual instruments from the FluidSynth library. Similarly, the positive and negative masks are simulated following the same procedure outlined in Section ...

    work page

  5. [5]

    The Slakh2100 dataset provides an official train-validation-test split, which we utilize as well

    EXPERIMENT We conduct experiments using the Slakh2100 dataset [58] augmented by MoisesDB [59] for training. The Slakh2100 dataset provides an official train-validation-test split, which we utilize as well. We evaluate our model’s performance using the widely adopted signal-to-distortion ratio (SDR) metrics [60, 61]. 3.1 Training and model details 3.1.1 Th...

    work page

  6. [6]

    D3Net: Densely con- nected multidilated densenet for music source separa- tion,

    N. Takahashi and Y . Mitsufuji, “D3Net: Densely con- nected multidilated densenet for music source separa- tion,” arXiv preprint arXiv:2010.01733, 2020

    arXiv 2010

  7. [7]

    Our re- sults demonstrate that this approach achieves high-quality separation while offering greater adaptability compared to traditional class-based methods

    CONCLUSION We introduced GuideSep, a diffusion-based music source separation model that enables flexible, instrument-agnostic separation using waveform mimicry conditions and mel- spectrogram masks, and released the codebase. Our re- sults demonstrate that this approach achieves high-quality separation while offering greater adaptability compared to tradi...

    work page

  8. [8]

    The 2015 signal separation evaluation cam- paign,

    N. Ono, Z. Rafii, D. Kitamura, N. Ito, and A. Li- utkus, “The 2015 signal separation evaluation cam- paign,” in Latent Variable Analysis and Signal Sep- aration, E. Vincent, A. Yeredor, Z. Koldovský, and P. Tichavský, Eds. Cham: Springer International Pub- lishing, 2015, pp. 387–395

    work page 2015

Show all 71 references
  1. [9]

    Hybrid trans- formers for music source separation,

    S. Rouard, F. Massa, and A. Défossez, “Hybrid trans- formers for music source separation,” inICASSP 2023- 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  2. [10]

    However, class labels can be vague, as instruments like the guitar may exhibit significant variability within the same label

    extends this approach to 13 instruments. However, class labels can be vague, as instruments like the guitar may exhibit significant variability within the same label. Moreover, new instrument classes require re-training. An- other approach, query-based MSS conditions the model...

    2025 arXiv

  3. [11]

    Music source sepa- ration based on a lightweight deep learning framework (dttnet: Dual-path tfc-tdf unet),

    J. Chen, S. Vekkot, and P. Shukla, “Music source sepa- ration based on a lightweight deep learning framework (dttnet: Dual-path tfc-tdf unet),” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 656–660

    2024

  4. [12]

    The whole is greater than the sum of its parts: improving music source separation by bridging networks,

    R. Sawata, N. Takahashi, S. Uhlich, S. Takahashi, and Y . Mitsufuji, “The whole is greater than the sum of its parts: improving music source separation by bridging networks,” EURASIP Journal on Audio, Speech, and Music Processing, vol. 2024, no. 1, p. 39, 2024

    2024

  5. [13]

    SCNet: Sparse compression net- work for music source separation,

    W. Tong, J. Zhu, J. Chen, S. Kang, T. Jiang, Y . Li, Z. Wu, and H. Meng, “SCNet: Sparse compression net- work for music source separation,” in ICASSP 2024- 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 1276–1280

    2024

  6. [14]

    Music source separation with band- split rnn,

    Y . Luo and J. Yu, “Music source separation with band- split rnn,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 31, pp. 1893–1901, 2023

    1901

  7. [15]

    Mu- sic source separation with band-split rope transformer,

    W.-T. Lu, J.-C. Wang, Q. Kong, and Y .-N. Hung, “Mu- sic source separation with band-split rope transformer,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 481–485

    2024

  8. [16]

    Conditioned- U-Net: Introducing a control mechanism in the u- net for multiple source separations,

    G. Meseguer-Brocal and G. Peeters, “Conditioned- U-Net: Introducing a control mechanism in the u- net for multiple source separations,” arXiv preprint arXiv:1907.01277, 2019

    1907 arXiv

  9. [17]

    End- to-end sound source separation conditioned on instru- ment labels,

    O. Slizovskaia, L. Kim, G. Haro, and E. Gomez, “End- to-end sound source separation conditioned on instru- ment labels,” in ICASSP 2019-2019 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019, pp. 306–310

    2019

  10. [18]

    Class-conditional embeddings for music source separation,

    P. Seetharaman, G. Wichern, S. Venkataramani, and J. Le Roux, “Class-conditional embeddings for music source separation,” in ICASSP 2019-2019 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019, pp. 301–305

    2019

  11. [19]

    Meta- learning extractors for music source separation,

    D. Samuel, A. Ganeshan, and J. Naradowsky, “Meta- learning extractors for music source separation,” in ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2020, pp. 816–820

    2020

  12. [20]

    User guided audio selection from com- plex sound mixtures,

    P. Smaragdis, “User guided audio selection from com- plex sound mixtures,” in Proceedings of the 22nd an- nual ACM symposium on User interface software and technology, 2009, pp. 89–92

    2009

  13. [21]

    Audio query-based music source separation,

    J. H. Lee, H.-S. Choi, and K. Lee, “Audio query-based music source separation,” arXiv preprint arXiv:1908.06593, 2019

    1908 arXiv

  14. [22]

    Hierarchi- cal musical instrument separation

    E. Manilow, G. Wichern, and J. Le Roux, “Hierarchi- cal musical instrument separation.” inISMIR, 2020, pp. 376–383

    2020

  15. [23]

    A stem-agnostic single-decoder system for music source separation be- yond four stems,

    K. N. Watcharasupat and A. Lerch, “A stem-agnostic single-decoder system for music source separation be- yond four stems,” arXiv preprint arXiv:2406.18747 , 2024

    2024 arXiv

  16. [24]

    Zero-shot audio source separation through query-based learning from weakly-labeled data,

    K. Chen, X. Du, B. Zhu, Z. Ma, T. Berg-Kirkpatrick, and S. Dubnov, “Zero-shot audio source separation through query-based learning from weakly-labeled data,” in Proceedings of the AAAI Conference on Arti- ficial Intelligence, vol. 36, no. 4, 2022, pp. 4441–4449

    2022

  17. [25]

    Few-shot musical source separation,

    Y . Wang, D. Stoller, R. M. Bittner, and J. P. Bello, “Few-shot musical source separation,” in ICASSP 2022-2022 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 121–125

    2022

  18. [26]

    Structured dropout for weak label and multi-instance learning and its appli- cation to score-informed source separation,

    S. Ewert and M. B. Sandler, “Structured dropout for weak label and multi-instance learning and its appli- cation to score-informed source separation,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2017, pp. 2277–2281

    2017

  19. [27]

    Monaural score- informed source separation for classical music using convolutional neural networks

    M. Miron, J. Janer, and E. Gómez, “Monaural score- informed source separation for classical music using convolutional neural networks.” in ISMIR, vol. 2017, 2017, pp. 55–62

    2017

  20. [28]

    Score-informed source separation of choral music,

    M. Gover, “Score-informed source separation of choral music,” 2020

    2020

  21. [29]

    Online/offline score informed music signal decompo- sition: application to minus one,

    A. J. Munoz-Montoro, J. J. Carabias-Orti, P. Vera- Candeas, F. J. Canadas-Quesada, and N. Ruiz-Reyes, “Online/offline score informed music signal decompo- sition: application to minus one,” EURASIP Journal on Audio, Speech, and Music Processing , vol. 2019, pp. 1–30, 2019

    2019

  22. [30]

    Transcrip- tion is all you need: Learning to separate musical mix- tures with score as supervision,

    Y .-N. Hung, G. Wichern, and J. Le Roux, “Transcrip- tion is all you need: Learning to separate musical mix- tures with score as supervision,” inICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2021, pp. 46–50

    2021

  23. [31]

    ISSE: An interactive source separation editor,

    N. J. Bryan, G. J. Mysore, and G. Wang, “ISSE: An interactive source separation editor,” inProceedings of the SIGCHI Conference on Human Factors in Comput- ing Systems, 2014, pp. 257–266

    2014

  24. [32]

    Interactive user- feedback for sound source separation,

    N. J. Bryan and G. J. Mysore, “Interactive user- feedback for sound source separation,” inInternational Conference on Intelligent User-Interfaces (IUI), Work- shop on Interactive Machine Learning. Santa Monica, 2013

    2013

  25. [33]

    An efficient posterior reg- ularized latent variable model for interactive sound source separation,

    N. Bryan and G. Mysore, “An efficient posterior reg- ularized latent variable model for interactive sound source separation,” in International conference on ma- chine learning. PMLR, 2013, pp. 208–216

    2013

  26. [34]

    Interactive refinement of supervised and semi-supervised sound source sep- aration estimates,

    N. J. Bryan and G. J. Mysore, “Interactive refinement of supervised and semi-supervised sound source sep- aration estimates,” in 2013 IEEE International Con- ference on Acoustics, Speech and Signal Processing . IEEE, 2013, pp. 883–887

    2013

  27. [35]

    Separation by “hum- ming

    P. Smaragdis and G. J. Mysore, “Separation by “hum- ming”: User-guided sound extraction from mono- phonic mixtures,” in 2009 IEEE Workshop on Appli- cations of Signal Processing to Audio and Acoustics . IEEE, 2009, pp. 69–72

    2009

  28. [36]

    EDMSound: Spectrogram based diffusion models for efficient and high-quality audio synthesis,

    G. Zhu, Y . Wen, M.-A. Carbonneau, and Z. Duan, “EDMSound: Spectrogram based diffusion models for efficient and high-quality audio synthesis,” arXiv preprint arXiv:2311.08667, 2023

    2023 arXiv

  29. [37]

    Diffusion models for audio restoration: A review [special issue on model- based and data-driven audio signal processing],

    J.-M. Lemercier, J. Richter, S. Welker, E. Moliner, V . Välimäki, and T. Gerkmann, “Diffusion models for audio restoration: A review [special issue on model- based and data-driven audio signal processing],” IEEE Signal Processing Magazine, vol. 41, no. 6, pp. 72–84, 2025

    2025

  30. [38]

    Conv-TasNet: Surpassing ideal time–frequency magnitude masking for speech separation,

    Y . Luo and N. Mesgarani, “Conv-TasNet: Surpassing ideal time–frequency magnitude masking for speech separation,” IEEE/ACM transactions on audio, speech, and language processing , vol. 27, no. 8, pp. 1256– 1266, 2019

    2019

  31. [39]

    DCCRN: Deep complex con- volution recurrent network for phase-aware speech en- hancement,

    Y . Hu, Y . Liu, S. Lv, M. Xing, S. Zhang, Y . Fu, J. Wu, B. Zhang, and L. Xie, “DCCRN: Deep complex con- volution recurrent network for phase-aware speech en- hancement,” arXiv preprint arXiv:2008.00264, 2020

    2008 arXiv

  32. [40]

    MANNER: Multi-view attention network for noise erasure,

    H. J. Park, B. H. Kang, W. Shin, J. S. Kim, and S. W. Han, “MANNER: Multi-view attention network for noise erasure,” in ICASSP 2022-2022 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 7842–7846

    2022

  33. [41]

    Evaluation metrics for generative speech enhancement methods: Issues and perspectives,

    J. Pirklbauer, M. Sach, K. Fluyt, W. Tirry, W. Wardah, S. Moeller, and T. Fingscheidt, “Evaluation metrics for generative speech enhancement methods: Issues and perspectives,” in Speech Communication; 15th ITG Conference. VDE, 2023, pp. 265–269

    2023

  34. [42]

    Latent autoregressive source separation,

    E. Postolache, G. Mariani, M. Mancusi, A. San- tilli, L. Cosmo, and E. Rodolà, “Latent autoregressive source separation,” inProceedings of the AAAI Confer- ence on Artificial Intelligence, vol. 37, no. 8, 2023, pp. 9444–9452

    2023

  35. [43]

    Generative adver- sarial source separation,

    Y . C. Subakan and P. Smaragdis, “Generative adver- sarial source separation,” in 2018 IEEE International Conference on Acoustics, Speech and Signal Process- ing (ICASSP). IEEE, 2018, pp. 26–30

    2018

  36. [44]

    SEPDIFF: Speech separation based on denoising diffusion model,

    B. Chen, C. Wu, and W. Zhao, “SEPDIFF: Speech separation based on denoising diffusion model,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2023, pp. 1–5

    2023

  37. [45]

    Separate and dif- fuse: Using a pretrained diffusion model for improving source separation,

    S. Lutati, E. Nachmani, and L. Wolf, “Separate and dif- fuse: Using a pretrained diffusion model for improving source separation,” arXiv preprint arXiv:2301.10752 , 2023

    2023 arXiv

  38. [46]

    Diffusion-based generative speech source separation,

    R. Scheibler, Y . Ji, S.-W. Chung, J. Byun, S. Choe, and M.-S. Choi, “Diffusion-based generative speech source separation,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Process- ing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  39. [47]

    Universal score-based speech enhancement with high content preservation,

    R. Scheibler, Y . Fujita, Y . Shirahata, and T. Ko- matsu, “Universal score-based speech enhancement with high content preservation,” arXiv preprint arXiv:2406.12194, 2024

    2024 arXiv

  40. [48]

    A variance-preserving interpolation approach for diffusion models with applications to single chan- nel speech enhancement and recognition,

    Z. Guo, Q. Wang, J. Du, J. Pan, Q.-F. Liu, and C.- H. Lee, “A variance-preserving interpolation approach for diffusion models with applications to single chan- nel speech enhancement and recognition,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, 2024

    2024

  41. [49]

    Speech enhancement and dere- verberation with diffusion-based generative models,

    J. Richter, S. Welker, J.-M. Lemercier, B. Lay, and T. Gerkmann, “Speech enhancement and dere- verberation with diffusion-based generative models,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 31, pp. 2351–2364, 2023

    2023

  42. [50]

    Music source separation with generative flow,

    G. Zhu, J. Darefsky, F. Jiang, A. Selitskiy, and Z. Duan, “Music source separation with generative flow,” IEEE Signal Processing Letters , vol. 29, pp. 2288–2292, 2022

    2022

  43. [51]

    Multi-source diffusion models for simultaneous music generation and sepa- ration,

    G. Mariani, I. Tallini, E. Postolache, M. Mancusi, L. Cosmo, and E. Rodolà, “Multi-source diffusion models for simultaneous music generation and sepa- ration,” arXiv preprint arXiv:2302.02257, 2023

    2023 arXiv

  44. [52]

    Si- multaneous music separation and generation using multi-track latent diffusion models,

    T. Karchkhadze, M. R. Izadi, and S. Dubnov, “Si- multaneous music separation and generation using multi-track latent diffusion models,” arXiv preprint arXiv:2409.12346, 2024

    2024 arXiv

  45. [53]

    Fluidsynth real-time and thread safety challenges,

    D. Henningsson and F. Team, “Fluidsynth real-time and thread safety challenges,” in Proceedings of the 9th International Linux Audio Conference, Maynooth University, Ireland, 2011, pp. 123–128

    2011

  46. [54]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020

    2020

  47. [55]

    Deep unsupervised learning us- ing nonequilibrium thermodynamics,

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning us- ing nonequilibrium thermodynamics,” in International conference on machine learning . PMLR, 2015, pp. 2256–2265

    2015

  48. [56]

    Score-based generative mod- eling through stochastic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative mod- eling through stochastic differential equations,” arXiv preprint arXiv:2011.13456, 2020

    2011 arXiv

  49. [57]

    Eluci- dating the design space of diffusion-based generative models,

    T. Karras, M. Aittala, T. Aila, and S. Laine, “Eluci- dating the design space of diffusion-based generative models,” Advances in neural information processing systems, vol. 35, pp. 26 565–26 577, 2022

    2022

  50. [58]

    Analysis of the decision- directed snr estimator for speech enhancement with re- spect to low-snr and transient conditions,

    C. Breithaupt and R. Martin, “Analysis of the decision- directed snr estimator for speech enhancement with re- spect to low-snr and transient conditions,” IEEE trans- actions on audio, speech, and language processing , vol. 19, no. 2, pp. 277–289, 2010

    2010

  51. [59]

    Universal speech enhancement with score-based dif- fusion,

    J. Serrà, S. Pascual, J. Pons, R. O. Araz, and D. Scaini, “Universal speech enhancement with score-based dif- fusion,” arXiv preprint arXiv:2206.03065, 2022

    2022 arXiv

  52. [60]

    Music ControlNet: Multiple time-varying controls for music generation,

    S.-L. Wu, C. Donahue, S. Watanabe, and N. J. Bryan, “Music ControlNet: Multiple time-varying controls for music generation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 2692– 2703, 2024

    2024

  53. [61]

    Sketch2Sound: Controllable audio generation via time-varying signals and sonic imita- tions,

    H. F. García, O. Nieto, J. Salamon, B. Pardo, and P. Seetharaman, “Sketch2Sound: Controllable audio generation via time-varying signals and sonic imita- tions,” arXiv preprint arXiv:2412.08550, 2024

    2024 arXiv

  54. [62]

    Photorealistic text-to-image diffusion models with deep language un- derstanding,

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gontijo Lopes, B. Karagol Ayan, T. Salimans et al. , “Photorealistic text-to-image diffusion models with deep language un- derstanding,” Advances in neural information process- ing systems, vol. ...

    2022

  55. [63]

    DPM-Solver: A fast ode solver for diffusion proba- bilistic model sampling in around 10 steps,

    C. Lu, Y . Zhou, F. Bao, J. Chen, C. Li, and J. Zhu, “DPM-Solver: A fast ode solver for diffusion proba- bilistic model sampling in around 10 steps,” Advances in Neural Information Processing Systems, vol. 35, pp. 5775–5787, 2022

    2022

  56. [64]

    DPM-Solver++: Fast solver for guided sam- pling of diffusion probabilistic models,

    ——, “DPM-Solver++: Fast solver for guided sam- pling of diffusion probabilistic models,”arXiv preprint arXiv:2211.01095, 2022

    2022 arXiv

  57. [65]

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

    E. Manilow, G. Wichern, P. Seetharaman, and J. Le Roux, “Cutting music source separation some Slakh: A dataset to study the impact of training data quality and quantity,” in Proc. IEEE Workshop on Ap- plications of Signal Processing to Audio and Acoustics (WASPAA). IEEE, 2019

    2019

  58. [66]

    MoisesDB: A dataset for source separation beyond 4- stems,

    I. Pereira, F. Araújo, F. Korzeniowski, and R. V ogl, “MoisesDB: A dataset for source separation beyond 4- stems,” arXiv preprint arXiv:2307.15913, 2023

    2023 arXiv

  59. [67]

    Perfor- mance measurement in blind audio source separation,

    E. Vincent, R. Gribonval, and C. Févotte, “Perfor- mance measurement in blind audio source separation,” IEEE transactions on audio, speech, and language pro- cessing, vol. 14, no. 4, pp. 1462–1469, 2006

    2006

  60. [68]

    SDR—medium rare with fast computa- tions,

    R. Scheibler, “SDR—medium rare with fast computa- tions,” in ICASSP 2022-2022 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 701–705

    2022

  61. [69]

    Why does music source separation benefit from ca- cophony?

    C.-B. Jeon, G. Wichern, F. G. Germain, and J. Le Roux, “Why does music source separation benefit from ca- cophony?” in 2024 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). IEEE, 2024, pp. 873–877

    2024

  62. [70]

    web- MUSHRA—a comprehensive framework for web- based listening tests,

    M. Schoeffler, S. Bartoschek, F.-R. Stöter, M. Roess, S. Westphal, B. Edler, and J. Herre, “web- MUSHRA—a comprehensive framework for web- based listening tests,” Journal of Open Research Soft- ware, vol. 6, no. 1, 2018

    2018

  63. [71]

    HumTrans: A novel open-source dataset for humming melody transcrip- tion and beyond,

    S. Liu, X. Li, D. Li, and Y . Shan, “HumTrans: A novel open-source dataset for humming melody transcrip- tion and beyond,” in ICASSP 2024-2024 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 7915–7919

    2024

Pith tools

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