Pith. sign in

REVIEW 4 major objections 6 minor 3 cited by

Source Separation by Flow Matching

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Flow matching with built-in mixture consistency separates overlapping speech better than prior generative and regression methods.

desk verdict FLOSS shows a clean projection-based flow matching construction for separation and strong LibriVox numbers, but the loss modifications break the posterior-sampling theory and the one-step result smells like regression. read the letter →

arxiv 2505.16119 v2 pith:E2TXEIX7 submitted 2025-05-22 cs.SD eess.AS

Robin Scheibler , John R. Hershey , Arnaud Doucet , Henry Li This is my paper · ORCID
classification cs.SDeess.AS MSC 68T0794A12
keywords sourceseparationflowmatchingpermutationequivariancemixtureconsistencyspeechgenerativeaudiosingle-channelflow-basedmodels
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 overlapping speech can be separated by a generative flow-matching model that enforces mixture consistency by construction, rather than by a network that directly predicts sources. The proposed method, FLOSS, learns an ordinary differential equation that transports the mixture, augmented with noise only in directions orthogonal to the constant-sum subspace, to a sample from the distribution of the K sources given that mixture. Because any permutation of the sources yields the same mixture, the authors pair a permutation-equivariant network with a permutation-invariant training loss, and their ablations show a decibel-valued objective is the largest single contributor to the gains. On the paper's audiobook evaluation set, FLOSS reaches 19.13 dB scale-invariant signal-to-distortion ratio (SI-SDR) with five sampling steps and 19.12 dB with one step, outperforming the diffusion baselines Diffsep and EDSep and the regression baselines Conv-TasNet and MB-Locoformer. The practical payoff is that high-quality generative separation can work with very few sampling steps.

What carries the argument

The load-bearing object is the orthogonal decomposition $S = \bar{S} + P^\perp S$, with projection matrices $P = \mathbf{1}\mathbf{1}^\top/K$ and $P^\perp = I_K - P$: the mixture occupies the one-dimensional subspace spanned by the all-ones vector, and everything that distinguishes the sources lies in the zero-sum subspace. Flow matching is confined to that subspace by sampling $Z \sim \mathcal{N}(0, \Sigma_Z)$ and writing the drift as $P^\perp \tilde{v}_\theta(t, P^\perp x_t, \bar{s})$, so the equality $P x_t = \bar{S}$ holds along the whole trajectory. The permutation symmetry is carried by two components: a network that is equivariant by design because attention over sources is computed without positional encoding, and the permutation equivariant training (PET) loss, eq. (17), which selects the best source assignment at $t=0$ and uses it for all times. Two additional design choices, the decibel-valued loss of eq. (19) and mixture-shaped noise from the envelope of $\bar{s}$, drive the main performance gains in the ablations.

What would settle it

Build a two-speaker mixture from the same sentence spoken with the same prosody, so the two true sources are acoustically interchangeable. If FLOSS reliably returns two tracks that sum to the mixture but do not match either true utterance individually (for instance by splitting energy symmetrically or swapping randomly), then the permutation-invariant loss has not learned the true conditional distribution $p(S \mid \bar{s})$; a method that had learned it should return the two actual utterances up to permutation.

Watch

Extended reading notes

Core claim

FLOSS frames separation as sampling from $p(S \mid \bar{s})$, where $S$ is the $K \times L$ matrix of sources and $\bar{s}$ is the average of the mixture repeated $K$ times. The flow is initialized at $x_0 = \bar{S} + P^\perp Z$, so the mixture part is fixed and noise lives only in the zero-sum subspace $P^\perp$; the drift is parameterized as $v_\theta(t, x_t, \bar{s}) = P^\perp \tilde{v}_\theta(t, P^\perp x_t, \bar{s})$, which forces $P x_t = \bar{S}$ at every time and makes mixture consistency an identity. To deal with the $K!$ permutation symmetry of the sources, the authors replace the exact equivariant optimal-transport matching of [31], which they report does not work well at this dimensionality, with a permutation-invariant training loss that fixes the permutation at $t=0$, and they train with a normalized decibel-valued loss, eq. (19), that the ablations identify as the decisive performance factor. The resulting model separates two overlapping speakers at 19.13 dB SI-SDR with five sampling steps and 19.12 dB with one step, and it beats all four baselines on every reported metric.

Load-bearing premise

The premise that carries the method is that the permutation-invariant training loss, together with the normalized and decibel-valued objectives, teaches the network to produce actual samples from the true conditional distribution of the sources given the mixture; the paper gives empirical support but no proof, and if this premise fails, mixture consistency only guarantees that the outputs sum to the observed mixture, not that they are the true sources.

Editorial extensions

If this is right

  • FLOSS reaches 19.12 dB SI-SDR with a single sampling step, nearly matching the 19.13 dB of five steps, so the method is usable in roughly the cost regime of a single regression pass while remaining generative.
  • Because $P x_t = \bar{S}$ is enforced by the drift parameterization, generated sources always sum to the input mixture; this removes the consistency drift that the authors identify in Diffsep.
  • The ablation order shows the decibel-valued loss (eq. 19) and the PIT-based permutation selection (eq. 17) are the main drivers of the gains, so both are transferable design choices for other generative inverse problems with permutation symmetry.
  • The band-split mel architecture operates on 24 kHz audio, meaning the approach is aimed at high-quality restoration and not only at intelligibility-oriented separation.

Reading between the lines

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

  • If the PET loss truly samples $p(S \mid \bar{s})$, the same subspace-constrained flow should extend to noisy or distorted mixtures, where the forward model is the mixture plus noise rather than an exact sum; the paper lists this as future work.
  • The near-identical one-step and five-step results suggest the learned ODE is nearly straight on the relevant subspace, so consistency distillation or simpler samplers may push FLOSS even closer to a one-pass separator.
  • The 'add noise only in the null space of the forward operator and constrain the drift to that null space' recipe is a general strategy for any linear inverse problem with a known null space, such as inpainting or super-resolution, with the permutation-equivariant network becoming optional when the unknown has no symmetry.
  • Testing on same-speaker or same-phrase mixtures would probe whether the permutation-invariant loss resolves content-level ambiguity or relies on acoustic differences between speakers; this is a stress test implied by the symmetry assumption rather than something the paper reports.
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 / 6 minor

Summary. The paper proposes FLOSS, a flow-matching method for single-channel speech source separation. The method represents K sources as rows of a matrix S and the mixture average as ̄S; it initializes the flow at x0 = ̄S + P⊥Z, so that noise is confined to the subspace orthogonal to the all-ones direction, and it parameterizes the learned velocity as P⊥s̃. A permutation-equivariant network and a permutation-invariant training (PET) loss are introduced to handle the permutation ambiguity of the sources. Experiments on LibriVox mixtures report SI-SDR 19.13 with a five-step custom Euler schedule, outperforming Diffsep, EDSep, Conv-TasNet, and MB-Locoformer. The paper also claims that FLOSS is a generative method that samples from the conditional distribution p(S|̄s).

Significance. If the central claims were fully supported, this would be a strong contribution: the subspace projection construction gives mixture consistency by construction, the equivariant architecture is a sensible and non-trivial adaptation of flow matching, and the reported separation quality is substantially better than prior generative baselines. The ablations in Table 1 are also informative. However, the theoretical claim that the trained ODE samples from p(S|̄s) is not established for the loss actually used, and the empirical support is based on a single evaluation set with no error bars. The contribution is therefore significant but currently conditional on additional analysis and validation.

major comments (4)
  1. [§3.2, Eqs. (16)–(19)] The paper's central claim is that FLOSS samples from the conditional distribution p(S|̄s), stated in §3.1. However, the training objective actually minimized no longer has the standard flow-matching minimizer. In Eq. (17), the permutation π^{PIT}_{x0,x1} is itself selected by minimizing Eq. (16) at t=0 using the current network, so the loss is a self-referential function of θ rather than a regression to a fixed target; Eq. (19) then applies a nonlinear dB transformation to the normalized MSE. Neither modification has a known minimizer equal to the conditional expectation in Eq. (13), so there is no argument that the ODE marginal at t=1 equals p(S|̄s). Table 1, rows 3–4, shows these choices are not innocuous reweightings (18.43 vs. 16.73 SI-SDR), so the gap between the theory and the implemented loss is material. I request either a proof or careful argument for distributional fidelity of the PET+dB objective, or a reframing of the contribution as a constrained flow-based separator with heuristic losses, together with a statistical check of the implied posterior.
  2. [§4.5, Table 2] The near-identical one-step and five-step results (SI-SDR 19.12 vs. 19.13) and the lower 25-step score (18.43) are consistent with the model behaving as a one-step conditional-mean estimator rather than as a generative ODE sampler. Because Eq. (17) includes a Dirac mass at t=0, the network may learn a direct posterior-mean map at initialization, making the multi-step integration incidental. This does not directly invalidate the separation metrics, but it undercuts the generative-sampling interpretation. Please report the full NFE/SI-SDR curve, examine whether the learned velocity field actually varies with t in a way that transports the noise distribution, and discuss explicitly whether one-step inference is a feature of the method or a symptom of regression behavior.
  3. [§4.2, §4.5] The empirical support rests on a single evaluation set of 1000 LibriVox mixtures, one training run per configuration, and no confidence intervals or significance tests. The baselines also differ in architecture and parameter count (Conv-TasNet 5M, MB-Locoformer 39M, NCSN++ 47M, FLOSS 36M), so the reported gaps could be affected by capacity or tuning rather than by the proposed mechanism. I recommend reporting at least three seeds for the Table 2 comparison, error bars for all metrics, and an additional corpus such as WSJ0-2mix to establish that the gains are not LibriVox-specific.
  4. [§3.4, §4.4] Several components that influence the final result are heuristic and their sensitivity is not quantified: the active-power threshold and envelope shaping T_env(̄s) in §3.4, the constants C and ε in the SNR-uniform schedule, and the custom five-step Euler schedule {0.95, 4·10^-2, ...} used for the best Table 2 result. Table 1 covers some loss and noise choices, but not the numerical constants or the evaluation schedule. Please add a sensitivity analysis or state clearly which choices were fixed and why.
minor comments (6)
  1. [§2.2, §4.1] The method name is spelled both “Sepdiff” and “SEPDIFF”; please standardize.
  2. [Fig. 1 caption] “Uni-dimensional subspace” should read “one-dimensional subspace.”
  3. [§4.4] The custom five-step schedule should specify whether the listed values are step sizes or ODE evaluation times; as written, the Euler step sequence is ambiguous.
  4. [§4.2] “Active power is normalized to a random value in −29 dB to −19 dB” should define the reference level (e.g., dB FS).
  5. [Eq. (17), Table 1] The notation δ_t for a Dirac delta in a weight function should be distinguished from any Kronecker delta used elsewhere, and the weight “1” in Table 1 row 5 is not defined in the caption.
  6. [References] References [12] and [39] are the same paper and should be consolidated.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the reported improvements are empirical and externally evaluated.

full rationale

The derivation chain is not circular. FLOSS's mixture consistency is guaranteed by the projection parameterization v_theta = P_perp v_tilde and the initialization x0 = \bar S + P_perp Z: from Eqs. (9)-(14), P x_t = \bar S at every t, so outputs sum to the mixture by construction; this is a geometric constraint, not an evaluation metric encoded into the loss. The conditional-flow-matching objective (10) is a standard regression on (x1 - x0), and the PET selection rule (17), normalized loss (18), and dB loss (19) are training heuristics. The paper explicitly acknowledges these deviate from theory: "deviating from the theory of flow matching, we found that training on a decibel valued version of (18) leads to significant improvement" (Sec. 3.2). Such heuristics may weaken the claim that the learned ODE samples p(S|\bar s), but they do not define the prediction as a fitted copy of the reported metrics, and the SI-SDR/ETOI/PESQ numbers are computed on an independent evaluation set. The only self-referential element is the Diffsep baseline [23], a prior paper by the first author; but it is used as a comparative baseline, not as a justification of FLOSS's correctness or as a uniqueness theorem. There is no self-citation chain forcing the conclusion, no imported uniqueness theorem, and no renaming of a known empirical result as a new derivation. The unsupported theoretical claim is a correctness risk, not a circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The core construction is not circular: the velocity target is the conditional expectation of S minus Z derived from data, and mixture consistency follows from projection geometry. The main free parameters and assumptions are training heuristics, not quantities fitted to reproduce the reported metrics. The largest unproven step is the PET/dB-loss training objective, which is a heuristic replacement for exact equivariant flow matching.

free parameters (4)
  • Dirac mass weight in training time schedule = 0.01 (best row #1)
    Section 3.2 and Table 1; the fraction of t=0 training examples is tuned and changes SI-SDR from 15.98 to 18.43 across rows.
  • Noise envelope shaping T_env(bar s) and active power threshold = not specified quantitatively
    Section 3.4; envelope from Hamming window and threshold for active power; mixture-shaped noise e(bar s) outperforms constant active-power noise sigma_act in rows #8 vs #9.
  • Custom 5-step Euler schedule = {0.95, 4e-2, 9e-3, 9e-4, 1e-4}
    Section 4.4; found by hand to work well and produces the best SI-SDR of 19.13.
  • SNR-uniform schedule constants C and epsilon = not reported
    Section 3.2, row #8; the C/(t(1-t)) + epsilon delta_t schedule is stated with unspecified constants.
assumptions (5)
  • standard math Conditional flow matching with linear interpolants transports p0(·|c) to p1(·|c) when the regression loss is minimized.
    Invoked in Section 2.3 and Section 3.1 to justify objective (10).
  • standard math The learned drift can be restricted to the mixture-invariant subspace P-perp without loss of generality.
    Derived from equations (11)-(14); P-perp(S minus Z) lies in P-perp.
  • ad hoc to paper Permutation invariant training with an equivariant network generates samples from the permutation-invariant conditional p(S|bar s).
    Section 3.2 introduces PET as a replacement for exact equivariant OT flow matching; no distributional proof is supplied.
  • ad hoc to paper Normalized inverse-SNR loss (18) and dB loss (19) preserve enough of the flow-matching target to train a valid velocity field.
    Section 3.2 explicitly says these losses deviate from the theory of flow matching; their effect is empirical.
  • domain assumption Euler sampling with 1, 5, or 25 steps accurately approximates the learned ODE.
    Section 4.4; no discretization error analysis is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Source Separation by Flow Matching." pith.science (2026). https://pith.science/paper/E2TXEIX7

@misc{pith2026250516119,
  author       = {Pith},
  title        = {Pith review of: Source Separation by Flow Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E2TXEIX7}},
  note         = {Machine review of arXiv:2505.16119}
}
abstract

We consider the problem of single-channel audio source separation with the goal of reconstructing $K$ sources from their mixture. We address this ill-posed problem with FLOSS (FLOw matching for Source Separation), a constrained generation method based on flow matching, ensuring strict mixture consistency. Flow matching is a general methodology that, when given samples from two probability distributions defined on the same space, learns an ordinary differential equation to output a sample from one of the distributions when provided with a sample from the other. In our context, we have access to samples from the joint distribution of $K$ sources and so the corresponding samples from the lower-dimensional distribution of their mixture. To apply flow matching, we augment these mixture samples with artificial noise components to match the dimensionality of the $K$ source distribution. Additionally, as any permutation of the sources yields the same mixture, we adopt an equivariant formulation of flow matching which relies on a neural network architecture that is equivariant by design. We demonstrate the performance of the method for the separation of overlapping speech.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

  1. Conditional Flow Matching for Visually-Guided Acoustic Highlighting

    eess.AS 2026-02 conditional novelty 6.0 of 10

    Conditional flow matching with a rollout loss and early audio-visual fusion achieves state-of-the-art results on visually-guided acoustic highlighting.

  2. UniFlow: Unifying Speech Front-End Tasks via Continuous Generative Modeling

    eess.AS 2025-08 conditional novelty 6.0 of 10

    UniFlow unifies four speech front-end tasks in one continuous-latent generative model with task-ID conditioning and reports competitive, but not uniformly superior, benchmark scores.

  3. SequenceLayers: Sequence Processing and Streaming Neural Networks Made Easy

    cs.LG 2025-07 conditional novelty 6.0 of 10

    SequenceLayers defines a layer contract with explicit state and step methods so any composed sequence model is immediately streamable with tested layer-step equivalence.

Reference graph

Works this paper leans on

45 extracted references · 39 canonical work pages · cited by 3 Pith papers

  1. [1]

    Makino, Ed., Audio Source Separation, ser

    S. Makino, Ed., Audio Source Separation, ser. Signals and Communication Technology. Springer International Publishing, Jan. 2018

    work page 2018

  2. [2]

    Static and dynamic source separation using nonnegative factorizations: A unified view,

    P. Smaragdis, C. F ´evotte, G. J. Mysore, N. Mohammadiha, and M. Hoffman, “Static and dynamic source separation using nonnegative factorizations: A unified view,” IEEE Signal Process. Mag., vol. 31, no. 3, pp. 66–75, Apr. 2014

    work page 2014

  3. [3]

    Deep clustering: Discriminative embeddings for segmentation and separation,

    J. R. Hershey, Z. Chen, J. Le Roux, and S. Watanabe, “Deep clustering: Discriminative embeddings for segmentation and separation,” in ICASSP, Mar. 2016, pp. 31–35

    work page 2016

  4. [4]

    Multitalker speech separation with utterance-level permutation invariant training of deep recurrent neural networks,

    M. Kolbaek, D. Yu, Z.-H. Tan, and J. Jensen, “Multitalker speech separation with utterance-level permutation invariant training of deep recurrent neural networks,” IEEE/ACM Trans. Audio Speech Lang. Process., vol. 25, no. 10, pp. 1901–1913, Aug. 2017

    work page 1901

  5. [5]

    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 Trans. Audio Speech Lang. Process., vol. 27, no. 8, pp. 1256–1266, Aug. 2019

    work page 2019

  6. [6]

    High fidelity speech enhancement with band-split RNN,

    J. Yu, H. Chen, Y . Luo, R. Gu, and C. Weng, “High fidelity speech enhancement with band-split RNN,” in INTERSPEECH, 2023, pp. 2483– 2487

    work page 2023

  7. [7]

    Mel-band RoFormer for music source separation,

    J.-C. Wang, W.-T. Lu, and M. Won, “Mel-band RoFormer for music source separation,” in ISMIR, Nov. 2023

    work page 2023

  8. [8]

    TF- locoformer: Transformer with local modeling by convolution for speech separation and enhancement,

    K. Saijo, G. Wichern, F. G. Germain, Z. Pan, and J. L. Roux, “TF- locoformer: Transformer with local modeling by convolution for speech separation and enhancement,” in IWAENC, Sep. 2024, pp. 205–209

    work page 2024

Show all 45 references
  1. [9]

    Autoencoding beyond pixels using a learned similarity metric,

    A. B. L. Larsen, S. K. Sønderby, H. Larochelle, and O. Winther, “Autoencoding beyond pixels using a learned similarity metric,” in ICML, 2016, pp. 1558–1566

    2016

  2. [10]

    Generative adversarial source separa- tion,

    Y . C. Subakan and P. Smaragdis, “Generative adversarial source separa- tion,” in ICASSP, Apr. 2018, pp. 26–30

    2018

  3. [11]

    Source separation with deep generative priors,

    V . Jayaram and J. Thickstun, “Source separation with deep generative priors,” in ICML, 2020, pp. 4724–4735

    2020

  4. [12]

    Deep unsupervised learning using nonequilibrium thermodynamics,

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” in ICML, 2015

    2015

  5. [14]

    Score-based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” in ICLR, 2021

    2021

  6. [15]

    WaveGrad: Estimating gradients for waveform generation,

    N. Chen, Y . Zhang, H. Zen, R. J. Weiss, M. Norouzi, and W. Chan, “WaveGrad: Estimating gradients for waveform generation,” in ICLR, May 2021

    2021

  7. [16]

    PriorGrad: Improving conditional denoising diffusion models with data-dependent adaptive prior,

    S.-g. Lee, H. Kim, C. Shin, X. Tan, C. Liu, Q. Meng, T. Qin, W. Chen, S. Yoon, and T.-Y . Liu, “PriorGrad: Improving conditional denoising diffusion models with data-dependent adaptive prior,” in ICLR, Apr. 2022

    2022

  8. [17]

    SpecGrad: Diffusion probabilistic model based neural vocoder with adaptive noise spectral shaping,

    Y . Koizumi, H. Zen, K. Yatabe, N. Chen, and M. Bacchiani, “SpecGrad: Diffusion probabilistic model based neural vocoder with adaptive noise spectral shaping,” in INTERSPEECH, Sep. 2022, pp. 803–807

    2022

  9. [18]

    Diffusion models for audio restoration: A review,

    J.-M. Lemercier, J. Richter, S. Welker, E. Moliner, V . V ¨alim¨aki, and T. Gerkmann, “Diffusion models for audio restoration: A review,” IEEE Signal Process. Mag. , vol. 41, no. 6, pp. 72–84, Nov. 2024

    2024

  10. [19]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,” in NeurIPS Workshop on Deep Generative Models and Downstream Applications , 2021

    2021

  11. [20]

    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, Jun. 2023, pp. 1–5

    2023

  12. [21]

    Multi-source diffusion models for simultaneous music generation and separation,

    G. Mariani, I. Tallini, E. Postolache, M. Mancusi, L. Cosmo, and E. Rodol `a, “Multi-source diffusion models for simultaneous music generation and separation,” in ICLR, 2024

    2024

  13. [22]

    Noise-robust speech separation with fast generative correction,

    H. Wang, J. Villalba, L. Moro-Velazquez, J. Hai, T. Thebaud, and N. Dehak, “Noise-robust speech separation with fast generative correction,” arXiv preprint arXiv:2406.07461 , 2024

    2024 arXiv

  14. [23]

    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, Jun. 2023, pp. 1–5

    2023

  15. [24]

    EDSep: An effective diffusion-based method for speech source separation,

    J. Dong, X. Wang, and Q. Mao, “EDSep: An effective diffusion-based method for speech source separation,” in ICASSP, Apr. 2025

    2025

  16. [25]

    Non-denoising forward-time diffusions,

    S. Peluchetti, “Non-denoising forward-time diffusions,” https:// openreview.net/forum?id=oVfIKuhqfC, 2021

    2021

  17. [26]

    Diffusion Schr¨odinger bridge with applications to score-based generative modeling,

    V . De Bortoli, J. Thornton, J. Heng, and A. Doucet, “Diffusion Schr¨odinger bridge with applications to score-based generative modeling,” in NeurIPS, 2021

    2021

  18. [27]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” in ICLR, 2023

    2023

  19. [28]

    Stochastic interpolants: A unifying framework for flows and diffusions,

    M. S. Albergo, N. M. Boffi, and E. Vanden-Eijnden, “Stochastic interpolants: A unifying framework for flows and diffusions,” arXiv preprint arXiv:2303.08797, 2023

    2023 arXiv

  20. [29]

    Let us build bridges: Under- standing and extending diffusion generative models,

    X. Liu, L. Wu, M. Ye, and Q. Liu, “Let us build bridges: Under- standing and extending diffusion generative models,” arXiv preprint arXiv:2208.14699, 2022

    2022 arXiv

  21. [30]

    Zero-shot image restoration using denoising diffusion null-space model,

    Y . Wang, J. Yu, and J. Zhang, “Zero-shot image restoration using denoising diffusion null-space model,” in ICLR, Sep. 2022

    2022

  22. [31]

    Equivariant flow matching,

    L. Klein, A. Kr¨amer, and F. No´e, “Equivariant flow matching,” inNeurIPS, vol. 36, 2024

    2024

  23. [32]

    Diffusion posterior sampling for general noisy inverse problems,

    H. Chung, J. Kim, M. T. Mccann, M. L. Klasky, and J. C. Ye, “Diffusion posterior sampling for general noisy inverse problems,” in ICLR, 2023

    2023

  24. [33]

    Improving and generalizing flow-based generative models with minibatch optimal transport,

    A. Tong, K. Fatras, N. Malkin, G. Huguet, Y . Zhang, J. Rector- Brooks, G. Wolf, and Y . Bengio, “Improving and generalizing flow-based generative models with minibatch optimal transport,” Transact. mach. learn. res., 2024

    2024

  25. [34]

    Conditional Wasserstein distances with applications in Bayesian OT flow matching,

    J. Chemseddine, P. Hagemann, G. Steidl, and C. Wald, “Conditional Wasserstein distances with applications in Bayesian OT flow matching,” arXiv preprint arXiv:2403.18705 , 2024

    2024 arXiv

  26. [35]

    Dynamic conditional optimal transport through simulation-free flows,

    G. Kerrigan, G. Migliorini, and P. Smyth, “Dynamic conditional optimal transport through simulation-free flows,” in NeurIPS, vol. 37, 2024, pp. 93 602–93 642

    2024

  27. [36]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” in ICCV, 2023, pp. 4195–4205

    2023

  28. [37]

    Neural diffusion processes,

    V . Dutordoir, A. Saul, Z. Ghahramani, and F. Simpson, “Neural diffusion processes,” in ICML, Jul. 2023

    2023

  29. [38]

    SDR–half-baked or well done?

    J. Le Roux, S. Wisdom, H. Erdogan, and J. R. Hershey, “SDR–half-baked or well done?” in ICASSP, May 2019, pp. 626–630

    2019

  30. [39]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in NeurIPS, vol. 33, 2020, pp. 6840–6851

    2020

  31. [40]

    Looking to listen at the cocktail party: Audio-visual speech separation,

    A. Ephrat, I. Mosseri, O. Lang, T. Dekel, K. Wilson, B. Freeman, and M. Rubinstein, “Looking to listen at the cocktail party: Audio-visual speech separation,” in CVPR, 2018

    2018

  32. [41]

    Miipher: A robust speech restoration model integrating self-supervised speech and text representations,

    Y . Koizumi, H. Zen, S. Karita, Y . Ding, K. Yatabe, N. Morioka, Y . Zhang, W. Han, A. Bapna, and M. Bacchiani, “Miipher: A robust speech restoration model integrating self-supervised speech and text representations,” in WASPAA, 2023, pp. 1–5

    2023

  33. [42]

    SDR – half- baked or well done?

    J. Le Roux, S. Wisdom, H. Erdogan, and J. R. Hershey, “SDR – half- baked or well done?” in ICASSP, May 2019

    2019

  34. [43]

    An algorithm for predicting the intelligibility of speech masked by modulated noise maskers,

    J. Jensen and C. H. Taal, “An algorithm for predicting the intelligibility of speech masked by modulated noise maskers,” IEEE/ACM Trans. Audio Speech Lang. Process., vol. 24, no. 11, pp. 2009–2022, Nov. 2016

    2009

  35. [44]

    Perceptual evaluation of speech quality (PESQ),

    A. W. Rix, J. G. Beerends, M. P. Hollier, and A. P. Hekstra, “Perceptual evaluation of speech quality (PESQ),” in ICASSP, Jan. 2001, pp. 749–752

    2001

  36. [45]

    Perceptual objective listening quality assessment (POLQA), the third generation ITU-T standard for end-to-end speech quality measurement part I–Temporal alignment,

    J. G. Beerends, A. E. S. Fellow, C. Schmidmer, J. Berger, M. Obermann, R. Ullmann, J. Pomy, M. Keyhl, and AES Member, “Perceptual objective listening quality assessment (POLQA), the third generation ITU-T standard for end-to-end speech quality measurement part I–Temporal align...

    2013

  37. [46]

    DNSMOS P.835: A Non- Intrusive Perceptual Objective Speech Quality Metric to Evaluate Noise Suppressors,

    C. K. A. Reddy, V . Gopal, and R. Cutler, “DNSMOS P.835: A Non- Intrusive Perceptual Objective Speech Quality Metric to Evaluate Noise Suppressors,” in ICASSP, 2022, pp. 886–890

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.