pith. sign in

arxiv: 2606.16551 · v2 · pith:Z5S7K4WYnew · submitted 2026-06-15 · 📡 eess.AS

Learning Input-Channel Permutation Equivariance for Multi-Channel Source Separation: Reducing Bleeding in Small Music Ensembles

Ruchi Pandey , Jaime Garcia-Martinez , Pablo Cabanas-Molero , David Diaz-Guerra , Ricardo Falcon Perez , Tuomas Virtanen , Julio J. Carabias-Orti , Pedro Vera-Candeas This is my paper

Pith reviewed 2026-07-02 22:36 UTC · model grok-4.3

classification 📡 eess.AS
keywords multi-channel source separationmicrophone bleedpermutation equivariancemusic ensemblessynthetic training dataSDR improvementURMP recordings
0
0 comments X

The pith

Applying the same random permutation to input channels and targets during training improves source separation and reduces bleeding in music ensembles.

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

The paper establishes that enforcing permutation equivariance as a training principle allows multi-channel separation models to avoid learning fixed associations between specific microphones and instruments. By applying identical random permutations to the input mixture channels and their reference targets on each example, the network becomes robust to arbitrary channel ordering. Training occurs on synthetic ensembles that vary room acoustics and microphone placements, with evaluation on both held-out simulations and real URMP recordings. This yields consistent gains in SDR and lower bleeding relative to non-permutation baselines. The approach matters because real-world recording setups frequently change microphone positions and instrument layouts, breaking assumptions of fixed channel assignments.

Core claim

The central claim is that permutation-aware training, achieved by applying the same random permutation to the input microphone channels and their corresponding reference targets, produces models that generalize better to unseen conditions and real recordings, resulting in higher SDR and reduced microphone bleed compared with standard training.

What carries the argument

The mechanism of channel-permutation-equivariance created by applying identical random permutations to inputs and targets, which removes dependence on fixed channel-instrument mappings.

If this is right

  • Higher SDR is obtained under unseen simulated acoustic conditions.
  • Bleeding is reduced on real URMP ensemble recordings relative to non-permutation baselines.
  • The model becomes more robust to changes in microphone placement and recorded instruments.
  • The strategy functions as a data-centric addition that does not require architectural changes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same matched-permutation idea could be tested on other multi-channel tasks such as spatial audio rendering or microphone array processing.
  • Performance on larger ensembles or highly reverberant real rooms would indicate whether the synthetic-to-real transfer holds beyond the URMP conditions.
  • If channel count varies across recordings, an extension that also handles variable numbers of channels could be examined.

Load-bearing premise

That training on synthetic ensembles with simulated acoustics and mic placements produces a model that generalizes to real URMP recordings without the simulation details dominating the learned behavior.

What would settle it

Measure SDR and a bleed metric on the real URMP recordings; if the permutation-trained model shows no improvement or worse results than the non-permutation baseline, the central claim is falsified.

Figures

Figures reproduced from arXiv: 2606.16551 by David Diaz-Guerra, Jaime Garcia-Martinez, Julio J. Carabias-Orti, Pablo Cabanas-Molero, Pedro Vera-Candeas, Ricardo Falcon Perez, Ruchi Pandey, Tuomas Virtanen.

Figure 1
Figure 1. Figure 1: Semicircular string-ensemble layout used in simula [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
read the original abstract

Microphone bleed is a persistent challenge in small ensembles and orchestral recordings, where close microphones intended for individual instruments also capture leakage from nearby sources. This overlap degrades track isolation and complicates mixing. This paper addresses the bleeding problem by making channel-permutation-equivariance a core learning principle. During training, we apply the same random permutation to the input microphone channels and their corresponding reference targets. This discourages reliance on fixed channel-instrument associations and improves robustness to changes in the recording setup and even in the recorded instruments. The proposed model is trained on synthetic ensembles with diverse simulated room acoustics and microphone placements, and evaluated on unseen simulated conditions and real URMP recordings. The results show that permutation-aware training consistently improves SDR and reduces bleeding under unseen conditions compared with non-permutation baselines. The findings highlight permutation-equivariance as a simple, data-centric strategy for robust debleeding and practical multi-channel source separation in music production workflows.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes enforcing input-channel permutation equivariance during training of multi-channel source separation models for music by applying identical random permutations to the microphone input channels and their corresponding reference targets. This is intended to discourage fixed channel-instrument associations, reduce microphone bleed, and improve robustness to unseen recording setups and instruments. The model is trained solely on synthetic ensembles with simulated room acoustics and microphone placements, then evaluated on both unseen simulated conditions and real URMP recordings, with the claim that permutation-aware training yields consistent SDR improvements and reduced bleeding relative to non-permutation baselines.

Significance. If the central empirical claim holds with adequate controls, the work offers a simple, architecture-agnostic training strategy that could enhance generalization in practical multi-microphone music separation without requiring new model components or loss terms. The focus on permutation equivariance directly targets a common failure mode in ensemble recordings.

major comments (2)
  1. [Evaluation] Evaluation section: the abstract and evaluation description state 'consistent SDR gains' and 'reduced bleeding' on real URMP recordings but supply no numerical SDR values, baseline model specifications, error bars, or statistical tests, preventing assessment of effect size or reliability.
  2. [Training and Data] Training and Data section: the generalization claim from synthetic training data to real URMP recordings is load-bearing for the central contribution, yet no domain-gap controls, simulation-fidelity ablations, or mismatched-parameter experiments are reported to rule out the possibility that learned equivariance exploits simulation-specific artifacts rather than true channel-permutation robustness.
minor comments (1)
  1. [Abstract] Abstract: consider adding one or two key quantitative results (e.g., average SDR delta) to make the performance claim concrete.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight important gaps in the presentation of results and controls. We address each major comment below and commit to revisions that strengthen the manuscript without altering its core claims.

read point-by-point responses

  1. Referee: [Evaluation] Evaluation section: the abstract and evaluation description state 'consistent SDR gains' and 'reduced bleeding' on real URMP recordings but supply no numerical SDR values, baseline model specifications, error bars, or statistical tests, preventing assessment of effect size or reliability.

    Authors: We agree that the evaluation section lacks the requested numerical details for the real URMP recordings. The manuscript text states that permutation-aware training improves SDR and reduces bleeding, but does not tabulate the actual values, baselines, or variability measures. In the revised version we will add a dedicated results table reporting SDR (and other metrics) for both simulated and real conditions, with means, standard deviations across runs or folds, baseline model architectures and hyperparameters, and paired statistical tests (e.g., Wilcoxon signed-rank) to quantify reliability and effect size. revision: yes

  2. Referee: [Training and Data] Training and Data section: the generalization claim from synthetic training data to real URMP recordings is load-bearing for the central contribution, yet no domain-gap controls, simulation-fidelity ablations, or mismatched-parameter experiments are reported to rule out the possibility that learned equivariance exploits simulation-specific artifacts rather than true channel-permutation robustness.

    Authors: The referee is correct that the current manuscript reports no explicit domain-gap controls, simulation-fidelity ablations, or mismatched-parameter experiments. While the training data already incorporate diverse room acoustics and microphone placements, and real URMP evaluation provides an external test, these additional controls would more rigorously isolate the contribution of permutation equivariance. We will therefore add, in the revision, a set of controlled experiments that vary simulation parameters (e.g., RT60 mismatch, microphone directivity mismatch) and report the resulting SDR differences, thereby addressing the concern that benefits might stem from simulation artifacts. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical augmentation and evaluation on held-out data

full rationale

The paper implements permutation-equivariance via a standard data-augmentation procedure (identical random permutation applied to inputs and targets) and reports empirical SDR gains on unseen simulated and real recordings. No equations, uniqueness theorems, or self-citations are invoked that would reduce the reported improvement to a quantity defined by the same fitted data or prior author work. The central claim rests on direct comparison to non-permutation baselines rather than any self-referential derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that synthetic data distributions are representative enough for real-world generalization and that the permutation operation alone suffices to induce the desired equivariance without further architectural constraints.

axioms (1)
  • domain assumption Synthetic ensembles with diverse simulated room acoustics and microphone placements are sufficiently representative of real recording conditions for generalization claims.
    Invoked to support evaluation on unseen simulated conditions and real URMP recordings.

pith-pipeline@v0.9.1-grok · 5731 in / 1136 out tokens · 26634 ms · 2026-07-02T22:36:36.803075+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 4 canonical work pages

  1. [1]

    Close-microhpone cross-talk cancellationin ensembler record- ings with statistical estimation,

    O. Das, “Close-microhpone cross-talk cancellationin ensembler record- ings with statistical estimation,” Ph.D. dissertation, Standford University, 2021

    2021

  2. [2]

    Automatic noise gate set- tings for drum recordings containing bleed from secondary sources,

    M. Terrell, J. D. Reiss, and M. Sandler, “Automatic noise gate set- tings for drum recordings containing bleed from secondary sources,” EURASIP J. Adv. Signal Process., vol. 2010, no. 1, Dec. 2011

    2010

  3. [3]

    All for one and one for all: Improving music separation by bridging networks,

    R. Sawata, S. Uhlich, S. Takahashi, and Y . Mitsufuji, “All for one and one for all: Improving music separation by bridging networks,” inIEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2021, pp. 51–55

    2021

  4. [4]

    Hybrid spectrogram and waveform source separation,

    A. D ´efossez, “Hybrid spectrogram and waveform source separation,” in Proceedings of the ISMIR 2021 Workshop on Music Source Separation, 2021

    2021

  5. [5]

    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

  6. [6]

    The MUSDB18 corpus for music separation,

    Z. Rafii, A. Liutkus, F. St ¨oter, S. I. Mimilakis, and R. Bittner, “The MUSDB18 corpus for music separation,” Dec. 2017. [Online]. Available: https://doi.org/10.5281/zenodo.1117372

    work page doi:10.5281/zenodo.1117372 2017

  7. [7]

    Score-informed source separation for multichannel orchestral record- ings,

    M. Miron, J. J. Carabias-Orti, J. J. Bosch, E. G ´omez, and J. Janer, “Score-informed source separation for multichannel orchestral record- ings,”Journal of Electrical and Computer Engineering, vol. 2016, no. 1, p. 8363507, 2016

    2016

  8. [8]

    30+ years of source separation research: Achievements and future challenges,

    S. Arakiet al., “30+ years of source separation research: Achievements and future challenges,” inIEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2025, pp. 1–5

    2025

  9. [9]

    End-to-end microphone permutation and number invariant multi-channel speech separation,

    Y . Luoet al., “End-to-end microphone permutation and number invariant multi-channel speech separation,” inIEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2020, pp. 6394– 6398

    2020

  10. [10]

    Multi-microphone complex spectral mapping for utterance-wise and continuous speech separation,

    Z. Q. Wang, P. Wang, and D. Wang, “Multi-microphone complex spectral mapping for utterance-wise and continuous speech separation,” IEEE/ACM transactions on audio, speech, and language processing, vol. 29, pp. 2001–2014, 2021

    2001

  11. [11]

    Multi-channel speech separation using spatially selective deep non-linear filters,

    K. Tesch and T. Gerkmann, “Multi-channel speech separation using spatially selective deep non-linear filters,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 542–553, 2023

    2023

  12. [12]

    Mixture to mixture: Leveraging close-talk mixtures as weak-supervision for speech separation,

    Z. Q. Wang, “Mixture to mixture: Leveraging close-talk mixtures as weak-supervision for speech separation,”IEEE Signal Processing Letters, vol. 31, pp. 1715–1719, 2024

    2024

  13. [13]

    Direction specific ambisonics source separation with end-to-end deep learning,

    F. Llu ´ıset al., “Direction specific ambisonics source separation with end-to-end deep learning,”Acta Acustica, vol. 7, p. 29, 2023

    2023

  14. [14]

    Integrating high order ambisonics and deep learning for advanced instrument separation in spatial audio applica- tions,

    J. Garcia-Martinezet al., “Integrating high order ambisonics and deep learning for advanced instrument separation in spatial audio applica- tions,” inEuropean Signal Processing Conference (EUSIPCO), 2025, pp. 1253–1257

    2025

  15. [15]

    Leveraging synthetic data for improving chamber ensemble separation,

    S. Sarkaret al., “Leveraging synthetic data for improving chamber ensemble separation,” inIEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2023, pp. 1–5

    2023

  16. [16]

    Synthsod: Developing an heterogeneous dataset for orchestra music source separation,

    J. Garcia-Martinezet al., “Synthsod: Developing an heterogeneous dataset for orchestra music source separation,”IEEE Open Journal of Signal Processing, 2025

    2025

  17. [17]

    The cadenza woodwind dataset: Synthesised quartets for music information retrieval and machine learning,

    G. R. Dabikeet al., “The cadenza woodwind dataset: Synthesised quartets for music information retrieval and machine learning,”Data in Brief, vol. 57, p. 111199, 2024

    2024

  18. [18]

    Creating a multitrack classical music performance dataset for multimodal music analysis: Challenges, insights, and applications,

    B. Li, X. Liu, K. Dinesh, Z. Duan, and G. Sharma, “Creating a multitrack classical music performance dataset for multimodal music analysis: Challenges, insights, and applications,”IEEE Transactions on Multimedia, vol. 21, no. 2, pp. 522–535, 2018

    2018

  19. [19]

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

    D. Yuet al., “Permutation invariant training of deep models for speaker-independent multi-talker speech separation,” inIEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP. , 2017, pp. 241–245. [Online]. Available: https://doi.org/10. 1109/ICASSP.2017.7952154

    work page arXiv 2017

  20. [20]

    Event-independent network for polyphonic sound event localization and detection,

    Y . Caoet al., “Event-independent network for polyphonic sound event localization and detection,” inProceedings of 5th the Workshop on Detection and Classification of Acoustic Scenes and Events (DCASE), 2020, pp. 11–15

    2020

  21. [21]

    End-to-end recurrent multi-object tracking and trajectory prediction with relational reasoning,

    F. B. Fuchset al., “End-to-end recurrent multi-object tracking and trajectory prediction with relational reasoning,”CoRR, vol. abs/1907.12887, 2019. [Online]. Available: http://arxiv.org/abs/1907. 12887

    work page arXiv 1907

  22. [22]

    Pyroomacoustics: A python package for audio room simulation and array processing algorithms,

    R. Scheibleret al., “Pyroomacoustics: A python package for audio room simulation and array processing algorithms,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Apr. 2018, pp. 351–355. [Online]. Available: http: //dx.doi.org/10.1109/ICASSP.2018.8461310

    work page doi:10.1109/icassp.2018.8461310 2018

  23. [23]

    Acoustic design of large rehearsal spaces,

    P. Adamset al., “Acoustic design of large rehearsal spaces,” in International Symposium on Room Acoustics (ISRA), 2019. [Online]. Available: https://publications.rwth-aachen.de/record/772251

    2019

  24. [24]

    The 2018 signal separation evaluation campaign,

    F. R. St ¨oteret al., “The 2018 signal separation evaluation campaign,” inInternational Conference on Latent Variable Analysis and Signal Separation. Springer, 2018, pp. 293–305

    2018