Pith. sign in

REVIEW 3 major objections 7 minor 104 references

A single configuration-driven framework can train, evaluate, and improve many music demixing models without rebuilding the pipeline each time.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 23:18 UTC pith:RAVEIMUT

load-bearing objection Useful open MSS training stack with real community uptake; the science claims are thin and the TTA table does not carry the abstract’s “improvements” language. the 3 major comments →

arxiv 2607.23395 v1 pith:RAVEIMUT submitted 2026-07-25 cs.SD cs.LGeess.AS

Music-Source-Separation-Training (MSST): A Unified Framework for Training and Evaluating Music Demixing Models

Roman Solovyev , Ilya Kiselev , Alexander Stempkovskiy , Tatiana Gabruseva This is my paper
classification cs.SD cs.LGeess.AS
keywords music source separationaudio demixingdeep learningopen-source frameworktest-time augmentationmodel ensemblingLoRAYAML configuration
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Music source separation quality is decided less by any one network architecture than by the full engineering pipeline around it: data preparation, augmentations, losses, metrics, long-file inference, and post-processing. This paper presents MSST, an open, YAML-configurable system that puts many modern demixing model families behind one training, validation, and inference loop. It also builds in practical quality levers—overlapping sliding-window inference with cross-fading, test-time augmentation, weighted model ensembling, and LoRA fine-tuning—and reports ablations in which those levers raise standard separation scores. The aim is to make systematic experiments cheap and reproducible, so an idea can move quickly from config change to measured result. A sympathetic reader cares because karaoke, remixing, restoration, and production work all live or die on pipeline choices that most single-model codebases leave scattered or fixed.

Core claim

The authors claim that consolidating diverse music demixing architectures, dataset layouts, losses, metrics, multi-GPU training, and inference post-processing into one reproducible configuration-driven framework lowers the barrier to ablation studies, and that the framework’s built-in techniques—sliding windows with cross-fading, test-time augmentation, ensembling, and LoRA—produce measurable gains in separation quality on the metrics they track.

What carries the argument

MSST’s unified YAML-configurable training–validation–inference loop: one interface that swaps model families, dataset types, augmentations, weighted losses, and metrics while attaching shared quality tools (overlap-add inference, TTA, ensembles, LoRA) without rewriting the experiment harness.

Load-bearing premise

The small reported metric gains, especially tiny test-time-augmentation SDR lifts with no uncertainty or repeated runs shown, are treated as real quality improvements rather than evaluation noise on the chosen checkpoints and stems.

What would settle it

Re-run the Table 3 TTA and Table 4 ensemble comparisons on fixed public checkpoints and stems with multiple random seeds or bootstrap intervals; if the SDR deltas vanish inside the run-to-run spread, the claimed quality improvements from those techniques do not hold as stated.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Researchers can swap architectures and loss mixes via config and run comparable ablations without rebuilding trainers.
  • Long real-world tracks can be separated with overlap-add cross-fading and optional TTA without custom inference code.
  • Combining several strong models by weighted time-domain averaging can raise SDR beyond any single model in the set.
  • Large demixing models can be domain-adapted with LoRA instead of full retraining.
  • Community models for non-canonical stems (crowd, reverb, aspiration, phantom centre, drum-kit parts) can share one training and release path.

Where Pith is reading between the lines

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

  • If pipeline configuration truly dominates architecture choice, leaderboard gains may increasingly come from shared inference and ensemble recipes rather than new backbones alone.
  • A common config surface makes multi-lab reproduction of demixing results easier, which could tighten how future challenges specify baselines.
  • The same harness could be stress-tested on mixtures that are not linear stem sums (heavy mastering) to see whether Type-6 explicit-mixture training closes the studio-to-release gap.
  • Tiny TTA gains suggest the next useful experiment is cheaper invariances or selective TTA only on hard segments rather than always-on triple inference.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The manuscript describes MSST, an open-source (MIT) PyTorch framework that unifies training, validation, and inference for a broad set of music source separation model families (BS/Mel-Band RoFormer, HTDemucs, SCNet, MDX23C, BSMamba2, and others) behind a single YAML-configured pipeline. The paper surveys competing toolkits (§2, Table 1), formalizes the MSS problem and the supported loss/metric inventory (§§3, 6), documents seven dataset types and augmentations (§5), and describes practical techniques: sliding-window inference with cross-fading (§7.1), test-time augmentation (§7.2), LoRA (§7.3), flexible checkpoint loading (§7.4), ONNX/TensorRT export (§7.5), and weighted output ensembling (§8). Two small empirical tables report TTA effects (Table 3) and ensemble compositions (Table 4). The central contribution is a software/systems description of a framework that is publicly available and demonstrably adopted by the MSS community (§9).

Significance. If the description is accurate, the contribution is a genuinely useful piece of community infrastructure: MSST is a real, public, MIT-licensed codebase with a visible downstream ecosystem (community checkpoints for crowd extraction, dereverberation, drum-kit decomposition, OpenVINO conversions; §9), which is unusually strong adoption evidence for a tool paper. The scope claim — many modern demixing families under one configuration-driven training/validation/inference loop — is verifiable against the repository and appears consistent with it. The framework's support for reproducible, YAML-driven ablations and its loss/metric inventory are practical strengths. The empirical support for the secondary claim that TTA and ensembling "improve separation quality" is thin (Tables 3–4), but this is a secondary claim; the framework contribution stands independently.

major comments (3)
  1. [§7.2, Table 3] Table 3 and the surrounding text claim that TTA 'provides a small but consistent improvement in SDR across different models and stems.' Two problems. (a) The reported deltas are +0.00, +0.01, +0.03, +0.03, +0.06 dB — at or below the two-decimal reporting precision of the table, with no evaluation protocol given (no test set identified, no track count, no per-track statistics, no repeated runs, no significance test). As presented, these numbers cannot support 'consistent improvement'; they may be evaluation noise. (b) The improvement is not consistent even within the table: Bleedless degrades under TTA in four of five rows, most sharply for BS Roformer keys (33.81 -> 29.86, a 3.95-point drop), and Fullness moves in both directions. The honest reading of Table 3 is that TTA trades negligible SDR change for measurable bleed degradation. Either provide the evaluation protocol (dataset, check
  2. [§8, Table 4] The text states that ensembling 'consistently yields gains in the SDR metric, typically ranging from +0.01 to +0.2 dB,' but the paper's own Table 4 contradicts this: the ensembles improve over the best member by +0.37 (vocals: 11.24 -> 11.61), +0.46 (wind: 6.76 -> 7.22), +0.21 (crowd: 6.06 -> 6.27), and +0.25 dB (bass: 14.62 -> 14.87) — every row exceeds the stated upper bound of +0.2 dB. The sentence should be corrected to match the table, or the table explained. Relatedly, Table 4 lacks an evaluation protocol entirely: no dataset, no track count, no checkpoint provenance, no per-track statistics. Also, the bass row (members 13.81 and 14.62, weights 11 and 24, ensemble 14.87) yields an ensemble SDR above any convex combination of the member SDRs; this is possible for time-domain averaging followed by SDR computation, but the manuscript should say so explicitly rather than leave the arit
  3. [§§7–8 (evaluation protocol)] Neither Table 3 nor Table 4 identifies the evaluation data (MUSDB18 test? a private set? how many tracks?), the exact checkpoints used (which community releases?), or the inference configuration (chunk_size, num_overlap). For a paper whose stated purpose is reproducibility and 'verifiable results,' the two empirical tables are currently not reproducible. A short paragraph specifying dataset, checkpoints, and config per table would resolve this.
minor comments (7)
  1. [Throughout] Numerous typos: 'Hovewer' (§6.2), 'supportes' (§6.2), 'implenented' (§6.2), 'whch' (§2), 'famyily' (Table 1), 'trainingdatais' (§2). A proofreading pass is needed.
  2. [Table 1] Table 1 leaves the Licence column blank ('–') for most libraries; several of these have known licences (e.g., Demucs MIT, Spleeter MIT, Asteroid MIT). Either fill in or remove the column.
  3. [§3, Table 2] Table 2 lists 'minimal artefacts' as an advantage of hybrid methods; this is an overstatement — hybrid models have their own artefact profiles. Suggest 'reduced phase-related artefacts' or similar.
  4. [§1 / References (Solovyev and Kiselev [2023])] The framework version described is not pinned. Since the paper is dated July 2026 and describes the ecosystem 'as of July 2026,' please cite a specific commit hash or release tag of the repository so the description remains checkable as the codebase evolves.
  5. [§7.2] §7.2 states TTA uses 'two augmentations: stereo channel swap and polarity inversion,' and the final result averages 'the three runs (original plus two TTA variants)' — this is clear, but the cost claim ('proportionally increased inference time') should state the factor (3x) explicitly.
  6. [§6.1–6.2] The LogWMSE loss/metric description (§6.1, §6.2) cites only the torch_log_wmse implementation; if there is a defining publication for the metric, it should be cited.
  7. [§6.2, Eq. (2)] Eq. (2) defines SDR but the manuscript should state whether it is computed over full tracks or chunked (chunked SDR, as in MDX/SDX challenges, gives systematically different values), since this affects comparability of Tables 3–4 with published benchmarks.

Circularity Check

0 steps flagged

No circular derivation: MSST is a toolkit paper whose quality claims rest on empirical on/off ablations, not on definitions or self-fitted inputs renamed as predictions.

full rationale

The paper’s load-bearing claims are (1) that MSST unifies many demixing families under one YAML-configurable train/val/infer loop, and (2) that supported techniques (sliding-window inference, TTA, ensembling, LoRA) improve separation quality as shown by ablations (Abstract; §§7–8; Tables 3–4). Claim (1) is a software-scope statement, not a derivation. Claim (2) is supported by direct metric comparisons (SDR, Bleedless, Fullness, L1_freq) under toggled settings—e.g., TTA on vs off in Table 3, and weighted time-domain ensembles vs single models in Table 4—not by fitting a parameter to the target quantity and re-reporting it as a prediction. There is no self-definitional loop (no metric defined in terms of the claimed improvement), no uniqueness theorem imported from the authors to forbid alternatives, and no ansatz smuggled in via self-citation that forces the reported deltas. Self-references (Solovyev & Kiselev 2023 GitHub; community checkpoint releases) document the artifact and pretrained weights; they do not substitute for the measured on/off differences. Weaknesses in the ablations (tiny TTA SDR deltas, missing uncertainty) are evidence-quality issues, not circularity. Steps list is empty by design.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper’s load-bearing content is engineering consolidation plus small empirical ablations. It inherits the standard supervised MSS setup (linear stem sum, STFT/waveform/hybrid models, SDR-family metrics) and treats many training/inference hyperparameters as user-chosen rather than derived. No new physical entities are postulated; free parameters are the usual ML knobs and ensemble weights that affect the quality claims.

free parameters (5)
  • TTA transform set and equal averaging over original + channel-swap + polarity inversion = 3-run average; two transforms
    Table 3’s gains depend on this fixed, hand-chosen augmentation policy and uniform average; no search or uncertainty is reported.
  • Ensemble member weights (e.g., vocals 40/40/25; bass 11/24) = per-table integer weights
    Table 4 SDR depends on manually specified weights and model picks; optimal weighted time-domain average is stated empirically.
  • Loss component weights and multi_stft_resolution_loss_weight = config-dependent
    Final training objective is a configurable weighted sum; reported quality inherits these choices.
  • Inference chunk_size and num_overlap / cross-fade windows = YAML inference section
    Long-form quality and boundary artifacts depend on these hand-set inference parameters (§7.1).
  • training.q quantile for masked/robust losses = config training.q
    Robust losses discard large errors by a configurable quantile, directly shaping optimization.
axioms (4)
  • domain assumption A mixture is modeled as an additive combination of stems, x(t)=∑ s_i(t), even though real commercial mixes include nonlinear mastering.
    Stated in §3 as the mathematical problem setup and noted limitation.
  • domain assumption SDR and related instrumental metrics are adequate primary indicators of separation quality improvements for ablations.
    Tables 3–4 and default validation emphasize SDR (plus bleedless/fullness); paper itself notes SDR’s perceptual limits in §6.2.
  • domain assumption Standard deep learning optimization and STFT-based spectral losses are valid training signals for waveform/spectral/hybrid MSS models.
    Assumed throughout §§4–6 when composing losses and training configs.
  • ad hoc to paper Comparable baseline quality among ensemble members is required for ensembling to help.
    Explicit prerequisite in §8; without it the ensemble claim does not hold.

pith-pipeline@v1.2.0-grok45-kimik3 · 21119 in / 3263 out tokens · 74272 ms · 2026-07-30T23:18:48.005463+00:00 · methodology

0 comments
read the original abstract

Music Source Separation (MSS), the task of recovering individual sound components (stems) from a polyphonic mixture, is central to applications ranging from karaoke and remixing to audio restoration and content production. The separation quality depends on engineering decisions across the entire pipeline: model choice, training data preparation and augmentation, loss function and metrics choice, training configuration, validation, and post-processing. This paper presents MSST (Music-Source-Separation-Training) - a universal open-source framework for MSS tasks, which unifies training, validation, and inference for a broad range of modern demixing model families under a single, configuration-driven interface. The framework supports various model architectures, data preprocessing and augmentations, multiple loss functions and evaluation metrics, which helps with fast iterations and ablation studies. Additionally, the framework supports a range of practical techniques that improve separation quality, such as sliding-window inference with cross-fading, test-time augmentation, model ensembling, and fine-tuning via Low-Rank Adaptation (LORA). Our ablation studies demonstrate improvements of MSS using the above techniques. By consolidating these components into a reproducible, YAML-configurable framework, MSST lowers the barrier to systematic experimentation and enables rapid iteration from idea to verifiable result.

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

104 extracted references · 3 canonical work pages

  1. [1]

    2023 , howpublished =

    Jensen, Kimberley , title =. 2023 , howpublished =

    2023

  2. [2]

    2024 , howpublished =

    aufr33 and viperx , title =. 2024 , howpublished =

    2024

  3. [3]

    crowd\_mel\_band\_roformer\_aufr33\_viperx\_openvino , year =

  4. [4]

    2024 , howpublished =

    Solovyev, Roman , title =. 2024 , howpublished =

    2024

  5. [5]

    2024 , howpublished =

    Aspiration. 2024 , howpublished =

    2024

  6. [6]

    2024 , howpublished =

    2024

  7. [7]

    3rd International Conference on Learning Representations (ICLR) , year =

    Adam: A Method for Stochastic Optimization , author =. 3rd International Conference on Learning Representations (ICLR) , year =

  8. [8]

    7th International Conference on Learning Representations (ICLR) , publisher =

    Decoupled Weight Decay Regularization , author =. 7th International Conference on Learning Representations (ICLR) , publisher =. 2019 , url =

    2019

  9. [9]

    8th International Conference on Learning Representations (ICLR) , publisher =

    On the Variance of the Adaptive Learning Rate and Beyond , author =. 8th International Conference on Learning Representations (ICLR) , publisher =. 2020 , url =

    2020

  10. [10]

    Proceedings of the 41st International Conference on Machine Learning (ICML) , editor =

    Prodigy: An Expeditiously Adaptive Parameter-Free Learner , author =. Proceedings of the 41st International Conference on Machine Learning (ICML) , editor =. 2024 , url =

    2024

  11. [11]

    Proceedings of the 40th International Conference on Machine Learning (ICML) , editor =

    Learning-Rate-Free Learning by D-Adaptation , author =. Proceedings of the 40th International Conference on Machine Learning (ICML) , editor =. 2023 , url =

    2023

  12. [12]

    10th International Conference on Learning Representations (ICLR) , publisher =

    8-bit Optimizers via Block-wise Quantization , author =. 10th International Conference on Learning Representations (ICLR) , publisher =. 2022 , url =

    2022

  13. [13]

    2023 , url =

    Dettmers, Tim and Pagnoni, Artidoro and Holtzman, Ari and Zettlemoyer, Luke , booktitle =. 2023 , url =

    2023

  14. [14]

    2024 , url =

    Muon: An Optimizer for Hidden Layers in Neural Networks , author =. 2024 , url =

    2024

  15. [15]

    High-Performance Computing at

    Kostenetskiy, Pavel and Kozyrev, Vyacheslav and Chulkevich, Roman and Salikova, Mariia and Mishenin, Roman , booktitle =. High-Performance Computing at. 2026 , doi =

    2026

  16. [16]

    2023 , howpublished =

    Solovyev, Roman and Kiselev, Ilya , title =. 2023 , howpublished =

    2023

  17. [17]

    Journal of Open Source Software , volume =

    Open-Unmix --- A Reference Implementation for Music Source Separation , author =. Journal of Open Source Software , volume =. 2019 , publisher =

    2019

  18. [18]

    Sound Demixing Challenge 2023 Technical Report:

    Kim, MinSeok and Lee, Jun Hyung and Jung, Soonyoung , journal =. Sound Demixing Challenge 2023 Technical Report:

    2023

  19. [19]

    Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , year =

    Hybrid Transformers for Music Source Separation , author =. Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , year =

  20. [20]

    IEEE/ACM Transactions on Audio, Speech, and Language Processing , year =

    A Generalised Bandsplit Neural Network for Cinematic Audio Source Separation , author =. IEEE/ACM Transactions on Audio, Speech, and Language Processing , year =

  21. [21]

    Bandit v2 --- Three-stem separation model (speech, music, effects) , year =

  22. [22]

    Kim, Eunwoo and Choi, Yong-Hwa , journal =

  23. [23]

    Proceedings of Interspeech , pages =

    Conformer: Convolution-augmented Transformer for Speech Recognition , author =. Proceedings of Interspeech , pages =

  24. [24]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages =

    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages =

  25. [25]

    Hu, Edward and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , journal =

  26. [26]

    open-unmix-pytorch , year =

  27. [27]

    Spleeter: Deezer Source Separation Library Including Pretrained Models , year =

  28. [28]

    arXiv preprint arXiv:1911.13254 , year =

    Music Source Separation in the Waveform Domain , author =. arXiv preprint arXiv:1911.13254 , year =

    Pith/arXiv arXiv 1911

  29. [29]

    Demucs: Code for the Papers ``Hybrid Spectrogram and Waveform Source Separation'' and ``Music Source Separation in the Waveform Domain'' , year =

  30. [30]

    Asteroid: The

    Pariente, Manuel and Cornell, Samuele and Cosentino, Joris and Sivasankaran, Sunit and Tzinis, Efthymios and Heitkaemper, Jens and Olvera, Michel and St. Asteroid: The. Proceedings of Interspeech , year =

  31. [31]

    2020 , howpublished =

    Asteroid: The. 2020 , howpublished =

    2020

  32. [32]

    2021 , howpublished =

    mdx-net:. 2021 , howpublished =

    2021

  33. [33]

    Manilow, Ethan and Seetharaman, Prem and Pardo, Bryan , booktitle =. The

  34. [34]

    2018 , howpublished =

    nussl: A Flexible Source Separation Library in. 2018 , howpublished =

    2018

  35. [35]

    arXiv preprint arXiv:2112.03752 , year =

    Danna-Sep: Unite to Separate Them All , author =. arXiv preprint arXiv:2112.03752 , year =

    Pith/arXiv arXiv

  36. [36]

    2021 , howpublished =

    Yu, Chin-Yun , title =. 2021 , howpublished =

    2021

  37. [37]

    Proceedings of the International Conference on Latent Variable Analysis and Signal Separation (LVA/ICA) , pages =

    Monoaural Audio Source Separation Using Deep Convolutional Neural Networks , author =. Proceedings of the International Conference on Latent Variable Analysis and Signal Separation (LVA/ICA) , pages =

  38. [38]

    DeepConvSep: Deep Convolutional Neural Networks for Musical Source Separation , year =

  39. [39]

    Decoupling Magnitude and Phase Estimation with Deep

    Kong, Qiuqiang and Cao, Yin and Liu, Haohe and Choi, Keunwoo and Wang, Yuxuan , booktitle =. Decoupling Magnitude and Phase Estimation with Deep

  40. [40]

    music\_source\_separation , year =

  41. [41]

    music-demixing-challenge-starter-kit , year =

  42. [42]

    Transactions of the International Society for Music Information Retrieval , volume =

    The Sound Demixing Challenge 2023 – Music Demixing Track , author =. Transactions of the International Society for Music Information Retrieval , volume =. 2024 , doi =

    2023

  43. [43]

    Medleyvox: An Evaluation Dataset for Multiple Singing Voices Separation , year=

    Jeon, Chang-Bin and Moon, Hyeongi and Choi, Keunwoo and Chon, Ben Sangbae and Lee, Kyogu , booktitle=. Medleyvox: An Evaluation Dataset for Multiple Singing Voices Separation , year=

  44. [44]

    2024 , issn =

    RoFormer: Enhanced transformer with Rotary Position Embedding , journal =. 2024 , issn =. doi:https://doi.org/10.1016/j.neucom.2023.127063 , url =

    arXiv 2024

  45. [45]

    and Reiss, Joshua D

    Steinmetz, Christian J. and Reiss, Joshua D. , booktitle=. auraloss:

  46. [46]

    On Loss Functions and Evaluation Metrics for Music Source Separation , year=

    Gusó, Enric and Pons, Jordi and Pascual, Santiago and Serrà, Joan , booktitle=. On Loss Functions and Evaluation Metrics for Music Source Separation , year=

  47. [47]

    , booktitle=

    Roux, Jonathan Le and Wisdom, Scott and Erdogan, Hakan and Hershey, John R. , booktitle=. SDR – Half-baked or Well Done? , year=

  48. [48]

    arXiv preprint arXiv:2401.13276 , year=

    SCNet: Sparse Compression Network for Music Source Separation , author=. arXiv preprint arXiv:2401.13276 , year=

    Pith/arXiv arXiv

  49. [49]

    arXiv preprint arXiv:2309.02612 , year=

    Music Source Separation with Band-Split RoPE Transformer , author=. arXiv preprint arXiv:2309.02612 , year=

    Pith/arXiv arXiv

  50. [50]

    arXiv preprint arXiv:2310.01809 , year=

    Mel-Band RoFormer for Music Source Separation , author=. arXiv preprint arXiv:2310.01809 , year=

    Pith/arXiv arXiv

  51. [51]

    The Cocktail Fork Problem: Three-Stem Audio Separation for Real-World Soundtracks , booktitle =

    Petermann, Darius and Wichern, Gordon and Wang, Zhong-Qiu and. The Cocktail Fork Problem: Three-Stem Audio Separation for Real-World Soundtracks , booktitle =

  52. [52]

    Mitsufuji, Yuki and Haraguchi, Tatsuya and Takashima, Yoshikazu , title =

  53. [53]

    Mitsufuji, Yuki and Uhlich, Stefan , title =

  54. [54]

    16th International Workshop on Acoustic Signal Enhancement (

    Naoya Takahashi and Nabarun Goswami and Yuki Mitsufuji , title =. 16th International Workshop on Acoustic Signal Enhancement (. 2018 , month = sep, pages=. doi:10.1109/iwaenc.2018.8521383 , url =

    arXiv 2018

  55. [55]

    Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA) , year=

    Multi-scale multi-band densenets for audio source separation , author=. Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA) , year=

  56. [56]

    2017 , month = mar, publisher =

    Stefan Uhlich and Marcello Porcu and Franck Giron and Michael Enenkl and Thomas Kemp and Naoya Takahashi and Yuki Mitsufuji , title =. 2017 , month = mar, publisher =. doi:10.1109/icassp.2017.7952158 , url =

    arXiv 2017

  57. [57]

    Long Short-Term Memory , journal =

    Sepp Hochreiter and J\". Long Short-Term Memory , journal =. 1997 , month = nov, publisher =. doi:10.1162/neco.1997.9.8.1735 , url =

    doi:10.1162/neco.1997.9.8.1735 1997

  58. [58]

    Advances in Neural Information Processing Systems , pages =

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N and Kaiser,. Advances in Neural Information Processing Systems , pages =

  59. [59]

    doi:10.5281/zenodo.3338373 , url =

    Rafii, Zafar and Liutkus, Antoine and Stöter, Fabian-Robert and Mimilakis, Stylianos Ioannis and Bittner, Rachel , title =. doi:10.5281/zenodo.3338373 , url =

    doi:10.5281/zenodo.3338373

  60. [60]

    Rafii, Zafar and Liutkus, Antoine and Fabian-Robert St. The. 2017 , doi =

    2017

  61. [61]

    2017 , booktitle =

    Audio Set: An ontology and human-labeled dataset for audio events , author =. 2017 , booktitle =

    2017

  62. [62]

    Proceedings of the ISMIR 2021 Workshop on Music Source Separation , year=

    Hybrid Spectrogram and Waveform Source Separation , author=. Proceedings of the ISMIR 2021 Workshop on Music Source Separation , year=

    2021

  63. [63]

    Wave-U-Net:

    Daniel Stoller and Sebastian Ewert and Simon Dixon , editor =. Wave-U-Net:. Proceedings of the 19th International Society for Music Information Retrieval Conference,. 2018 , url =

    2018

  64. [64]

    Sound Demixing Challenge 2023 , year =

    2023

  65. [65]

    Music Demixing Challenge 2021 , journal =

    Yuki Mitsufuji and Giorgio Fabbro and Stefan Uhlich and Fabian-Robert St\". Music Demixing Challenge 2021 , journal =. 2022 , month = jan, publisher =. doi:10.3389/frsip.2021.808395 , url =

    arXiv 2021

  66. [66]

    arXiv preprint , year =

    Kim, Minseok and Choi, Woosung and Chung, Jaehwa and Lee, Daewon and Jung, Soonyoung , title =. arXiv preprint , year =. doi:10.48550/ARXIV.2111.12203 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2111.12203

  67. [67]

    The 2018 Signal Separation Evaluation Campaign

    St \"o ter, Fabian-Robert and Liutkus, Antoine and Ito, Nobutaka. The 2018 Signal Separation Evaluation Campaign. Latent Variable Analysis and Signal Separation: 14th International Conference, LVA/ICA 2018, Surrey, UK. 2018

    2018

  68. [68]

    2015 , eprint =

    Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun , title =. 2015 , eprint =

    2015

  69. [69]

    Ensembles in machine learning applications , year =

    Okun, Oleg and Valentini, Giorgio and Re, Matteo , publisher =. Ensembles in machine learning applications , year =

  70. [70]

    , booktitle =

    Gabruseva, Tatiana and Poplavskiy, Dmytro and Kalinin, Alexandr A. , booktitle =. Deep Learning for Automatic Pneumonia Detection , year =

  71. [71]

    The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Hu, Jie and Shen, Li and Sun, Gang , title =. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  72. [72]

    2018 , volume =

    Simone Bianco and Remi Cadene and Luigi Celona and Paolo Napoletano , title =. 2018 , volume =. doi:10.1109/access.2018.2877890 , publisher =

    arXiv 2018

  73. [73]

    2019 , month = may, publisher =

    Laure Pretet and Romain Hennequin and Jimena Royo-Letelier and Andrea Vaglio , title =. 2019 , month = may, publisher =. doi:10.1109/icassp.2019.8683555 , url =

    arXiv 2019

  74. [74]

    2020 , month = jun, publisher =

    Romain Hennequin and Anis Khlif and Felix Voituret and Manuel Moussallam , title =. 2020 , month = jun, publisher =. doi:10.21105/joss.02154 , url =

    work page doi:10.21105/joss.02154 2020

  75. [75]

    GUI for a Vocal Remover that uses Deep Neural Networks , year =

    Anjok07 and other contributors , howpublished =. GUI for a Vocal Remover that uses Deep Neural Networks , year =. GitHub repository , publisher =

  76. [76]

    MVSEP MDX23 music separation model , year =

    Solovyev, Roman , howpublished =. MVSEP MDX23 music separation model , year =. GitHub repository , publisher =

  77. [77]

    MVSEP CDX23 Cinematic Sound Demixing , year =

    Solovyev, Roman , howpublished =. MVSEP CDX23 Cinematic Sound Demixing , year =. GitHub repository , publisher =

  78. [78]

    2022 , copyright =

    Luo, Yi and Yu, Jianwei , title =. 2022 , copyright =. doi:10.48550/ARXIV.2209.15174 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2209.15174 2022

  79. [79]

    Kim\_Vocal\_1.onnx , year =

    Kimberley\_Jensen , howpublished =. Kim\_Vocal\_1.onnx , year =. GitHub repository , publisher =

  80. [80]

    Kim\_Inst.onnx , year =

    Kimberley\_Jensen , howpublished =. Kim\_Inst.onnx , year =. GitHub repository , publisher =

Showing first 80 references.