Pith. sign in

REVIEW 3 major objections 5 minor 44 references

The paper claims that treating an audio effect chain as a token sequence and predicting it autoregressively on source-separated stems yields a mixing style representation that outperforms contrastive learning and supports near-real-time sty

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 · deepseek-v4-flash

2026-08-01 22:43 UTC pith:FDSWQV6F

load-bearing objection StemFX is a solid within-subfield contribution with a controlled comparison that supports its main claim, though the headline speed and MUSHRA numbers need more uncertainty quantification. the 3 major comments →

arxiv 2607.15634 v2 pith:FDSWQV6F submitted 2026-07-17 cs.SD eess.AS

StemFX: Learning Mixing Style Representations via Autoregressive FX Chain Prediction on Source-Separated Stems

Yuan-Chiao Cheng , Jui-Te Wu , Brian Chen , Yen-Tung Yeh , Yu-Hua Chen , Yi-Hsuan Yang This is my paper
classification cs.SD eess.AS
keywords mixing style representationFX chain predictionaudio effectssource separationautoregressive generationmusic mixing style transferband-split encoderstem audio
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.

The paper aims to show that a mixing engineer's style—which effects are applied to each stem, in what order, with what settings—can be captured by treating the effect chain as a sentence and predicting it autoregressively from audio. StemFX trains a band-split encoder and a Transformer decoder end-to-end on paired original/augmented stems, where the augmented stems come from a pipeline that source-separates roughly 105,000 songs and applies random chains of 85 effects. Under matched architecture and data, this autoregressive objective beats a contrastive objective on mixing-style retrieval (86.8% vs 77.8% top-1 at eight effects). On paired real-mix transfer, the predicted chains achieve the best spectral fidelity and the highest listener preference among compared methods, and run in 0.24 seconds rather than roughly 1,000 seconds for iterative optimization. If true, the framework offers a scalable way to learn, retrieve, and transfer mixing style without scarce multitrack recordings.

Core claim

StemFX's central claim is that the right learning signal for mixing style is not a similarity contrast between augmented and unaugmented audio, but the exact reconstruction of the effect chain that produced the augmented audio. The paper shows that when the encoder and decoder are trained jointly to predict tokenized, variable-length, per-stem FX chains, the resulting embeddings retrieve the correct chain from a pool of 500 candidates more accurately than contrastively trained encoders—86.8% vs 77.8% top-1 at eight effects under matched architecture and data—and that the same model can transfer a real mix's style by executing its predicted chain, reaching an MRSTFT of 1.44 and a mean MUSHRA

What carries the argument

The machinery has two coupled halves. On the audio side, a shared band-split multi-band CNN encoder with feature-wise linear modulation (FiLM) conditioning maps the original and augmented four-stem inputs into two 512-dimensional embeddings, which are projected to conditioning vectors. On the language side, a six-layer Transformer decoder treats the FX chain as a flat token sequence drawn from a 358-token vocabulary (stem names, effect names, parameter names, quantized value bins) and generates it token by token with teacher forcing and cross-entropy loss, attending to both conditioning vectors. The Sep-Aug pipeline is the data machinery: source separation on about 105K songs produces pseudo

Load-bearing premise

The load-bearing premise is that pseudo-stems produced by source separation are faithful enough stand-ins for true multitrack stems that a model trained on them generalizes to real mixing decisions.

What would settle it

Train StemFX on clean, true multitrack stems (for example, professionally recorded sessions) using the same FX augmentation and evaluation protocol, and compare retrieval and transfer to the pseudo-stem-trained model. If accuracy drops markedly or the model learns to rely on separation artifacts, the Sep-Aug pipeline's proxy assumption fails; if performance holds, the representation is genuinely about mixing decisions.

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

If this is right

  • Under matched architecture and training data, an autoregressive FX-chain prediction objective produces stronger mixing-style embeddings than contrastive learning, with top-1 retrieval rising from 77.8% to 86.8% at eight effects.
  • Choosing which effects to apply and in what order matters more than fine-tuning parameters of a fixed universal chain: free generation beats all fixed-chain methods on both synthetic and real-mix transfer.
  • Mixing style transfer can run in near real time (0.24 seconds per example) instead of iterative optimization that takes about 1,000 seconds, making interactive use plausible.
  • Paired training data for mixing style can be scaled to roughly 105,000 songs through separation and augmentation, with retrieval accuracy improving monotonically as the training set grows (58.6% to 86.8% top-1).
  • The predicted chains are human-readable and executable, so an engineer can inspect, edit, and apply the suggested effect chain rather than receiving a black-box audio transformation.

Where Pith is reading between the lines

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

  • Editorial inference: the paper leaves open whether separation artifacts distort the representation; a head-to-head training run on true multitrack stems with the same FX augmentation would isolate that effect.
  • Editorial inference: the autoregressive formulation naturally extends to predicting genre-specific or engineer-specific chain priors, since the token vocabulary is extensible; that would test whether the representation captures structured mixing conventions rather than random compositions.
  • Editorial inference: the encoder's role as a style summarizer could be reused outside chain prediction, for example as a retrieval index over a mix library or as conditioning for a generative mixing model—applications the paper mentions but does not build.

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 / 5 minor

Summary. The paper introduces StemFX, a framework that learns mixing-style representations by autoregressively predicting tokenized FX chains on four source-separated stems. Training data are generated by a Sep-Aug pipeline: ~105K FMA songs are separated with SCNet into pseudo-stems, cross-song stem mixing is applied, and each stem is augmented with 1–10 effects from a new MultiAFx library. A BSFiLM encoder with FiLM conditioning maps original and augmented audio into conditioning vectors for a Transformer decoder, trained with teacher forcing and cross-entropy. Experiments cover mixing-style retrieval against AFx-Rep, Fx-Encoder++, CLAP, and a matched BSFiLM-CL baseline; paired mixing-style transfer against iterative optimization baselines; ablations removing FiLM, per-stem input, and the BSFiLM encoder; and dataset-scale analysis. The paper claims that autoregressive FX-chain prediction outperforms contrastive learning under matched architecture/data, and that near-real-time style transfer is achievable with best spectral fidelity and listener preference.

Significance. If the empirical claims hold, this is a significant contribution: it is the first end-to-end autoregressive FX-chain generator for mixing-style representation, it introduces a scalable pseudo-stem augmentation pipeline and an open-source MultiAFx toolkit, and it provides a controlled comparison against a contrastive baseline trained on the same data and architecture. The evaluation is well designed in several respects: retrieval and transfer use held-out MUSDB18 audio, the transfer targets use pedalboard effects never seen in training, and the ablations cleanly attribute gains to FiLM conditioning and per-stem input. The paper also ships source code and trained weights. However, the headline claims rest on quantitative gaps that are currently reported without uncertainty estimates, and the retrieval protocol leaves a load-bearing ambiguity about which embedding is used. The Sep-Aug premise is explicitly admitted to be unvalidated. These are fixable but require additional experiments or careful rescoping.

major comments (3)
  1. [§4.2, Fig. 2; §4.3, Table 2] The two load-bearing quantitative claims — that the autoregressive objective beats contrastive learning under matched conditions (STEMFX 86.8% vs BSFiLM-CL 77.8% Top-1 at 8 effects, Fig. 2) and that StemFX achieves best real-mix transfer (MRSTFT 1.44 and MUSHRA 60.6 vs low anchor 54.9, Table 2) — are reported without uncertainty estimates, number of training seeds, or significance tests. With 20 listeners, a 5.7-point MUSHRA margin is not obviously above noise; the MRSTFT margin over the FX-normalized input is only 0.05. The retrieval gap likewise needs seed-to-seed variance. Please provide multiple-seed means ± std, bootstrap or per-listener confidence intervals, and appropriate significance tests for the Fig. 2 curves, Fig. 4 scaling curves, and Table 2, and state the number of runs used for each.
  2. [§4.2, Fig. 1] The retrieval protocol is incompletely specified. The model has two encoder outputs e_orig and e_aug and projected conditioning vectors m1,m2 (Fig. 1), but §4.2 simply says retrieval is performed by 'cosine similarity between embeddings.' It is not stated which embedding is used for the query and candidate, nor which embedding BSFiLM-CL uses. Because the paired-input paradigm and the contrastive baseline may define embeddings differently, this ambiguity undermines the matched-architecture comparison. Please specify the exact embedding and retrieval protocol, and report results for each plausible variant (e_orig, e_aug, m1/m2, or their difference/concat) so the comparison is unambiguous.
  3. [§5; §3.7] As acknowledged in §5, the Sep-Aug pipeline's core premise is not validated: 'Pseudo-stems additionally inherit the errors of the separation model, and we do not quantify how these artifacts propagate into the learned representation; a comparison against a model trained on clean multitrack stems would isolate this effect.' Section 3.7 also notes that cross-song stem mixing may produce disharmonious combinations in different keys or tempos. Because the claims are about 'mixing style' generally rather than only pseudo-stem processing, this admitted gap is load-bearing. Add a clean-multitrack training comparison (e.g., on MUSDB18 or MedleyDB/MoisesDB stems) or a pseudo-stem artifact-injection study, or explicitly scope the claims to source-separated pseudo-stems in the title/abstract.
minor comments (5)
  1. [§3.1, Eq. (1)] Eq. (1) uses F_s for each stem-level chain and F for the full chain, but the serialization into a single token sequence y in §3.3 should be spelled out with the stem/effect separators used by the tokenizer. The exact order in which stems and effects are serialized affects the learning task.
  2. [Fig. 3] The legend label 'w/o sep. + FiLM' is ambiguous: the text says this variant removes both source separation and FiLM conditioning. Consider renaming it 'mixture input (w/o sep., w/o FiLM)' or 'stereo mixture input'.
  3. [§4.1] The retrieval candidate pool construction (N=500; how are candidates drawn and are queries excluded?) and the details of the MUSHRA statistical design are missing. A few sentences clarifying these would help reproducibility.
  4. [§3.3] The per-effect parameter normalization ranges (v_min, v_max) and the 101-bin quantization are free design choices. A sensitivity analysis or at least a brief discussion of their effect on retrieval/transfer accuracy would strengthen the paper.
  5. [General] There are minor typesetting issues: the title in the PDF has an unnatural line break ('Representa tions'), and some figure legend abbreviations are not expanded in captions. Please proofread.

Circularity Check

0 steps flagged

No load-bearing circularity: the headline comparisons are out-of-domain and not fitted; the only self-citation is a baseline, not a supporting premise.

full rationale

STEMFX's central derivation is not circular by construction. The training objective (Eq. 3) is supervised cross-entropy on tokenized FX chains for pseudo-stems. The retrieval experiments use MUSDB18 audio and pedalboard effects that are explicitly never used in training: 'pedalboard is not included in MultiAFx and was not used by any of the baseline methods during training.' The 86.8% vs. 77.8% figure is therefore not a fitted value relabeled as a prediction. The matched comparison BSFiLM-CL uses the same encoder and data with only the training objective changed, which is the correct control for the claim that autoregressive FX chain prediction beats contrastive learning. Transfer results likewise evaluate on real professional mixes and out-of-domain effects; no parameter is fitted to the test targets. The only self-citation is [13] (Fx-Encoder++, which shares co-authors Y.-T. Yeh and Y.-H. Yang); it is used as a baseline and as part of the retrieval protocol alongside [1], not as the justification for the framework's premise or any uniqueness claim, so it is non-load-bearing. The paper's own limitation about pseudo-stem artifacts (Sec. 5: 'we do not quantify how these artifacts propagate into the learned representation') is an honest missing control, not a circular step. Similarly, the absence of error bars in Figure 2 and Table 2 is a statistical robustness concern, not a form of circularity under this rubric.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central claims rest on three domain assumptions that the paper itself flags as untested: pseudo-stem fidelity, random chains approximating real mixing conventions, and the completeness of FX chains as a style representation. The free parameters are mostly hand-set thresholds and tokenization choices that shape the training and evaluation distributions. No physical entities are postulated; the two invented artifacts (MultiAFx, BSFiLM encoder) are software releases with independent handles.

free parameters (4)
  • Per-effect parameter normalization ranges (v_min, v_max)
    Hand-chosen for each of the 85 effects in MultiAFx (Section 3.3); log vs linear scaling changes how parameter values are tokenized and how hard value prediction is.
  • Stem exclusion threshold, integrated loudness < -40 dB = -40 dB
    Hand-set filter in the Sep-Aug pipeline (Section 4.1) that removes near-silent pseudo-stems and changes the training distribution.
  • MUSDB18 balanced-window selection threshold (>=10% energy per stem) = 10%
    Post-hoc selection criterion on evaluation clips (Section 4.1); affects the difficulty and representativeness of the Real Mix evaluation.
  • Value quantization (101 bins), 10-second clips, 1-10 effects/stem, macro-category constraint
    Architectural and data-generation choices (Sections 3.3, 4.1) that bound what the model can express and what effect structures it can learn.
axioms (4)
  • domain assumption SCNet source separation of FMA mixtures yields pseudo-stems faithful enough for learning generalizable mixing-style representations.
    The entire Sep-Aug pipeline depends on this. Section 5 admits: 'Pseudo-stems additionally inherit the errors of the separation model, and we do not quantify how these artifacts propagate into the learned representation.'
  • domain assumption Random uniform sampling of FX chains (1-10 effects, no consecutive same-category) is representative enough of real mixing style decisions.
    Section 5: 'training chains are sampled at random, whereas mix engineers apply structured, genre-dependent processing. Whether the representation captures such conventions beyond random effect compositions remains open.'
  • domain assumption Mixing style is adequately represented by per-stem FX chain identity/order/parameters; level-balancing and spatialization dimensions are not part of the learned objective.
    The abstract defines mixing style as including level balancing and spatialization, but the training objective (Section 3.5) and retrieval evaluation (Section 4.2) only involve FX chains, so the representation may be blind to parts of the style the paper itself names.
  • ad hoc to paper The 64 hand-crafted FiLM features (RMS, crest, band energies, stereo cues, etc.) are a sufficient conditioning summary of the mixing state.
    These are manually designed (adapted from Diff-MST, Table 1) and their sufficiency is assumed. The Section 4.4 ablation shows FiLM contributes about 12.4 percentage points, so this assumption matters but is partial.
invented entities (2)
  • MultiAFx toolkit independent evidence
    purpose: Unified wrapper over 85 audio effects from 7 Python libraries for random chain augmentation of pseudo-stems (Section 3.7).
    A new software artifact, not a physical entity. It is released on GitHub/pip and any chain can be executed and checked, so it carries an independent falsifiable handle.
  • BSFiLM encoder independent evidence
    purpose: Band-split multi-band CNN with FiLM conditioning for per-stem spectral structure (Section 3.2).
    A named architecture with released weights whose contribution is isolated by ablations (Figure 3); externally checkable, though it is a design rather than a postulated physical/conceptual entity.

pith-pipeline@v1.3.0-alltime-deepseek · 147 in / 15653 out tokens · 185007 ms · 2026-08-01T22:43:10.794905+00:00 · methodology

0 comments
read the original abstract

Audio mixing style encompasses the artistic and technical decisions a mix engineer makes, including level balancing, spatialization, and the choice, ordering, and parameterization of audio effects (FX) on each stem. FX chains are a key determinant of this style, yet existing approaches to modeling them remain limited. Some operate on stereo mixtures without explicit per-stem FX chain modeling, others fix the number or type of effects per track, and many require differentiable effect implementations or scarce multitrack datasets. We present StemFX, a framework that learns mixing style representations by autoregressively predicting variable-length FX chains on source-separated stems. A Transformer decoder predicts tokenized FX chains autoregressively, while a band-split multi-band CNN encoder with FiLM conditioning captures per-stem spectral structure. To enable large-scale paired training, we extract pseudo-stems from about 105K songs via source separation and augment them using MultiAFx, a toolkit unifying 85 audio effects from 7 Python libraries. Evaluated on mixing style retrieval, StemFX outperforms all baseline models across all tested chain lengths. On paired mixing style transfer, StemFX achieves the best spectral fidelity and the highest listener preference, over 4000 times faster than iterative optimization.

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

44 extracted references · 20 linked inside Pith

  1. [1]

    StemFX: Learning Mixing Style Representations via Autoregressive FX Chain Prediction on Source-Separated Stems

    INTRODUCTION A mixing engineer applies a distinct chain of audio effects to each instrument track: an equalizer to brighten the vo- cals, a compressor to tighten the drums, reverb to place the guitar in a room. The collective choice of effects, their ordering, and their parameter settings across all stems de- fines themixing style. An FX chain can be thou...

    Pith/arXiv arXiv 2026

  2. [2]

    RELA TED WORK 2.1 Automatic Mixing and Style Transfer Automatic mixing systems transfer the sonic character- istics of a reference mix to new material, through ei- ther representation learning or direct parameter predic- tion. FxEncoder [1] disentangles mixing style from con- tent via contrastive learning on multitrack audio, its suc- cessor Fx-Encoder++ ...

  3. [3]

    The frame- work jointly trains a BSFiLM Encoder and a Transformer- based FX Chain Generator end-to-end

    METHOD 3.1 Framework Overview and Problem Formulation Figure 1 illustrates the STEMFX architecture. The frame- work jointly trains a BSFiLM Encoder and a Transformer- based FX Chain Generator end-to-end. The encoder maps both the original and augmented stems into conditioning vectors that summarize the mixing style difference, and the generator autoregres...

    2048

  4. [4]

    Each song is augmented with 1–10 random effects per stem from MultiAFx

    EXPERIMENTS 4.1 Setup Training Data.We apply the Sep-Aug Pipeline to the open FMA dataset [32] using SCNet [22] for source sep- aration, producing about 105K songs with 4 pseudo-stems each. Each song is augmented with 1–10 random effects per stem from MultiAFx. Because FMA contains pre- dominantly mono or stereo mixtures rather than true multi- tracks, ma...

    2048

  5. [5]

    LIMITA TIONS AND FUTURE WORK STEMFX can only predict effects contained in the FX set it was trained on, and supporting additional effects requires retraining on data that includes them. STEMFX also op- erates on the four stems produced by current separation systems, a step beyond estimating mixing style from the stereo mixture alone, though extending it t...

  6. [6]

    CONCLUSION We presented STEMFX, a framework for learning mix- ing style representations through autoregressive FX chain prediction on source-separated stems, with three contribu- tions: an end-to-end autoregressive formulation that out- performs contrastive learning under matched conditions, 86.8% vs. 77.8% Top-1; the BSFiLM Encoder, whose FiLM conditioni...

  7. [7]

    ACKNOWLEDGMENTS The work is supported by grants from Google Asia Pacific, the National Science and Technology Council of Taiwan (NSTC 114-2628-E-002-013-MY3), and the Ministry of Education (MOE) of Taiwan (for Taiwan Centers of Ex- cellence in Artificial Intelligence)

  8. [8]

    Music mixing style transfer: A contrastive learning approach to disentangle audio effects,

    J. Koo, M. A. Martinez-Ramirez, W.-H. Liao, S. Uh- lich, K. Lee, and Y . Mitsufuji, “Music mixing style transfer: A contrastive learning approach to disentangle audio effects,” inProc. IEEE Interna- tional Conference on Acoustics, Speech and Sig- nal Processing (ICASSP), 2023. [Online]. Available: https://arxiv.org/abs/2211.02247

    Pith/arXiv arXiv 2023

  9. [9]

    Diff-MST: Differentiable mixing style transfer,

    S. S. Vanka, C. J. Steinmetz, J.-B. Rolland, J. D. Reiss, and G. Fazekas, “Diff-MST: Differentiable mixing style transfer,” inProc. International Society for Music Information Retrieval Conference (ISMIR),

  10. [10]

    LLM2Fx-Tools: Tool calling for music post- production,

    S. Doh, J. Koo, M. A. Martinez-Ramirez, W. Choi, W.-H. Liao, Q. Wu, J. Nam, and Y . Mitsu- fuji, “LLM2Fx-Tools: Tool calling for music post- production,” inProc. International Conference on Learning Representations (ICLR), 2026. [Online]. Available: https://arxiv.org/abs/2512.01559

    arXiv 2026

  11. [11]

    Au- tomatic multitrack mixing with a differentiable mixing console of neural audio effects,

    C. J. Steinmetz, J. Pons, S. Pascual, and J. Serrà, “Au- tomatic multitrack mixing with a differentiable mixing console of neural audio effects,” inProc. IEEE Inter- national Conference on Acoustics, Speech and Sig- nal Processing (ICASSP), 2021. [Online]. Available: https://arxiv.org/abs/2010.10291

    Pith/arXiv arXiv 2021

  12. [12]

    Style transfer of audio effects with differentiable signal pro- cessing,

    C. J. Steinmetz, N. J. Bryan, and J. D. Reiss, “Style transfer of audio effects with differentiable signal pro- cessing,”Journal of the Audio Engineering Society, vol. 70, no. 9, pp. 708–721, 2022. [Online]. Available: https://arxiv.org/abs/2207.08759

    Pith/arXiv arXiv 2022

  13. [13]

    dasp-pytorch: Differentiable au- dio signal processors in PyTorch,

    C. J. Steinmetz, “dasp-pytorch: Differentiable au- dio signal processors in PyTorch,” https://github.com/ csteinmetz1/dasp-pytorch, 2023

    2023

  14. [14]

    ST- ITO: Controlling audio effects for style transfer with inference-time optimization,

    C. J. Steinmetz, S. Singh, M. Comunità, I. Ib- nyahya, S. Yuan, E. Benetos, and J. D. Reiss, “ST- ITO: Controlling audio effects for style transfer with inference-time optimization,” inProc. International Society for Music Information Retrieval Conference (ISMIR), 2024, best Paper Award. [Online]. Avail- able: https://arxiv.org/abs/2410.21233

    Pith/arXiv arXiv 2024

  15. [15]

    MUSDB18 – a corpus for music sep- aration,

    Z. Rafii, A. Liutkus, F.-R. Stöter, S. I. Mimilakis, and R. Bittner, “MUSDB18 – a corpus for music sep- aration,” https://sigsep.github.io/datasets/musdb.html, 2017

    2017

  16. [16]

    MedleyDB: A multi- track dataset for annotation-intensive MIR research,

    R. M. Bittner, J. Salamon, M. Tierney, M. Mauch, C. Cannam, and J. P. Bello, “MedleyDB: A multi- track dataset for annotation-intensive MIR research,” inProc. International Society for Music Information Retrieval Conference (ISMIR), 2014

    2014

  17. [17]

    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,” inProc. International Society for Music In- formation Retrieval Conference (ISMIR), 2023. [On- line]. Available: https://arxiv.org/abs/2307.15913

    Pith/arXiv arXiv 2023

  18. [18]

    MERT: Acoustic music understanding model with large-scale self-supervised training,

    Y . Li, R. Yuan, G. Zhang, Y . Ma, X. Chen, H. Yin, C. Xiao, C. Lin, A. Ragni, E. Benetos, N. Gyenge, R. Dannenberg, R. Liu, W. Chen, G. Xia, Y . Shi, 7 W. Huang, Z. Wang, Y . Guo, and J. Fu, “MERT: Acoustic music understanding model with large-scale self-supervised training,” inProc. International Con- ference on Learning Representations (ICLR), 2024. [O...

    Pith/arXiv arXiv 2024

  19. [19]

    Multi-band CNN architecture using adaptive frequency filter for acoustic event classification,

    D. Kim, S. Park, D. K. Han, and H. Ko, “Multi-band CNN architecture using adaptive frequency filter for acoustic event classification,”Applied Acoustics, vol. 172, p. 107579, 2021

    2021

  20. [20]

    FiLM: Visual reasoning with a gen- eral conditioning layer,

    E. Perez, F. Strub, H. de Vries, V . Dumoulin, and A. Courville, “FiLM: Visual reasoning with a gen- eral conditioning layer,” inProc. AAAI Conference on Artificial Intelligence, 2018. [Online]. Available: https://arxiv.org/abs/1709.07871

    Pith/arXiv arXiv 2018

  21. [21]

    Fx-Encoder++: Extracting instrument-wise audio effects representa- tions from mixtures,

    Y .-T. Yeh, J. Koo, M. A. Martinez-Ramirez, W.-H. Liao, Y .-H. Yang, and Y . Mitsufuji, “Fx-Encoder++: Extracting instrument-wise audio effects representa- tions from mixtures,” inProc. International Society for Music Information Retrieval Conference (ISMIR),

  22. [22]

    SCNet: Sparse compression network for music source separation,

    W. Tong, J. Zhu, J. Chen, S. Kang, T. Jiang, Y . Li, Z. Wu, and H. Meng, “SCNet: Sparse compression network for music source separation,” inProc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024. [Online]. Avail- able: https://arxiv.org/abs/2401.13276

    Pith/arXiv arXiv 2024

  23. [23]

    Automatic music mixing using a gen- erative model of effect embeddings,

    E. Moliner, M. A. Martinez-Ramirez, J. Koo, W.- H. Liao, K. W. Cheuk, J. Serrà, V . Välimäki, and Y . Mitsufuji, “Automatic music mixing using a gen- erative model of effect embeddings,” inProc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026. [Online]. Avail- able: https://arxiv.org/abs/2511.08040

    arXiv 2026

  24. [24]

    CLAP: Learning audio concepts from natural language supervision,

    B. Elizalde, S. Deshmukh, M. Al Ismail, and H. Wang, “CLAP: Learning audio concepts from natural language supervision,” inProc. IEEE Inter- national Conference on Acoustics, Speech and Sig- nal Processing (ICASSP), 2023. [Online]. Available: https://arxiv.org/abs/2206.04769

    Pith/arXiv arXiv 2023

  25. [25]

    Contrastive learning of general-purpose audio representations,

    A. Saeed, D. Grangier, and N. Zeghidour, “Contrastive learning of general-purpose audio representations,” in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021. [On- line]. Available: https://arxiv.org/abs/2010.10915

    Pith/arXiv arXiv 2021

  26. [26]

    Contrastive learn- ing of musical representations,

    J. Spijkervet and J. A. Burgoyne, “Contrastive learn- ing of musical representations,” inProc. Interna- tional Society for Music Information Retrieval Con- ference (ISMIR), 2021. [Online]. Available: https: //arxiv.org/abs/2103.09410

    Pith/arXiv arXiv 2021

  27. [27]

    Torchaudio: Building blocks for audio and speech processing,

    Y .-Y . Yang, M. Hira, Z. Ni, A. Chourdia, A. Asta- furov, C. Chen, C.-F. Yeh, C. Puhrsch, D. Pollack, D. Genzelet al., “Torchaudio: Building blocks for audio and speech processing,” inProc. IEEE Inter- national Conference on Acoustics, Speech and Sig- nal Processing (ICASSP), 2022. [Online]. Available: https://arxiv.org/abs/2110.15018

    Pith/arXiv arXiv 2022

  28. [28]

    Music2Latent: Consistency autoencoders for latent audio compres- sion,

    M. Pasini, S. Lattner, and G. Fazekas, “Music2Latent: Consistency autoencoders for latent audio compres- sion,” inProc. International Society for Music Infor- mation Retrieval Conference (ISMIR), 2024. [Online]. Available: https://arxiv.org/abs/2408.06500

    Pith/arXiv arXiv 2024

  29. [29]

    HTS-AT: A hierarchical token- semantic audio transformer for sound classification and detection,

    K. Chen, X. Du, B. Zhu, Z. Ma, T. Berg-Kirkpatrick, and S. Dubnov, “HTS-AT: A hierarchical token- semantic audio transformer for sound classification and detection,” inProc. IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP), 2022. [Online]. Available: https://arxiv.org/ abs/2202.00874

    Pith/arXiv arXiv 2022

  30. [30]

    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. [Online]. Available: https://arxiv.org/abs/ 2209.15174

    Pith/arXiv arXiv 1901

  31. [31]

    pyloudnorm: A simple yet flexible loudness meter in Python,

    C. J. Steinmetz and J. D. Reiss, “pyloudnorm: A simple yet flexible loudness meter in Python,” in150th Audio Engineering Society Convention, 2021

    2021

  32. [32]

    Music source separation with band-split RoPE trans- former,

    W.-T. Lu, J.-C. Wang, Q. Kong, and Y .-N. Hung, “Music source separation with band-split RoPE trans- former,” inProc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 481–485. [Online]. Available: https://arxiv. org/abs/2309.02612

    Pith/arXiv arXiv 2024

  33. [33]

    Mel-band Ro- Former for music source separation,

    J.-C. Wang, W.-T. Lu, and M. Won, “Mel-band Ro- Former for music source separation,”arXiv preprint arXiv:2310.01809, 2023. [Online]. Available: https: //arxiv.org/abs/2310.01809

    Pith/arXiv arXiv 2023

  34. [34]

    SoX – sound exchange,

    L. Norskog and C. Bagwell, “SoX – sound exchange,” https://sox.sourceforge.net/, 2005

    2005

  35. [35]

    audiomentations: A Python library for audio data augmentation,

    I. Jordal, “audiomentations: A Python library for audio data augmentation,” https://github.com/iver56/ audiomentations, 2020

    2020

  36. [37]

    SciPy 1.0: Funda- mental algorithms for scientific computing in Python,

    P. Virtanen, R. Gommers, T. E. Oliphant, M. Haber- land, T. Reddy, D. Cournapeau, E. Burovski, P. Peter- son, W. Weckesser, J. Brightet al., “SciPy 1.0: Funda- mental algorithms for scientific computing in Python,” Nature Methods, vol. 17, no. 3, pp. 261–272, 2020

    2020

  37. [38]

    librosa: Audio and mu- sic signal analysis in Python,

    B. McFee, C. Raffel, D. Liang, D. P. Ellis, M. McVicar, E. Battenberg, and O. Nieto, “librosa: Audio and mu- sic signal analysis in Python,” inProc. 14th Python in Science Conference, 2015, pp. 18–24

    2015

  38. [39]

    Array programming with NumPy,

    C. R. Harris, K. J. Millman, S. J. van der Walt, R. Gom- mers, P. Virtanen, D. Cournapeau, E. Wieser, J. Tay- lor, S. Berg, N. J. Smithet al., “Array programming with NumPy,”Nature, vol. 585, no. 7825, pp. 357–362, 2020

    2020

  39. [41]

    FMA: A dataset for music analysis,

    M. Defferrard, K. Benzi, P. Vandergheynst, and X. Bresson, “FMA: A dataset for music analysis,” inProc. International Society for Music Information Retrieval Conference (ISMIR), 2017. [Online]. Avail- able: https://arxiv.org/abs/1612.01840

    Pith/arXiv arXiv 2017

  40. [42]

    Auto- matic music mixing with deep learning and out-of- domain data,

    M. A. Martinez-Ramirez, W.-H. Liao, G. Fabbro, S. Uhlich, C. Nagashima, and Y . Mitsufuji, “Auto- matic music mixing with deep learning and out-of- domain data,” inProc. International Society for Mu- sic Information Retrieval Conference (ISMIR), 2022. [Online]. Available: https://arxiv.org/abs/2208.11428

    Pith/arXiv arXiv 2022

  41. [43]

    Pedalboard: A python library for adding ef- fects to audio,

    P. Sobot, “Pedalboard: A python library for adding ef- fects to audio,” https://github.com/spotify/pedalboard, 2021

    2021

  42. [44]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” inProc. International Conference on Machine Learning (ICML), 2020. 8

    2020

  43. [2024]

    Available: https://arxiv.org/abs/2407

    [Online]. Available: https://arxiv.org/abs/2407. 08889

  44. [2025]

    Available: https://arxiv.org/abs/2507

    [Online]. Available: https://arxiv.org/abs/2507. 02273