Pith. sign in

REVIEW 5 major objections 7 minor 28 references

WildFX: A DAW-Powered Pipeline for In-the-Wild Audio FX Graph Modeling

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A DAW pipeline turns real plugin chains into AI training data.

desk verdict A genuinely useful data-generation tool with honest limitations, but the validation section doesn't yet support the 'state of the art' or the ecological-validity claims. read the letter →

arxiv 2507.10534 v2 pith:LXJ3JMPY submitted 2025-07-14 cs.SD cs.AI

Qihui Yang , Taylor Berg-Kirkpatrick , Julian McAuley , Zachary Novack This is my paper · ORCID
classification cs.SDcs.AI
keywords audioeffectgraphsdatasetgenerationpipelineDockerREAPERVSTpluginssidechainroutingmixinggraphestimationneuralmodeling
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

WildFX claims to be the first end-to-end, Python-driven pipeline that generates multitrack audio datasets whose effect graphs run through a real DAW hosting arbitrary commercial plugins (VST/VST3/LV2/CLAP), including sidechain and splitter routing. The point of the paper is that datasets like this let AI research on audio effect modeling train on the same tools professionals use, rather than simplified differentiable stand-ins. If the pipeline works as described, researchers can generate large, reproducible, professionally processed multitrack corpora and use them for plugin classification, parameter estimation, and mixing-graph inference. The paper demonstrates the claim by using WildFX to build two datasets from Slakh stems and running a blind mixing-graph estimation model on them.

What carries the argument

The load-bearing mechanism is a layer-based topological scheduler over FXChain hyper-nodes. Each project's effect graph is a DAG whose nodes are FXChain objects (ordered plugin runs) and whose edges are gain-labeled audio routings. Kahn's algorithm groups chains into dependency layers; within a layer, WildFX pre-mixes raw audio in Python, renders each chain through REAPER, and passes outputs to the next layer. This construction makes sidechain and splitter routing renderable in headless batch while keeping the space of representable mixing graphs complete.

What would settle it

Render a small set of real professional multitracks through known plugin chains in a commercial DAW, or recover ground-truth chains from actual session files, then train the same graph estimator on WildFX data and evaluate on those real recordings; if accuracy falls to near chance while in-distribution held-out performance stays high, the ecological-validity premise is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that WildFX is a complete containerized data-generation pipeline that controls REAPER through Python, sums raw stems in advance per layer, renders audio through arbitrary commercial plugins via Wine and yabridge on Linux, and records the full heterogeneous effect graph as networkx objects alongside the audio. The key validation is that a blind audio-processing-graph estimation model trained on WildFX-generated projects can recover graph edges and plugin parameters with edge error rates near 0.07 to 0.09, showing that the generated audio carries enough structure for neural graph inference even though node-type recovery remains harder. The authors present this as evidence that ecologically valid, industry-grade plugin-based datasets can be produced at scale and that such data is usable for NeuralAFx tasks.

Load-bearing premise

The generated datasets are treated as ecologically valid stand-ins for real professional mixing workflows, yet the paper never tests whether models trained on WildFX audio transfer to real recordings.

Editorial extensions

If this is right

  • Researchers can generate large-scale multitrack datasets whose audio passes through real commercial plugins, with ground-truth graph and parameter labels included.
  • Neural models trained on WildFX data can be evaluated on industry-standard effects with sidechains and splitters, closing the gap to professional workflows.
  • The pipeline supports principled data augmentation via plugin-driven transformations in professional mixing style, addressing scarcity of high-quality music datasets.
  • The same infrastructure supports plugin classification, parameter estimation, graph inference, grey-box modeling, and musically informed source separation.
  • Layer-based batch processing makes rendering practical at scale, with average processing times of 12 to 15 seconds per project on a 64-core CPU.

Reading between the lines

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

  • The graph-completeness argument implies WildFX could serve as a simulator for studying generalization by varying chain-depth distributions and sidechain probabilities to map how model performance degrades with structural complexity, an experiment the paper does not run.
  • Because parameter presets are discretized and sampled, the generated parameter distribution is a modeling choice; fitting the sampler to real DAW session logs could produce even more ecologically valid data.
  • The blind-estimation results suggest node-type identity is the hardest part of graph recovery, so WildFX could generate targeted probes with spectrally similar plugins to test what features the estimator relies on.
  • Transfer to real professionally mixed audio is the untested bridge; a natural next step is to evaluate models trained on WildFX output against real DAW session recordings with known chains.
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

5 major / 7 minor

Summary. The paper introduces WildFX, a Dockerized pipeline that uses REAPER as a DAW backend to render multitrack audio from programmatically specified audio effect (AFx) graphs, including commercial VST/VST3/LV2/CLAP plugins, sidechain routing, and splitter-based parallel chains. The authors contribute a metadata schema (YAML/JSON plus pickled networkx graphs), a layer-based batch rendering scheduler, and CLI tooling for preset and project generation. They evaluate the pipeline by generating two datasets from Slakh2100 stems and training an existing blind graph-estimation method (Lee et al., ICASSP 2023) on the rendered audio and metadata, reporting graph/parameter estimation metrics. The paper claims to be the first comprehensive end-to-end pipeline for generating such datasets and states that the reproduced model remains state of the art for the task.

Significance. If the pipeline works as described, it would be a useful and reproducible resource for NeuralAFx research: it is containerized, supports multiple plugin formats including commercial Windows plugins via Wine/yabridge, releases code, and offers a structured metadata interface that goes beyond existing differentiable-plugin or simple-Linux-effect packages. The layer-based scheduling and pre-mixing in Python are sensible engineering choices that address real scalability concerns. However, the paper's central validity claims are not adequately supported by the current experiments: there is no verification that headless rendering actually matches the metadata, no transfer test to real-world audio, no baseline comparisons on the generated data, no error bars, and the state-of-the-art claim is contradicted by the paper's own admission that results fall short of the original Lee et al. work. The strengths are the engineering contribution and code release, but the scientific validation needs substantial strengthening.

major comments (5)
  1. [Section 3.2, Section 4.2] The headless rendering path provides no built-in error reporting or inspection tools (Section 3.2), and correctness was established only by manual GUI validation before deployment. In batch, a commercial VST failing to load under Wine/yabridge, a preset not recalled, or a sidechain send not connected would produce audio that does not match the ground-truth graph, yet the pipeline would record it as a valid rendering. The Section 4 experiments train and evaluate on the same generated audio/metadata pairs, so even a systematically corrupted renderer could yield learnable but wrong structure. I recommend adding automated rendering-fidelity checks (e.g., rendering a probe signal through a known graph and comparing against expected output, checking sidechain tap points, reporting plugin load/skip rates) and reporting these numbers for both datasets.
  2. [Section 4.2, Table 3, Section 5] The experimental validation is too weak to support the claim that WildFX generates realistic data suitable for training and evaluating neural audio systems. Table 3 reports single runs without error bars or multiple seeds, the reproduced model underperforms the original Lee et al. results, no alternative graph-estimation baseline is evaluated on WildFX data, and no transfer test is performed on real audio recordings. The statement in Section 5 that 'the reproduced model achieves competitive results and remains the state of the art for this task' is therefore unsupported, and it sits in tension with the same section's admission that 'our results fall short of those reported in the original work.' Please add baselines, repeated-seed statistics, and an external validation (e.g., real recordings processed through known graph configurations) before making generalization claims.
  3. [Section 3.4.2, Section 3.3.4] The semantics of sidechain routing are not fully specified. Section 3.4.2 states that sidechain sources and consumers reside in the same layer, but it does not specify whether the control signal is tapped from the source chain's input, from a point within the chain, or from the chain's fully processed output; the metadata schema in Section 3.3.3 likewise has no field for a tap point or for pre/post-FX routing. Without this specification, two renderings of the same metadata could differ, and ground-truth graphs for downstream models are ambiguous. Please define the tap semantics and add an explicit field if needed.
  4. [Section 3.5] The claim that 'our data structure can actually represent all possible mixing graph structures' is not established. The restrictions in Section 3.3.4 exclude sidechains originating from splitter outputs and allow at most one sidechain-enabled plugin per chain; Section 3.5 sketches transformations (splitting chains, inserting empty pseudo-layers) but does not prove that these transformations preserve the audio signal and graph semantics for arbitrary DAGs. A formal argument or a constructive enumeration for all graph classes up to a given size would be needed to support the completeness claim.
  5. [Section 4.1] The dataset configurations (sidechain probability 0.2, splitter probability 0.1, chain-depth distributions, parameter discretizations) are hand-chosen, and no sensitivity analysis or comparison with real mixing practices is provided. The experiments only test the pipeline on these generated configurations, so the claim that WildFX produces 'ecologically valid' data for bridging AI research and practical DSP is not directly supported. I recommend reporting how results vary across configurations and, if possible, validating the graph distributions against a small corpus of real session data or expert annotation.
minor comments (7)
  1. [Title] The title in the PDF renders as 'DA W-Powered' with a spurious space; the Abstract correctly uses 'DAW-powered'.
  2. [Section 3.3.4] There is a typo: 'at less one input soruce' should be 'at least one input source'.
  3. [Section 3.4.1] Minor typos: 'dicretize' should be 'discretize' and 'underperforment' (in the Introduction) should be 'underperforming'.
  4. [Section 4.2] The sentence 'By leveraging two available model configurations. autoencoding and prototype decoding' has a misplaced period and a non-capitalized continuation; it should read as one sentence.
  5. [Section 3.3.2] The JSON example in Figure 2 has inconsistent indentation, making the nested structure harder to read; please reformat it consistently.
  6. [Section 3.4.2, Figure 4 caption] The caption contains 'output requried', which should be 'output required'; also the grammar of the surrounding sentence could be tightened.
  7. [References] Several references (e.g., [20]) lack full bibliographic details such as volume, issue, or page numbers; please complete them.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: WildFX's experiments are a standard synthetic-benchmark evaluation against an external method and corpus, not a prediction that reduces to its own inputs.

full rationale

The paper's central claim is that the WildFX pipeline can generate multitrack audio datasets whose effect graphs are encoded in metadata and rendered by a DAW. The validity experiments then train the external method of Lee et al. [12] on held-out WildFX-generated projects and test on separate validation projects. This is a conventional supervised evaluation: the target graph is the metadata used for rendering, but the model must infer it from audio alone, and no parameter of the pipeline is fitted to make the evaluation succeed. The audio inputs come from the external Slakh2100 corpus, the baseline method is external, and the reported results are actually worse than the original method's, which is inconsistent with any claim that the evaluation is forced by construction. The authors' self-citations [18,19] appear only as examples of generative music systems in the introduction and are not load-bearing for the WildFX derivation. The weak assumption of ecological validity, and the unmonitored headless rendering noted in Section 3.2, are correctness and robustness risks rather than circularity: they concern whether rendered audio matches metadata, not whether a result is equivalent to its inputs by definition, fit, or self-citation chain. No step in the derivation reduces to its own inputs.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests mainly on domain assumptions about rendering fidelity and ecological validity. The many user-chosen dataset configuration values (probabilities, ranges, distributions) are free parameters that shape the data, though they are not fitted to any target result. No new physical or conceptual entities are introduced.

free parameters (7)
  • sidechain_probability = 0.2
    User-chosen probability in topology synthesis (Section 4.1); directly controls graph complexity and dataset difficulty.
  • splitter_probability = 0.1
    User-chosen probability in topology synthesis (Section 4.1); controls how often chains branch into parallel paths.
  • chain_depth_distribution = shallow [0.1, 0.7, 0.2]; deep [0.1, 0.3, 0.4, 0.2]
    User-defined categorical distribution for per-chain depth (Section 3.4.1, Table 2); arbitrary choices affect structure.
  • chain_count_range = shallow [3, 5]; deep [3, 10]
    User-specified min/max number of FX chains per project (Table 2); arbitrary configuration.
  • stem_count_range = shallow [1, 2]; deep [1, 4]
    User-specified min/max number of input stems (Table 2); arbitrary configuration.
  • plugin_parameter_discretizations = e.g., 3 Band EQ Low/Mid/High in [0.0, 0.01, ..., 1.0]
    Hand-written valid_params ranges and preset lists in JSON files (Section 3.3.2); define the full parameter sampling space.
  • variable_density_flag = enabled
    Flag that randomizes graph density, depth distribution, and probabilities per project (Section 3.4.1); the randomization distribution is not specified.
assumptions (5)
  • domain assumption Summing raw audio waveforms in Python before plugin processing reproduces DAW bus mixing for nodes with multiple input chains.
    Invoked in Section 3.4.2 (layer-based processing). Assumes no differences from DAW sample-accurate delay compensation or channel routing; no audio-level comparison is provided.
  • domain assumption REAPER running inside Docker with Wine and yabridge renders Windows VST plugins with fidelity equivalent to a native DAW host.
    Stated in Section 3.1 as the deployment architecture; no empirical audio comparison to native rendering is given, yet dataset validity depends on it.
  • domain assumption Every possible mixing graph can be represented under the splitter and sidechain restrictions by using empty chains and layer separation.
    Argued in prose in Section 3.5; the construction is informal and not proved, and it relies on a pseudo-layer equivalence for sidechain sources.
  • domain assumption Slakh2100 stems are representative input audio for professional-style mixing research.
    Section 4.1 builds both datasets from Slakh2100 without justifying that its stems cover real-world recording conditions.
  • standard math Kahn's topological sort provides a correct dependency order for the audio-effect DAG layers.
    Used in Section 3.4.2 for layer-based scheduling; standard algorithm, accepted as sound.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WildFX: A DAW-Powered Pipeline for In-the-Wild Audio FX Graph Modeling." pith.science (2026). https://pith.science/paper/LXJ3JMPY

@misc{pith2026250710534,
  author       = {Pith},
  title        = {Pith review of: WildFX: A DAW-Powered Pipeline for In-the-Wild Audio FX Graph Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LXJ3JMPY}},
  note         = {Machine review of arXiv:2507.10534}
}
read the original abstract

Despite rapid progress in end-to-end AI music generation, AI-driven modeling of professional Digital Signal Processing (DSP) workflows remains challenging. In particular, while there is growing interest in neural black-box modeling of audio effect graphs (e.g. reverb, compression, equalization), AI-based approaches struggle to replicate the nuanced signal flow and parameter interactions used in professional workflows. Existing differentiable plugin approaches often diverge from real-world tools, exhibiting inferior performance relative to simplified neural controllers under equivalent computational constraints. We introduce WildFX, a pipeline containerized with Docker for generating multi-track audio mixing datasets with rich effect graphs, powered by a professional Digital Audio Workstation (DAW) backend. WildFX supports seamless integration of cross-platform commercial plugins or any plugins in the wild, in VST/VST3/LV2/CLAP formats, enabling structural complexity (e.g., sidechains, crossovers) and achieving efficient parallelized processing. A minimalist metadata interface simplifies project/plugin configuration. Experiments demonstrate the pipeline's validity through blind estimation of mixing graphs, plugin/gain parameters, and its ability to bridge AI research with practical DSP demands. The code is available on: https://github.com/IsaacYQH/WildFX.

Figures

Figures reproduced from arXiv: 2507.10534 by the authors.

Figure 1
Figure 1. Deployment Environment of WildFX Docker is a platform for developing, shipping, and running applications in lightweight, isolated containers. Using the provided Dockerfile, users can easily build a self-contained environ￾ment for the WildFX data generation pipeline, eliminating the need for manual dependency management—a critical feature for reproducible open-source research. To further simplify setup, we include a … view at source ↗
Figure 2
Figure 2. Mixing Graph with the Provided Sample Note that in the provided project metadata ex￾ample, the params field is empty, as the plugin uses the preset indexed by 2 from its correspond￾ing JSON preset file. All plugin parameters are sampled from the discretized set defined in valid_params. A null value in the JSON file indicates use of the plugin’s default setting. Additionally, the values in next_chains represent gain … view at source ↗
Figure 3
Figure 3. Overview of the WildFX workflow. Users begin by specifying plugin names to generate corresponding presets. These presets are then used to synthesize project metadata, which defines audio effect graphs. The metadata is rendered using REAPER in a headless environment, after which the output can be exported in the desired format (e.g., waveform, compressed HDF5, YAML graph specification or pickled networkx objects). Ex… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Layer-based Multi-Project Batch Processing: the 4 layers across 3 projects are determined [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 22 canonical work pages

  1. [1]

    MusicLM: Generating music from text

    Andrea Agostinelli, Timo I Denk, Zalán Borsos, Jesse Engel, Mauro Verzetti, Antoine Caillon, Qingqing Huang, Aren Jansen, Adam Roberts, Marco Tagliasacchi, et al. MusicLM: Generating music from text. arXiv:2301.11325, 2023

    arXiv 2023

  2. [2]

    Guitar Effects Recognition and Parameter Estimation with Convolutional Neural Networks

    Marco Comunità, Dan Stowell, and Joshua D Reiss. Guitar effects recognition and parameter estimation with convolutional neural networks. arXiv preprint arXiv:2012.03216, 2020

    work page Pith review arXiv 2012

  3. [3]

    Differentiable Black-box and Gray-box Modeling of Nonlinear Audio Effects

    Marco Comunità, Christian J. Steinmetz, and Joshua D. Reiss. Differentiable Black-box and Gray-box Modeling of Nonlinear Audio Effects, February 2025. arXiv:2502.14405 [cs]

    work page Pith review arXiv 2025

  4. [4]

    NablAFx: A Framework for Differentiable Black-box and Gray-box Modeling of Audio Effects

    Marco Comunità, Christian J. Steinmetz, and Joshua D. Reiss. NablAFx: A Framework for Differentiable Black-box and Gray-box Modeling of Audio Effects, February 2025. arXiv:2502.11668 [cs]

    work page Pith review arXiv 2025

  5. [5]

    SingSong: Generating musical accompaniments from singing

    Chris Donahue, Antoine Caillon, Adam Roberts, Ethan Manilow, Philippe Esling, Andrea Agostinelli, Mauro Verzetti, et al. SingSong: Generating musical accompaniments from singing. arXiv:2301.12662, 2023

    arXiv 2023

  6. [6]

    DDSP: Differentiable digital signal processing

    Jesse Engel, Lamtharn Hantrakul, Chenjie Gu, and Adam Roberts. DDSP: Differentiable digital signal processing. In ICLR, 2020

    work page 2020

  7. [7]

    Stable audio open

    Zach Evans, Julian D Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons. Stable audio open. arXiv:2407.14358, 2024

    arXiv 2024

  8. [8]

    Riffusion: Stable diffusion for real-time music generation, 2022

    Seth Forsgren and Hayk Martiros. Riffusion: Stable diffusion for real-time music generation, 2022

    work page 2022

Show all 28 references
  1. [9]

    Signaltrain: Profiling audio compressors with deep neural networks

    Scott H Hawley, Benjamin Colburn, and Stylianos I Mimilakis. Signaltrain: Profiling audio compressors with deep neural networks. arXiv preprint arXiv:1905.11928, 2019

    1905 arXiv

  2. [10]

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

    Junghyun Koo, Marco A Martínez-Ramírez, Wei-Hsiang Liao, Stefan Uhlich, Kyogu Lee, and Yuki Mitsufuji. Music mixing style transfer: A contrastive learning approach to disentangle audio effects. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal P...

    2023

  3. [11]

    Grafx: an open-source library for audio processing graphs in pytorch

    Sungho Lee, Marco Martínez-Ramírez, Wei-Hsiang Liao, Stefan Uhlich, Giorgio Fabbro, Kyogu Lee, and Yuki Mitsufuji. Grafx: an open-source library for audio processing graphs in pytorch. arXiv preprint arXiv:2408.03204, 2024

    2024 arXiv

  4. [12]

    Blind Estimation of Audio Processing Graph

    Sungho Lee, Jaehyun Park, Seungryeol Paik, and Kyogu Lee. Blind Estimation of Audio Processing Graph. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5, June 2023. ISSN: 2379-190X

    2023

  5. [13]

    Decoupled weight decay regularization, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019

    2019

  6. [14]

    Cutting music source separation some slakh: A dataset to study the impact of training data quality and quantity, 2019

    Ethan Manilow, Gordon Wichern, Prem Seetharaman, and Jonathan Le Roux. Cutting music source separation some slakh: A dataset to study the impact of training data quality and quantity, 2019

    2019

  7. [15]

    Deep learning for black-box modeling of audio effects

    Marco A Martínez Ramírez, Emmanouil Benetos, and Joshua D Reiss. Deep learning for black-box modeling of audio effects. Applied Sciences, 10(2):638, 2020

    2020

  8. [16]

    Modula- tion extraction for lfo-driven audio effects

    Christopher Mitcheltree, Christian J Steinmetz, Marco Comunità, and Joshua D Reiss. Modula- tion extraction for lfo-driven audio effects. arXiv preprint arXiv:2305.13262, 2023

    2023 arXiv

  9. [17]

    Instrumentgen: Generating sample-based musical instruments from text

    Shahan Nercessian and Johannes Imort. Instrumentgen: Generating sample-based musical instruments from text. arXiv preprint arXiv:2311.04339, 2023

    2023 arXiv

  10. [18]

    Fast text-to-audio generation with adversarial post-training

    Zachary Novack, Zach Evans, Zack Zukowski, Josiah Taylor, CJ Carr, Julian Parker, Adnan Al-Sinan, Gian Marco Iodice, Julian McAuley, Taylor Berg-Kirkpatrick, and Jordi Pons. Fast text-to-audio generation with adversarial post-training. arXiv:2505.08175, 2025. 10

    2025 arXiv

  11. [19]

    Zachary Novack, Ge Zhu, Jonah Casebeer, Julian McAuley, Taylor Berg-Kirkpatrick, and Nicholas J. Bryan. Presto! distilling steps and layers for accelerating music generation. In ICLR, 2025

    2025

  12. [20]

    Style transfer of audio effects with differentiable signal processing

    Christian J Steinmetz, Nicholas J Bryan, and Joshua D Reiss. Style transfer of audio effects with differentiable signal processing. Journal of the Audio Engineering Society (JAES), 2022

    2022

  13. [21]

    Automatic multitrack mixing with a differentiable mixing console of neural audio effects

    Christian J Steinmetz, Jordi Pons, Santiago Pascual, and Joan Serrà. Automatic multitrack mixing with a differentiable mixing console of neural audio effects. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 71–75. IEEE, 2021

    2021

  14. [22]

    Efficient neural networks for real-time analog audio effect modeling

    Christian J Steinmetz and Joshua D Reiss. Efficient neural networks for real-time analog audio effect modeling. arXiv preprint arXiv:2102.06200, 2021

    2021 arXiv

  15. [23]

    Steinmetz, Shubhr Singh, Marco Comunità, Ilias Ibnyahya, Shanxin Yuan, Em- manouil Benetos, and Joshua D

    Christian J. Steinmetz, Shubhr Singh, Marco Comunità, Ilias Ibnyahya, Shanxin Yuan, Em- manouil Benetos, and Joshua D. Reiss. ST-ITO: Controlling Audio Effects for Style Transfer with Inference-Time Optimization, October 2024. arXiv:2410.21233 [cs]

    2024 arXiv

  16. [24]

    High-fidelity noise reduction with differentiable signal processing

    Christian J Steinmetz, Thomas Walther, and Joshua D Reiss. High-fidelity noise reduction with differentiable signal processing. arXiv preprint arXiv:2310.11364, 2023

    2023 arXiv

  17. [25]

    Diffmoog: a differentiable modular synthesizer for sound matching

    Noy Uzrad, Oren Barkan, Almog Elharar, Shlomi Shvartzman, Moshe Laufer, Lior Wolf, and Noam Koenigstein. Diffmoog: a differentiable modular synthesizer for sound matching. arXiv preprint arXiv:2401.12570, 2024

    2024 arXiv

  18. [26]

    Pyneuralfx: A python package for neural audio effect modeling

    Yen-Tung Yeh, Wen-Yi Hsiao, and Yi-Hsuan Yang. Pyneuralfx: A python package for neural audio effect modeling. arXiv preprint arXiv:2408.06053, 2024

    2024 arXiv

  19. [27]

    Singing voice synthesis using differentiable lpc and glottal- flow-inspired wavetables

    Chin-Yun Yu and György Fazekas. Singing voice synthesis using differentiable lpc and glottal- flow-inspired wavetables. arXiv preprint arXiv:2306.17252, 2023

    2023 arXiv

  20. [28]

    Yue: Scaling open foundation models for long-form music generation

    Ruibin Yuan, Hanfeng Lin, Shuyue Guo, Ge Zhang, Jiahao Pan, Yongyi Zang, Haohe Liu, Yiming Liang, Wenye Ma, Xingjian Du, et al. Yue: Scaling open foundation models for long-form music generation. arXiv:2503.08638, 2025. 11

    2025

Pith tools

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