REVIEW 4 major objections 6 minor 33 references
A Statistics-Driven Differentiable Approach for Sound Texture Synthesis and Analysis
T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read The paper claims that texture sounds—rain, fire, wind—can be compared, trained, and validated through a small set of auditory summary statistics, realized as a differentiable, time-invariant, noise-robust loss called TexStat, with a…
desk verdict A genuinely useful differentiable texture loss with honest limitations, but the perceptual-validity claim outruns the evidence. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the summary-statistics feature vector extracted from the auditory periphery model: an ERB cochlear filterbank $F=\{f_j\}$ splits the signal into subbands, the Hilbert transform yields each subband's amplitude envelope $e_j$, and a logarithmic modulation filterbank $G=\{g_k\}$ decomposes each envelope into modulation subbands $m_{j,k}$. From these, five statistic sets $S_1,\dots,S_5$ are computed—normalized envelope moments, envelope correlations, modulation-energy proportions, within-envelope modulation correlations, and across-envelope modulation correlations—and TexStat is the weighted MSE between two signals' statistic vectors. A practical design choice does much of the work: the same filterbank is shared between TexStat and the TexEnv synthesizer, so gradients flow coherently from loss to synthesis parameters, while configurability of filterbank size, moment count, and weights keeps the statistic count below the sample count of the analysis window, which the paper identifies as a needed control for training.
What would settle it
A listener experiment would settle it: take a texture class TexStat is claimed to capture (rain or fire), construct pairs of sounds whose TexStat statistic vectors are nearly identical but whose sample-level structure differs, and run an ABX test in which listeners pick which reference each candidate matches; if listeners can reliably tell the candidates apart, the claim that matching statistics means matching texture identity fails. A cheaper probe is to compute TexStat on a pitched texture such as wind chimes at two clearly different pitches and show the loss stays near zero even though listeners trivially hear different sounds.
Extended reading notes
Core claim
The central discovery the paper is trying to establish is that a fixed vector of auditory summary statistics—not the waveform, not a learned embedding—can carry enough of a texture's perceptual identity to drive both synthesis and evaluation. The statistics are computed by splitting the signal into ERB auditory subbands, taking each subband's Hilbert amplitude envelope, decomposing the envelopes through a logarithmic modulation filterbank, and then assembling five sets: normalized moments of each envelope, Pearson correlations between envelopes, per-band modulation energy ratios, within-envelope modulation correlations, and across-envelope modulation correlations. The TexStat distance between two signals is the weighted mean squared error over these statistic vectors, $L_{\alpha,\beta}(x,y)=\sum_{j=1}^{5}\beta_j\,\mathrm{MSE}(S_j(x),S_j(y))$, with weights chosen so no single statistic set dominates. The paper argues this quantity is time-invariant for continuous signals (and effectively so for noisy discrete signals), degrades gracefully under additive noise, and is flexible in the sense that many temporally different signals map to nearby statistic vectors. Empirically, the paper shows the statistic vector beats VGGish embeddings at texture classification and that TexDSP models trained with TexStat as the sole loss produce indefinitely long, recognizable textures of the target type.
Load-bearing premise
The load-bearing bet is that a sound's identity as a texture is carried entirely by a fixed set of envelope and modulation statistics, a premise the paper inherits from McDermott and Simoncelli and concedes fails for pitched, rhythmic, and reverberant sounds—so the central claim covers only a subset of textures.
Editorial extensions
If this is right
- A TexDSP model trained with TexStat as its only loss produces indefinitely long texture sounds that stay recognizably within the target type without reproducing the input waveform, which the paper demonstrates across bubbles, fire, keyboard, rain, river, shards, waterfall, and wind.
- The same summary-statistics vector works as an evaluation feature: it beat VGGish embeddings in texture classification on all three MicroTex subsets (accuracy 0.94–1.0 versus 0.71–0.98) and is proposed as the embedding inside Fréchet Audio Distance.
- TexStat is affordable enough for training: about 93 ms forward and 155 ms backward per batch of 32 one-and-a-half-second signals on an RTX 4090, roughly twenty times slower than MSS but with comparable memory use.
- For pitched, rhythmic, or reverberant textures, where the statistics are insufficient, the paper positions TexStat as a regularization loss or as a validation metric rather than the sole objective.
- A shared filterbank between TexStat and TexEnv lets the loss and the synthesizer be optimized coherently, and the paper shows some TexDSP models beat the reconstruction-focused NoiseBandNet baseline even on metrics that favor exact reconstruction.
Reading between the lines
- Beyond the paper: TexStat's interpretable statistic vector could double as a diagnostic—the specific statistic set that diverges between a synthesis and its target would point to whether the failure is spectral, cross-channel, or in the modulation (rhythmic) domain, a debugging use the authors do not develop.
- Beyond the paper: because the statistics are time-invariant, the same machinery could compare textures of different durations or enable alignment-free comparisons for environmental audio monitoring (e.g., classifying recordings by texture type without segmentation), which the paper does not test.
- Beyond the paper: the stated need for subjective validation suggests a concrete next experiment—an ABX listening study on the MicroTex classes comparing TexStat's ranking of candidates against human rankings—which would either confirm or bound the perceptual claim.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces TexStat, a differentiable loss function defined as a weighted sum of mean squared errors between five sets of auditory summary statistics (moments of subband amplitude envelopes, envelope correlations, modulation band energies, and modulation correlations), inspired by the McDermott–Simoncelli framework. It also presents TexEnv, a lightweight differentiable synthesizer that multiplies a fixed filtered-noise seed by amplitude envelopes, and TexDSP, a DDSP-style generative model that combines TexEnv with TexStat as the sole training loss. The authors report experiments on a curated MicroTex dataset covering time-shift and noise robustness (§5.1), computational benchmarks (§5.2), classification using TexStat features versus VGGish embeddings (§5.3), resynthesis exploration (§5.4), and model evaluation with FAD and loss metrics (§5.5). The central claim is that TexStat is perceptually meaningful, time-invariant, and noise-robust, making it effective both as a training loss and a validation metric.
Significance. If the perceptual claim were established, the contribution would be valuable: TexStat offers a fully differentiable, interpretable, and open-source loss for texture sounds, with a computational cost roughly two orders of magnitude higher than MSS but manageable on modern GPUs. The paper also provides a lightweight synthesizer (TexEnv) and a complete generative pipeline (TexDSP), along with transparent discussion of limitations inherited from the summary-statistics approach. However, the evidence presented does not support the strongest assertion that TexStat is perceptually meaningful: there is no human listening test, and the primary metric used in §5.5 (FAD on TexStat statistics) is partly circular for a model trained with TexStat. Therefore, the significance at the level claimed in the abstract is not yet established, though the underlying tools and data are potentially useful to the community.
major comments (4)
- [§5.5, Table 5, §2.5] The central claim that TexStat is a perceptually grounded validation metric is not independently supported. In §5.5, the 'Ours' FAD is computed on the TexStat summary statistics, and the TexDSP model was trained to minimize the TexStat loss; low FAD-Ours scores for TexDSP are therefore partly a tautology. The paper's own admission in §5.5 that 'a subjective evaluation would be necessary—an analysis that was beyond the scope of this work' explicitly identifies the missing link. Please provide a human listening test comparing TexStat rankings with perceptual judgments, or alternatively evaluate with a pretrained embedding that was not used in training and show agreement between TexStat and human similarity ratings. Without such evidence, the 'perceptually meaningful' label is an assertion rather than a demonstrated property.
- [§5.3, Table 4] The conclusion that TexStat summary statistics are 'strictly more informative' than VGGish embeddings is overstated. The comparison is based on a single train/test run on three small hand-curated classes with no cross-validation, error bars, or statistical significance testing. For the Freesound class, the accuracy difference (0.99 vs 0.98) is within typical random variation. Please report results over multiple random seeds or cross-validation folds, include standard deviations, and perform a significance test (e.g., paired bootstrap). Also clarify the sizes of the datasets, the number of segments, and whether classes are balanced, to allow the reader to gauge the strength of the claim.
- [§5.1, Table 2] The time-invariance and noise-robustness results are presented as general properties of TexStat, but they are computed only on the Freesound subset of MicroTex; the exclusion of the other two classes is reasonable given the stated difficulty of applying transformations, yet the conclusions in the abstract and §5.1 are not qualified. Moreover, the noise-add row shows large standard deviations relative to the means (e.g., 2.08 ± 1.99 at 10% noise), so the claim of a 'sublinear increase' in loss is not statistically substantiated. Please either restrict the claims to the tested subset or provide a statistical analysis (e.g., confidence intervals or a trend test) to support the sublinearity statement.
- [§2.4, abstract, conclusions] The paper acknowledges in §2.4 that TexStat inherits McDermott and Simoncelli's failure for pitched, rhythmic, and reverberant sounds, which narrows the domain of the central claim. However, the abstract and conclusions state that TexStat is effective for 'texture sounds' without this caveat. Please qualify the claims to aperiodic, non-pitched, non-reverberant textures, or explicitly state in the abstract that the method is intended for such textures only. This is important because the evaluation in §5.5 is conducted on a selection of sounds that are predominantly in this favorable class, so the results do not support unqualified generalization.
minor comments (6)
- [§2.2, Eq. (1)-(5)] In the definition of S1(s), the expression 'concat(α1S1,1, . . . , αN S1,L)' uses inconsistent indices: the weight should be subscripted by the moment index l, not N, and the last vector should be S1,L, so write 'concat(α1 S1,1, . . . , αL S1,L)'.
- [§2.3] The phrase 'Schwarz class' should be spelled 'Schwartz class', and the notation \mathcal{S}(\mathbb{R}) would avoid confusion with the signal s. Also, the parenthetical 's ∈ C(R)' is not standard; the intended meaning is probably continuous functions, but clarify.
- [§5.1, Table 2] The definition of 'noise percentage' is vague: specify whether it is the ratio of noise RMS to signal RMS or peak amplitude, and report the corresponding SNR in dB to make the experiment reproducible.
- [§5.3] Provide details on the classifier training: the proportion of data used for training vs testing, whether stratification was applied, the number of segments per class, the optimizer and number of epochs, and whether results are averaged over multiple runs.
- [§5.5, Table 5] The footnotes '(1) Energy bands were imposed post-resynthesis' and '(2) A loudness tracker was added post-resynthesis' are not referenced in the table or the text. Please specify which textures these modifications apply to and explain how they affect the comparison between TexDSP and NoiseBandNet.
- [References] The reference [8] lists 'D.F. Rosenthal, H.G. Okuno, H. Okuno, and D. Rosenthal' with duplicate author entries; the correct list should be 'D.F. Rosenthal and H.G. Okuno' (editors).
Circularity Check
Partial circularity: FAD computed on TexStat's own summary statistics is used to validate models trained with the TexStat loss, so the 'Ours' FAD column partly restates the training objective; external VGGish-FAD, MSS, and property tests keep the rest independent.
-
fitted input called prediction
[Section 5.5 (Validation Method and Training Details), with Definition 1 in Section 2.2]
"We then segmented the original and resynthesized signals, and measured Fréchet Audio Distance (FAD) using both VGGish embeddings and our custom summary statistics, along with frame-level TexStat and MSS losses."
TexStat (Definition 1) is the weighted MSE between the summary-statistics vectors S1..S5 of output and target. The TexDSP models in Section 5.5 are trained 'with TexStat as the sole loss function to guide the learning process.' The 'FAD ... using ... our custom summary statistics' column then measures the distributional distance of exactly those same S-vectors. A model optimized to match per-frame S-vectors will, by construction, tend to score well on FAD computed from those S-vectors, so this column is a near restatement of the training objective rather than independent evidence that TexStat is a perceptually valid metric. The paper mitigates this by also reporting VGGish-FAD and MSS, which are external to TexStat.
full rationale
The paper's central derivation of TexStat as a weighted MSE over McDermott-Simoncelli-style summary statistics is not itself circular: it openly builds on the external framework of references [4-6] and inherits its known limitations (Section 2.4). The time-invariance and noise-robustness claims are tested against explicit transformations (Section 5.1), and the classification experiment (Section 5.3) uses a downstream MLP against VGGish, both independent of the loss. The main circular step is the validation protocol in Section 5.5: models trained with TexStat as the sole loss are evaluated with FAD computed on 'our custom summary statistics' and with frame-level TexStat itself. Because TexStat is defined as MSE between those same summary statistics, the 'Ours' FAD results are partly forced by construction. However, the paper also reports VGGish-based FAD and MSS losses as external benchmarks, and its conclusions acknowledge that a subjective perceptual evaluation was beyond scope: 'a subjective evaluation would be necessary—an analysis that was beyond the scope of this work.' Thus the circularity is real but partial: the strong claim that TexStat is 'perceptually meaningful' remains an assertion supported mainly by the self-same statistics, while the tool's other properties rest on independent tests. Overall score 4 reflects partial, not total, circularity.
Assumptions & free parameters
free parameters (3)
- alpha (moment weights) =
default values in repository
- beta (statistics set weights) =
default values in repository
- filterbank sizes NF and NG =
varied per experiment
assumptions (3)
- domain assumption Summary statistics of the auditory periphery capture perceptually relevant similarities for stochastic texture sounds.
- domain assumption Time-shift invariance holds for sufficiently regular continuous signals; discrete finite signals may introduce clicks.
- ad hoc to paper The fixed seed (precomputed filtered noise) is a source of deterministic randomness that preserves texture statistics when multiplied by amplitude envelopes.
Cite this review
Pith. "Pith review of A Statistics-Driven Differentiable Approach for Sound Texture Synthesis and Analysis." pith.science (2026). https://pith.science/paper/OG252UOI
@misc{pith2026250604073,
author = {Pith},
title = {Pith review of: A Statistics-Driven Differentiable Approach for Sound Texture Synthesis and Analysis},
year = {2026},
howpublished = {\url{https://pith.science/paper/OG252UOI}},
note = {Machine review of arXiv:2506.04073}
}
read the original abstract
In this work, we introduce TexStat, a novel loss function specifically designed for the analysis and synthesis of texture sounds characterized by stochastic structure and perceptual stationarity. Drawing inspiration from the statistical and perceptual framework of McDermott and Simoncelli, TexStat identifies similarities between signals belonging to the same texture category without relying on temporal structure. We also propose using TexStat as a validation metric alongside Frechet Audio Distances (FAD) to evaluate texture sound synthesis models. In addition to TexStat, we present TexEnv, an efficient, lightweight and differentiable texture sound synthesizer that generates audio by imposing amplitude envelopes on filtered noise. We further integrate these components into TexDSP, a DDSP-inspired generative model tailored for texture sounds. Through extensive experiments across various texture sound types, we demonstrate that TexStat is perceptually meaningful, time-invariant, and robust to noise, features that make it effective both as a loss function for generative tasks and as a validation metric. All tools and code are provided as open-source contributions and our PyTorch implementations are efficient, differentiable, and highly configurable, enabling its use in both generative tasks and as a perceptually grounded evaluation metric.
Figures
Reference graph
Works this paper leans on
-
[1]
A Statistics-Driven Differentiable Approach for Sound Texture Synthesis and Analysis
INTRODUCTION Defining audio textures is a complex problem that has been ex- plored by various authors. The concept originated as an analogy to visual textures. In [1], Julesz, one of the pioneers in this field, pro- posed the so-called "Julesz conjecture," suggesting that humans cannot distinguish between visual textures with similar second- order statist...
work page Pith review arXiv 2025
-
[2]
TEXSTAT: A LOSS FUNCTION SPECIFICALLY TAILORED FOR TEXTURE SOUNDS In this section, we discuss some desirable properties of a texture sound loss function and then introduce TexStat, a loss function specifically designed for texture sounds that fulfills, to some ex- tent, the properties outlined here. 2.1. What Should a Texture Loss Be? In the context of de...
work page 2025
-
[3]
source of deterministic randomness
TEXENV: A DIFFERENTIABLE SIGNAL PROCESSOR TAILORED FOR TEXTURE SOUNDS The foundations of TexStat are based on the idea that summary statistics of amplitude envelopes derived from a low-size filterbank subband decomposition are sufficient to compare certain types of texture sounds. Implicit in this concept is the fact that directly im- posing amplitude env...
work page 2025
-
[4]
TEXDSP: A DDSP-BASED ARCHITECTURE TAILORED FOR TEXTURE SOUNDS In this section we introduce TexDSP, a relatively simple texture sound generative model based on DDSP [21] that showcases the capabilities of TexStat and TexEnv. The original DDSP model requires an encoder, decoder, sig- nal processor, and loss function, with each component designed to generate...
-
[5]
EXPERIMENTS AND RESULTS In this section, we briefly explain a series of experiments con- ducted to provide proof of concept for the models proposed in this work. For these experiments, we hand-curated MicroTex1, a dataset made from a selection of texture sounds from the follow- ing sources: the BOReilly dataset, containing textures made using 1MicroTex Hu...
-
[6]
CONCLUSIONS This paper introduced a novel framework for advancing the analy- sis and synthesis of texture sounds through deep learning. Central to our contribution is TexStat, a loss function grounded in au- ditory perception and statistical modeling. By explicitly encoding key properties such as time invariance, perceptual robustness, and long-term struc...
work page 2025
-
[7]
IA y Música: Cá- tedra en Inteligencia Artificial y Música (TSI-100929-2023-1)
ACKNOWLEDGMENTS This work has been supported by the project "IA y Música: Cá- tedra en Inteligencia Artificial y Música (TSI-100929-2023-1)", funded by the "Secretaría de Estado de Digitalización e Inteligen- cia Artificial and the Unión Europea-Next Generation EU"
work page 2023
-
[8]
D.F. Rosenthal, H.G. Okuno, H. Okuno, and D. Rosenthal, Computational Auditory Scene Analysis: Proceedings of the Ijcai-95 Workshop, CRC Press, 1st edition, 1998
work page 1998
Show all 33 references
-
[9]
In the context of texture sound analysis and resynthesis, this paper has three main goals
and its use of Generative Adversarial Networks (GANs), where the objective of the loss function is to train the model to generate sounds that can deceive a classifier, thereby biasing the training in a more nuanced manner than direct signal comparison. In the context of textur...
-
[10]
Visual pattern discrimination,
Bela Julesz, “Visual pattern discrimination,” IRE Transac- tions on Information Theory , vol. 8, no. 2, pp. 84–92, Feb. 1962
1962
-
[11]
On perceptual analyzers un- derlying visual texture discrimination: Part i,
Terry Caelli and Bela Julesz, “On perceptual analyzers un- derlying visual texture discrimination: Part i,” Biological Cybernetics, vol. 28, no. 3, pp. 167–175, Sep. 1978
1978
-
[12]
Texture feature extraction meth- ods: A survey,
Anne Humeau-Heurtier, “Texture feature extraction meth- ods: A survey,” IEEE Access, vol. 7, pp. 8975–9000, 2019
2019
-
[13]
Sound texture synthesis via filter statistics,
Josh H. McDermott, Andrew J. Oxenham, and Eero P. Si- moncelli, “Sound texture synthesis via filter statistics,” in 2009 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics , New Paltz, NY , October 2009
2009
-
[14]
Sound texture perception via statistics of the auditory periphery: evidence from sound synthesis,
Josh H. McDermott and Eero P. Simoncelli, “Sound texture perception via statistics of the auditory periphery: evidence from sound synthesis,” Neuron, vol. 71, no. 5, pp. 926–940, 2011
2011
-
[15]
Summary statistics in auditory perception,
Josh H. McDermott, Michael Schemitsch, and Eero P. Si- moncelli, “Summary statistics in auditory perception,” Na- ture Neuroscience, vol. 16, no. 4, pp. 493–498, 2013
2013
-
[16]
Classification of sound textures,
Nicholas Saint-Arnaud, “Classification of sound textures,” M.S. thesis, Massachusetts Institute of Technology, Cam- bridge, MA, Sep. 1995
1995
-
[17]
An audio texture lutherie,
Lonce Wyse, “An audio texture lutherie,” Annual Art Jour- nal, vol. 43, no. 54, 2022
2022
-
[18]
Trends in audio texture analysis, synthesis, and appli- cations,
Garima Sharma, Karthikeyan Umapathy, and Sridhar Krish- nan, “Trends in audio texture analysis, synthesis, and appli- cations,” J. Audio Eng. Soc. , vol. 70, no. 3, pp. 108–127, March 2022
2022
-
[19]
Analysis and synthe- sis of sound textures,
Nicolas Saint-Arnaud and Kris Popat, “Analysis and synthe- sis of sound textures,” in Readings in Computational Audi- tory Scene Analysis, pp. 125–131. 1995
1995
-
[20]
Synthesizing sound textures through wavelet tree learning,
S. Dubnov, Z. Bar-Joseph, R. El-Yaniv, D. Lischinski, and M. Werman, “Synthesizing sound textures through wavelet tree learning,” IEEE Computer Graphics and Applications , vol. 22, no. 4, pp. 38–48, 2002
2002
-
[21]
thesis, Université Paris 6 – Pierre et Marie Curie, Paris, France, 2004
Diemo Schwarz, Data-Driven Concatenative Sound Synthe- sis, Ph.D. thesis, Université Paris 6 – Pierre et Marie Curie, Paris, France, 2004
2004
-
[22]
The synthesis of environmental sound textures by iterated nonlinear functions, and its ecological relevance to perceptual modeling,
Agostino Di Scipio, “The synthesis of environmental sound textures by iterated nonlinear functions, and its ecological relevance to perceptual modeling,” Journal of New Music Research, vol. 31, no. 2, pp. 109–117, 2002
2002
-
[23]
Synthesizing sounds from rigid-body simulations,
James F. O’Brien, Chen Shen, and Christine M. Gatchalian, “Synthesizing sounds from rigid-body simulations,” in ACM SIGGRAPH/Eurographics Symposium on Computer Anima- tion, 2002, pp. 175–181
2002
-
[24]
Sound texture synthe- sis using convolutional neural networks,
Hugo Caracalla and Axel Roebel, “Sound texture synthe- sis using convolutional neural networks,” in Proceedings of the 22nd International Conference on Digital Audio Effects (DAFx-19), Birmingham, UK, September 2019
2019
-
[25]
Noisebandnet: Controllable time-varying neural synthesis of sound effects using filterbanks,
Adrián Barahona-Ríos and Tom Collins, “Noisebandnet: Controllable time-varying neural synthesis of sound effects using filterbanks,” IEEE/ACM Trans. Audio, Speech and Lang. Proc., vol. 32, pp. 1573–1585, Feb. 2024
2024
-
[26]
Suggested formu- lae for calculating auditory-filter bandwidths and excitation patterns,
Brian C. Moore and Brian R. Glasberg, “Suggested formu- lae for calculating auditory-filter bandwidths and excitation patterns,” The Journal of the Acoustical Society of America , vol. 74, no. 3, pp. 750–753, 1983
1983
-
[27]
An efficient auditory filterbank based on the gam- matone function,
Roy Patterson, Ian Nimmo-Smith, John Holdsworth, and Pe- ter Rice, “An efficient auditory filterbank based on the gam- matone function,” in Speech-Group Meeting of the Institute of Acoustics on Auditory Modelling , RSRE, Malvern, Dec. 1987, Institute of Acoustics, Meeting held...
1987
-
[28]
Frechet audio distance: A reference-free metric for evaluating music enhancement algorithms,
Kevin Kilgour, Mauricio Zuluaga, Dominik Roblek, and Matthew Sharifi, “Frechet audio distance: A reference-free metric for evaluating music enhancement algorithms,” inIN- TERSPEECH 2019, Graz, Austria, September 2019, ISCA
2019
-
[29]
Ddsp: Differentiable digital signal processing,
Jesse Engel, Lalit Hantrakul, Chenjie Gu, and Adam Roberts, “Ddsp: Differentiable digital signal processing,” in Interna- tional Conference on Learning Representations , 2020
2020
-
[30]
Spectral modeling synthesis: A sound analysis/synthesis system based on a deterministic plus stochastic decomposition,
Xavier Serra and Julius Smith, “Spectral modeling synthesis: A sound analysis/synthesis system based on a deterministic plus stochastic decomposition,” Computer Music Journal , vol. 14, no. 4, pp. 12–24, 1990
1990
-
[31]
Freesound technical demo,
Frederic Font, Gerard Roma, and Xavier Serra, “Freesound technical demo,” in Proceedings of the 21st ACM Inter- national Conference on Multimedia , New York, NY , USA, 2013, MM ’13, p. 411–412, Association for Computing Ma- chinery
2013
-
[32]
Syntex: parametric audio texture datasets for conditional training of instrumental interfaces.,
Lonce Wyse and Prashanth Thattai Ravikumar, “Syntex: parametric audio texture datasets for conditional training of instrumental interfaces.,” in Proceedings of the International Conference on New Interfaces for Musical Expression , The University of Auckland, New Zealand, jun 2022
2022
-
[33]
Cnn architectures for large-scale audio classification,
Shawn Hershey, Sourish Chaudhuri, Daniel P. W. Ellis, Jort F. Gemmeke, Aren Jansen, R. Channing Moore, Manoj Plakal, Devin Platt, Rif A. Saurous, Bryan Seybold, Malcolm Slaney, Ron J. Weiss, and Kevin Wilson, “Cnn architectures for large-scale audio classification,” in 2017 IE...
2017
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.