Pith. sign in

REVIEW 3 major objections 6 minor 31 references

Spectrotemporal Modulation: Efficient and Interpretable Feature Representation for Classifying Speech, Music, and Environmental Sounds

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Spectrotemporal modulation features, computed with no pretraining, classify speech, music, and environmental sounds with accuracy comparable to pretrained audio DNNs.

desk verdict Useful large-scale benchmark showing hand-crafted STM features nearly match frozen DNN embeddings on long sounds, but the 'comparable' claim needs error bars and the model-size note has a numerical error. read the letter →

arxiv 2505.23509 v1 pith:AVDAFAA5 submitted 2025-05-29 cs.SD cs.LGeess.AS

Andrew Chang , Yike Li , Iran R. Roman , David Poeppel This is my paper
classification cs.SDcs.LGeess.AS
keywords spectrotemporalmodulationaudioclassificationspeechmusicenvironmentalsoundspretraining-freerepresentationinterpretabilityauditorycortex
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

This paper tries to establish that spectrotemporal modulation (STM) — a two-dimensional modulation spectrum describing how a sound's spectral and temporal modulations change — can replace pretrained deep embeddings as the input to an audio classifier. On a large test set spanning speech, music, and environmental sounds, a simple multilayer perceptron trained on STM features achieves a ROC-AUC of 0.988 and a macro-F1 of 0.808, with no pretraining whatsoever. Those numbers sit close to classifiers trained on AST (0.995, 0.884), YAMNet (0.990, 0.849), and VGGish (0.993, 0.871) embeddings, and far above mel-spectrogram features (0.944, 0.479). The reason to care is that STM is both cheap — a fixed signal-processing step with no learned parameters — and neurophysiologically interpretable, so a machine listener built on it is directly linked to how the human auditory cortex represents sound.

What carries the argument

The central object is the spectrotemporal modulation (STM) spectrum: a two-dimensional power spectrum whose axes are temporal modulation rate (in Hz) and spectral modulation scale (in cycles per octave), obtained by taking a two-dimensional Fourier transform of an auditory spectrogram. It is designed to mimic the modulation-selective representation of sound in the human auditory cortex. In this paper it does double duty: it is the sole input representation to a small MLP classifier, and it also serves as an interpretable domain in which class differences can be inspected before any supervised modeling — the class-averaged STMs and a t-SNE projection already show separation between speech, music, and environmental sounds. The reduced STM subspace (≤ ±4 Hz, ≤ 6 cyc/oct) is the part of the representation that does the heavy lifting for classification.

What would settle it

Recompute Table 2 with 10 random 80/10/10 splits (or repeated seeds) and report means and confidence intervals for ROC-AUC and macro-F1; the 'comparable' claim is falsified if STM falls more than roughly 0.03 below AST or VGGish in macro-F1 on a majority of splits.

Watch

Extended reading notes

Core claim

The central claim is that the spectrotemporal modulation spectrum of an audio clip carries enough information to separate speech, music, and environmental sounds at a level close to that of deep embeddings from pretrained audio networks. The authors compute STM by filtering a four-second waveform through 128 cochlear-style Gaussian bands, taking the Hilbert envelope, and applying a two-dimensional Fourier transform to obtain 2,420 modulation features; they then train a single multilayer perceptron on those features with no pretraining. On the held-out test set, this model reaches ROC-AUC 0.988 and macro-F1 0.808, compared with AST at 0.995/0.884, YAMNet at 0.990/0.849, and VGGish at 0.993/0.871, while mel-spectrogram features achieve only 0.944/0.479 under the same MLP training pipeline. An ablation shows that low temporal (≤ ±4 Hz) and low spectral (≤ 6 cyc/oct) modulation regions carry most of the classification-relevant information, matching the regions where class-averaged STMs differ most.

Load-bearing premise

The load-bearing premise is that one fixed 80/10/10 split of the data gives accurate performance estimates; the paper reports no confidence intervals, repeated seeds, or cross-validation folds, so the observed closeness between STM and pretrained models could shift under a different split.

Editorial extensions

If this is right

  • Audio classification for speech, music, and environmental sounds can be done without pretraining on millions of labeled samples; a fixed STM front-end plus a small MLP is enough.
  • Because the STM feature extractor has zero learned parameters, models built on it cost orders of magnitude less to train and deploy than AST, YAMNet, or VGGish pipelines.
  • The low-modulation band (≤ ±4 Hz temporal, ≤ 6 cyc/oct spectral) is a sufficient STM subspace, so future extraction can drop high-modulation bins without losing classification performance.
  • STM's direct correspondence to auditory cortical modulation tuning means a classifier's decisions can be traced back to a physiological feature space, offering an interpretable alternative to deep embeddings.

Reading between the lines

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

  • The paper does not test sub-second transient events; if STM features were computed with higher temporal resolution and shorter windows, the same framework could be extended to sound-event detection, a domain where the authors note DNN embeddings may hold an advantage.
  • A natural follow-up is to use STM-based classifiers as encoding models for human auditory cortex responses; because their input features are modulation spectra, their internal states can be compared directly with neural tuning without the confound of pretrained visual-model features.
  • The absence of confidence intervals means the reported 'comparable' performance is a point estimate; an outside reader should treat the exact gap as provisional until repeated-split or repeated-seed results are available.
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

3 major / 6 minor

Summary. The paper proposes spectrotemporal modulation (STM) features, a fixed signal-processing representation inspired by auditory-cortex modulation analysis, and trains a small MLP classifier on them for a seven-class task (tonal/non-tonal speech, vocal/non-vocal music, urban/wildlife environmental sounds) using roughly 978k recordings. The authors report that the STM-based model achieves ROC-AUC 0.988 and macro-F1 0.808 without pretraining, compared with AST (0.995/0.884), YAMNet (0.990/0.849), and VGGish (0.993/0.871) evaluated with the same MLP head and no fine-tuning; a mel-spectrogram control performs much worse. They also present class-averaged STM visualizations, Cohen's d effect sizes, a tSNE projection, an undersampling check, and an STM ablation study.

Significance. If the comparability claim holds, this is a valuable result: fixed, interpretable, neurally motivated features reaching near-DNN accuracy on a large, naturalistic corpus without pretraining, with a small MLP head and clear visualization of the discriminative modulation regions. The manuscript has concrete strengths: a public reproducibility repository with a DOI, a shared MLP head and hyperparameter tuning protocol across all feature types, an explicit mel-spectrogram control for the MLP architecture, and STM features that are fixed signal-processing constructs not fitted to labels, so the classification result does not inherit circularity from the neuroscience framing. The main caveat is that the headline comparison rests on a single grouped split with no variance estimates, and part of the music labels were themselves generated with one of the comparator models, which can bias the comparison.

major comments (3)
  1. [Section 2.3 and Table 2] All reported metrics come from one fixed 8:1:1 StratifiedGroupKFold split with a single Bayesian hyperparameter search per feature; no confidence intervals, repeated seeds, cross-validation folds, or paired significance tests are reported. Because the split groups by speaker, musician, and site, the effective sample size for assessing generalization is the number of groups, not the ~978k recordings; a different grouping could plausibly move the observed macro-F1 gaps between STM (0.808) and the pretrained embeddings (0.849–0.884) by an amount comparable to the gaps themselves. The authors should report repeated stratified group k-fold results or bootstrap intervals, and ideally define an equivalence margin for the claim that STM is 'comparable' to pretrained audio DNNs.
  2. [Section 2.1.2 and Section 2.4] For the Free Music Archive, ISMIR04, MagnaTagATune, and MTG-Jamendo corpora, the vocal/non-vocal labels were produced by running Demucs followed by YAMNet, and the same YAMNet model later provides one of the compared feature embeddings. Evaluating YAMNet on labels that YAMNet itself helped create can only bias the comparison in YAMNet's favor, inflating the apparent gap between STM and YAMNet on the music subclasses. The authors should either exclude or relabel those instances for the comparison, or report results separately for corpora with human-provided labels.
  3. [Table 2 note and Table 1] The note in Table 2 claims that STM 'reduc[es] model size by 7 orders of magnitude' compared with the pretrained DNNs. This is numerically contradicted by Table 1: the STM (full) model uses 1.3M MLP parameters, YAMNet uses 3.8M backbone plus 0.8M MLP, VGGish uses 72.1M plus 0.9M, and AST uses 86.2M plus 0.7M. The actual reduction is roughly 3.5x to 67x, i.e., about 0.5 to 1.8 orders of magnitude, not seven. This quantitative claim should be corrected or removed.
minor comments (6)
  1. [Section 2.1.1] The corpus name 'MozillaCommonV oice' contains an erroneous space; it should be 'MozillaCommonVoice'.
  2. [Section 2.4] The model name 'Y AMNet' appears with a space in several places; it should be 'YAMNet' throughout.
  3. [Table 2] The mel-spectrogram row reports a macro-F1 of 0.000 for 'urban env' in both the full and undersampled settings; this is surprising and should be explained or verified, as it may indicate a collapsed class.
  4. [Section 2.3] The text refers to 'keras.BayesianOptimization()'; the standard implementation is in keras_tuner, and the specific package/version should be cited for reproducibility.
  5. [Figure 3] The caption notes that plots are truncated at 6 cyc/oct and ±6 Hz for visualization, but the model still included features beyond these limits; the caption should also state the full STM range so readers can interpret the ablation correctly.
  6. [References] Reference [6] contains 'inICASSP 2024' without a space; it should read 'in ICASSP 2024'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the STM feature pipeline is a fixed, label-free signal transform and the classification comparison is an independent benchmark.

full rationale

The paper's central empirical claim is that fixed spectrotemporal-modulation (STM) features, fed to a simple MLP without pretraining, achieve performance comparable to frozen embeddings from pretrained audio DNNs (Table 2). The STM features are computed by a deterministic signal-processing pipeline (Section 2.2): a filter-Hilbert spectrogram followed by a 2D FFT, cropping, and downsampling; no class labels or test-set information enter feature construction. The MLP is trained with standard train/validation/test separation (Section 2.3), and hyperparameters are tuned on validation only. No fitted parameter is renamed as a prediction, and no equation defining STM depends on the classification target. The paper does cite prior work by the same group to justify the pipeline and the neurophysiological framing (e.g., Flinker et al. 2019 and Chang et al. 2024), but those citations are not load-bearing for the classification result: even if the neuroscience framing were set aside, Table 2 stands as an external benchmark against AST, YAMNet, and VGGish. The single-split evaluation without confidence intervals is a robustness/correctness limitation, not a circularity. Accordingly, no circular step is present.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physical or conceptual entities are introduced. The free parameters are preprocessing and architecture choices, while the assumptions are domain-level claims about feature adequacy, label reliability, and split integrity.

free parameters (6)
  • STM temporal modulation range = -15 to 15 Hz
    Chosen by the authors to cover the modulation spectrum of speech and music while reducing feature count (Section 2.2).
  • STM spectral modulation range = 0 to 7.09 cyc/oct
    Selected to reduce features and avoid overfitting while covering speech and music modulation spectra (Section 2.2).
  • Audio chunk length = 4 seconds
    Non-overlapping 4-second chunks are used for STM analysis; chunks shorter than 4 seconds are excluded (Section 2.2).
  • Maximum audio used per sample = 120 seconds
    Up to 120 seconds of signal are used per audio sample, with chunks averaged (Section 2.2).
  • Silence exclusion threshold = 1 second
    Chunks containing 1 second or more of silence are excluded from averaging (Section 2.2).
  • PCA target dimension = 1024
    PCA is trained on the training set to reduce STM and mel feature dimensions to 1024 before the MLP (Section 2.3).
assumptions (5)
  • domain assumption STM spectrum captures task-relevant auditory information for speech, music, and environmental sounds.
    Core premise of the paper, supported by prior neurophysiology and signal processing work, not proven here.
  • domain assumption An MLP is expressive enough to learn from fixed STM features.
    Section 2.3 justifies this by noting STM axes are fixed and lack translational equivalence, so an MLP is used rather than a CNN or RNN.
  • domain assumption The dataset labels and grouping metadata prevent speaker, musician, or site leakage.
    StratifiedGroupKFold is used to keep recordings from the same source in the same split; this assumes the metadata groups are correct and complete.
  • domain assumption Automated vocal/nonvocal labels from Demucs and YAMNet are reliable enough for training and evaluation.
    Used for music corpora that lack vocal annotations; the Garland corpus is manually labeled by the first author (Section 2.1.2).
  • standard math Standard signal processing operations (filter-Hilbert spectrogram, 2D FFT, PCA) are valid for this purpose.
    These are standard tools in audio analysis; the paper does not introduce new mathematics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spectrotemporal Modulation: Efficient and Interpretable Feature Representation for Classifying Speech, Music, and Environmental Sounds." pith.science (2026). https://pith.science/paper/AVDAFAA5

@misc{pith2026250523509,
  author       = {Pith},
  title        = {Pith review of: Spectrotemporal Modulation: Efficient and Interpretable Feature Representation for Classifying Speech, Music, and Environmental Sounds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AVDAFAA5}},
  note         = {Machine review of arXiv:2505.23509}
}
read the original abstract

Audio DNNs have demonstrated impressive performance on various machine listening tasks; however, most of their representations are computationally costly and uninterpretable, leaving room for optimization. Here, we propose a novel approach centered on spectrotemporal modulation (STM) features, a signal processing method that mimics the neurophysiological representation in the human auditory cortex. The classification performance of our STM-based model, without any pretraining, is comparable to that of pretrained audio DNNs across diverse naturalistic speech, music, and environmental sounds, which are essential categories for both human cognition and machine perception. These results show that STM is an efficient and interpretable feature representation for audio classification, advancing the development of machine listening and unlocking exciting new possibilities for basic understanding of speech and auditory sciences, as well as developing audio BCI and cognitive computing.

Figures

Figures reproduced from arXiv: 2505.23509 by the authors.

Figure 1
Figure 1. A processing pipeline for converting an audio signal into a spectrogram, then into STM features for training an MLP classifier. 2.1.1. Speech corpora We included 737,948 single-voice recordings (102,729 and 635,219 of tonal and nontonal language recordings), totaling 1,148.17 hours. Among the 71,768 speakers, 32.83% were male, 19.99% were female, and 47.19% were not reported. The speech recordings included 97 langua… view at source ↗
Figure 2
Figure 2. The STM of the different classes in the real-world audio that we study shows clearly separable features. (A) Class-averaged STMs: clear difference between the major classes of speech, music and environmental sounds; (B) Cohen’s d: the differences between subclasses (tonal vs nontonal speech, vocal vs non-vocal music, and urban vs wildlife environmental sounds) is also captured by STM: speech dominates in higher temp… view at source ↗
Figure 3
Figure 3. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages

  1. [1]

    A critical step in developing powerful audio deep neu- ral networks (DNNs) is converting audio signals into meaning- ful acoustic feature representations

    Introduction Sound classification is a fundamental task in machine listen- ing. A critical step in developing powerful audio deep neu- ral networks (DNNs) is converting audio signals into meaning- ful acoustic feature representations. Common approaches in- clude spectrograms, mel-frequency cepstral coefficients, linear predictive coding, wavelet decomposi...

    work page

  2. [2]

    Spectrotemporal Modulation: Efficient and Interpretable Feature Representation for Classifying Speech, Music, and Environmental Sounds

    Approach Dataset and experiments are completely reproducible by using all the code and documentation in the repository.1 2.1. Data preparation The dataset consists of audio tracks featuring speech (tonal and non-tonal), music (vocal and non-vocal), and environmental sounds/others (urban and wildlife). Standard multi-domain au- dio datasets and benchmarks ...

    work page Pith review arXiv

  3. [3]

    Results 3.1. STM feature representation and statistical analyses The class-averaged STMs and the distribution of individual STMs in tSNE space (Figure 2) showed that, even without su- pervised neural network modeling, the STM feature represen- tation can effectively distinguish sound categories in the tSNE space. Specifically, the speech audio was more di...

    work page

  4. [4]

    Discussion and Conclusion We demonstrate the potential of STM features for building pow- erful, efficient, and interpretable audio model to classify speech, music, and environmental sounds. Our STM-based neural net- work model, without any pretraining, achieved exceptional per- formance, comparable to the models trained on the deep embed- ded features of ...

    work page

  5. [5]

    Acknowledgements A.C. is supported by National Research Service Award, NIDCD/NIH (F32DC018205), Leon Levy Scholarships in Neu- roscience, Leon Levy Foundation and New York Academy of Sciences, and GRAMMY Museum Grants Program. The fun- ders have no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. T...

    work page

  6. [6]

    A survey of audio classification using deep learning,

    K. Zaman, M. Sah, C. Direkoglu, and M. Unoki, “A survey of audio classification using deep learning,” IEEE Access, vol. 11, pp. 106 620–106 649, Sep. 2023

    work page 2023

  7. [7]

    AST: Audio spectrogram transformer,

    Y . Gong, Y .-A. Chung, and J. Glass, “AST: Audio spectrogram transformer,” Interspeech 2021, 2021

    work page 2021

  8. [8]

    Audio set: An ontology and human-labeled dataset for audio events,

    J. F. Gemmeke, D. P. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter, “Audio set: An ontology and human-labeled dataset for audio events,” in ICASSP 2017- 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2017

    work page 2017

Show all 31 references
  1. [9]

    CNN architectures for large-scale audio classification,

    S. Hershey, S. Chaudhuri, D. P. Ellis, J. F. Gemmeke, A. Jansen, R. C. Moore, M. Plakal, D. Platt, R. A. Saurous, B. Seybold et al., “CNN architectures for large-scale audio classification,” in ICASSP 2017-2017 IEEE International Conference on Acoustics, Speech and Signal Proc...

    2017

  2. [10]

    Richard, V

    G. Richard, V . Lostanlen, Y .-H. Yang, and M. M ¨uller, “Model- based deep learning for music information research: Leveraging diverse knowledge sources to enhance explainability, controlla- bility, and resource efficiency [special issue on model-based and data-driven audio s...

    2025

  3. [11]

    Robust DOA esti- mation from deep acoustic imaging,

    A. S. Roman, I. R. Roman, and J. P. Bello, “Robust DOA esti- mation from deep acoustic imaging,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). IEEE, 2024, pp. 1321–1325

    2024

  4. [12]

    A review of differentiable digital signal processing for music and speech synthesis,

    B. Hayes, J. Shier, G. Fazekas, A. McPherson, and C. Saitis, “A review of differentiable digital signal processing for music and speech synthesis,” Frontiers in Signal Processing, vol. 3, p. 1284100, 2024

    2024

  5. [13]

    The modulation spectrogram: In pursuit of an invariant representation of speech,

    S. Greenberg and B. E. Kingsbury, “The modulation spectrogram: In pursuit of an invariant representation of speech,” in ICASSP 1997-1997 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), vol. 3. IEEE, 1997, pp. 1647– 1650

    1997

  6. [14]

    Speech intel- ligibility prediction using spectro-temporal modulation analysis,

    A. Edraki, W.-Y . Chan, J. Jensen, and D. Fogerty, “Speech intel- ligibility prediction using spectro-temporal modulation analysis,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 29, pp. 210–225, 2020

    2020

  7. [15]

    Com- paring different flavors of spectro-temporal features for ASR

    B. T. Meyer, S. V . Ravuri, M. R. Sch¨adler, and N. Morgan, “Com- paring different flavors of spectro-temporal features for ASR.” in Interspeech, 2011, pp. 1269–1272

    2011

  8. [16]

    Automatic mu- sic genre classification based on modulation spectral analysis of spectral and cepstral features,

    C.-H. Lee, J.-L. Shih, K.-M. Yu, and H.-S. Lin, “Automatic mu- sic genre classification based on modulation spectral analysis of spectral and cepstral features,”IEEE Transactions on Multimedia, vol. 11, no. 4, pp. 670–682, 2009

    2009

  9. [17]

    Discrimination of speech from nonspeech based on multiscale spectro-temporal modulations,

    N. Mesgarani, M. Slaney, and S. A. Shamma, “Discrimination of speech from nonspeech based on multiscale spectro-temporal modulations,” IEEE Transactions on audio, speech, and language processing, vol. 14, no. 3, pp. 920–930, 2006

    2006

  10. [18]

    Distinct sensitivity to spectrotemporal modulation supports brain asym- metry for speech and melody,

    P. Albouy, L. Benjamin, B. Morillon, and R. J. Zatorre, “Distinct sensitivity to spectrotemporal modulation supports brain asym- metry for speech and melody,” Science, vol. 367, no. 6481, pp. 1043–1047, 2020

    2020

  11. [19]

    Spectrotemporal modulation provides a unifying framework for auditory cortical asymmetries,

    A. Flinker, W. K. Doyle, A. D. Mehta, O. Devinsky, and D. Poep- pel, “Spectrotemporal modulation provides a unifying framework for auditory cortical asymmetries,” Nature Human Behaviour , vol. 3, no. 4, pp. 393–405, Apr. 2019

    2019

  12. [20]

    Spectro-temporal acoustical markers differenti- ate speech from song across cultures,

    P. Albouy, S. A. Mehr, R. S. Hoyer, J. Ginzburg, Y . Du, and R. J. Zatorre, “Spectro-temporal acoustical markers differenti- ate speech from song across cultures,” Nature Communications, vol. 15, no. 1, 2024

    2024

  13. [21]

    The human auditory system uses amplitude modulation to distinguish music from speech,

    A. Chang, X. Teng, M. F. Assaneo, and D. Poeppel, “The human auditory system uses amplitude modulation to distinguish music from speech,” PLOS Biology, vol. 22, no. 5, p. e3002631, May 2024

    2024

  14. [22]

    Distinct cortical pathways for music and speech revealed by hypothesis-free voxel decomposition,

    S. Norman-Haignere, N. G. Kanwisher, and J. H. McDermott, “Distinct cortical pathways for music and speech revealed by hypothesis-free voxel decomposition,” Neuron, vol. 88, no. 6, pp. 1281–1296, 2015

    2015

  15. [23]

    Hybrid Transformers for Music Source Separation,

    S. Rouard, F. Massa, and A. D ´efossez, “Hybrid Transformers for Music Source Separation,” in ICASSP 2023-2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  16. [24]

    SONYC Urban Sound Tagging (SONYC-UST): a multilabel dataset from an urban acoustic sen- sor network,

    M. Cartwright, J. Cramer, A. E. M. Mendez, Y . Wang, H.-H. Wu, V . Lostanlen, M. Fuentes, G. Dove, C. Mydlarz, J. Sala- mon, O. Nov, and J. P. Bello, “SONYC Urban Sound Tagging (SONYC-UST): a multilabel dataset from an urban acoustic sen- sor network,” Sep. 2020

    2020

  17. [25]

    eBird: A citizen-based bird observation network in the biological sciences,

    B. L. Sullivan, C. L. Wood, M. J. Iliff, R. E. Bonney, D. Fink, and S. Kelling, “eBird: A citizen-based bird observation network in the biological sciences,” Biological Conservation , vol. 142, no. 10, pp. 2282–2292, 2009

    2009

  18. [26]

    The cortical organization of speech processing,

    G. Hickok and D. Poeppel, “The cortical organization of speech processing,” Nature Reviews Neuroscience, vol. 8, no. 5, pp. 393– 402, 2007

    2007

  19. [27]

    Facing Imbalanced Data Recommendations for the Use of Performance Metrics,

    L. A. Jeni, J. F. Cohn, and F. De La Torre, “Facing Imbalanced Data Recommendations for the Use of Performance Metrics,”Hu- maine Association Conference on Affective Computing and Intel- ligent Interaction, pp. 245–251, 2013

    2013

  20. [28]

    Survey on deep learning with class imbalance,

    J. M. Johnson and T. M. Khoshgoftaar, “Survey on deep learning with class imbalance,” Journal of Big Data , vol. 6, no. 1, Mar. 2019

    2019

  21. [29]

    Many but not all deep neural network audio models capture brain responses and exhibit correspondence between model stages and brain regions,

    G. Tuckute, J. Feather, D. Boebinger, and J. H. McDermott, “Many but not all deep neural network audio models capture brain responses and exhibit correspondence between model stages and brain regions,” PLOS Biology, vol. 21, no. 12, p. e3002366, Dec. 2023

    2023

  22. [30]

    Reconstructing the spectrotem- poral modulations of real-life sounds from fMRI response pat- terns,

    R. Santoro, M. Moerel, F. De Martino, G. Valente, K. Ugurbil, E. Yacoub, and E. Formisano, “Reconstructing the spectrotem- poral modulations of real-life sounds from fMRI response pat- terns,” Proceedings of the National Academy of Sciences, vol. 114, no. 18, 2017

    2017

  23. [31]

    Decoding spectrotemporal features of overt and covert speech from the hu- man cortex,

    S. Martin, P. Brunner, C. Holdgraf, H.-J. Heinze, N. E. Crone, J. Rieger, G. Schalk, R. T. Knight, and B. N. Pasley, “Decoding spectrotemporal features of overt and covert speech from the hu- man cortex,” Frontiers in Neuroengineering, vol. 7, p. 14, 2014

    2014

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.