Pith. sign in

REVIEW 3 major objections 4 minor 49 references

Estimating Musical Surprisal in Audio

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The information content of a Transformer's next-step predictions on compressed audio latents predicts EEG brain responses to songs, supporting an audio-domain proxy for musical surprisal.

desk verdict A genuinely new audio-level surprisal estimator with a real EEG validation, but the perceptual claim currently rests on a single acoustic nuisance regressor, so the result is promising rather than conclusive. read the letter →

arxiv 2501.07474 v1 pith:WUDE36R2 submitted 2025-01-13 cs.SD cs.AIeess.AS

Mathias Rose Bjare , Giorgia Cantisani , Stefan Lattner , Gerhard Widmer This is my paper · ORCID
classification cs.SDcs.AIeess.AS
keywords musicalsurprisalinformationcontentautoregressiveTransformeraudiolatentcompressionMusic2LatentGaussianmixturemodelEEGpredictionmusicretrieval
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

The paper tries to establish that musical surprisal can be estimated directly from full-length audio recordings by training an autoregressive Transformer to predict the next frame of a compressed latent audio representation, then reading the information content of that prediction as a surprisal score. If true, this would let researchers study musical expectation on any audio recording, including timbre and dynamics that symbolic score representations miss. The paper supports the claim by showing that the estimated information content drops when musical material repeats, rises for later contrasting sections, correlates with timbre-related and loudness features, and significantly predicts EEG brain responses to songs after controlling for the energy envelope.

What carries the argument

The central object is the information content of one-step predictions from a GIVT, a causal Transformer whose output head parameterizes a 32-component Gaussian mixture model over the next 64-dimensional frame of a Music2Latent audio latent sequence at roughly 11 frames per second. The information content is the negative log-density of the observed next frame under the predicted mixture, replacing the softmax over a finite token vocabulary used in symbolic music models. This lets the surprisal estimate flow from the raw audio content itself, including timbre and dynamics, and the same quantity is then compared with repetition structure, section type, musical features, and EEG responses.

What would settle it

Run the EEG prediction with spectral flux, loudness, onset density, and timbre descriptors added as competing regressors before the information content; if the information content no longer improves prediction, its apparent perceptual role is an acoustic confound rather than surprisal.

Watch

Extended reading notes

Core claim

The paper claims that musical surprisal in audio can be modeled without preselected audio features by computing the information content of next-step predictions from a GIVT, a causal Transformer whose output head parameterizes a 32-component Gaussian mixture model over the next frame of a Music2Latent latent sequence. The resulting audio-domain information content behaves like symbolic-domain surprisal: it decreases with repetition on the model's training distribution, it is higher for segment types that appear later in a piece while outros show lower values, and it correlates more strongly with loudness and spectral flux than with dissonance, rhythmic complexity, or onset density. The decisive validation is that this information content, competing with the energy envelope as a predictor in a linear encoding model, significantly improves prediction of EEG responses to songs, supporting the methodology's potential as a perceptual model of human musical surprisal.

Load-bearing premise

The result assumes that the EEG variance explained by the information content after removing only the energy envelope is due to musical surprisal, not to other acoustic properties, like timbre or loudness changes, that happen to move together with the information content.

Editorial extensions

If this is right

  • Repetition lowers the mean information content on a large popular-music training set, showing that the model learns musical structure from audio alone.
  • Segment types appearing later in a piece have higher information content than earlier ones, while outros have lower information content, supporting the idea that later sections are composed to contrast earlier material.
  • Information content correlates most strongly with loudness and spectral flux, and less with dissonance, rhythmic complexity, and onset density; these correlations are highest at the start of a piece and decay over time.
  • Information content explains a small but significant portion of EEG variance beyond the energy envelope, so the audio-based surprisal signal tracks human neural responses.
  • Because the method needs no preselected features, it can be applied to arbitrary full-length audio recordings, including timbral and dynamic content unavailable in symbolic scores.

Reading between the lines

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

  • A natural next test is whether the information content's EEG contribution survives when spectral flux and loudness are added as competing regressors; if it does not, the perceptual claim would need to be attributed to acoustic surprise rather than musical expectation.
  • The same GIVT-plus-Music2Latent recipe could be used to condition generative music models on surprisal curves, turning the estimator into a controllable knob for compositional contrast.
  • Applying the method to genres outside popular music, or to recordings with strong production effects, would clarify whether the repetition and segment-contrast effects generalize beyond the training distribution.
  • Because the model is trained on audio itself rather than on score-like events, its information content likely mixes acoustic novelty and music-theoretic expectation; separating these would require a model whose predictions are evaluated over semantically distinct event types.
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 / 4 minor

Summary. The paper proposes a method for estimating musical surprisal directly from audio by training an autoregressive GIVT model to predict Music2Latent latent representations and defining information content (IC) as the negative log density of a Gaussian mixture model's next-frame prediction. The authors evaluate the resulting IC in four ways: it decreases with repetition on the largest dataset, later-occurring segment types have higher IC than earlier ones (while outros have lower IC), it correlates with several audio and musical features (dissonance, IOI entropy, onset density, loudness, spectral flux), and it significantly improves prediction of EEG responses to sung music beyond an energy-envelope baseline. Code is released, and evaluation includes public datasets.

Significance. If the results hold, the paper extends information-theoretic surprisal modeling from the symbolic domain to full-length audio without relying on hand-selected audio features, which is a useful step for music cognition and MIR. The experimental design has notable strengths: the model is evaluated on multiple datasets, the EEG test uses an external benchmark with a shuffled-IC null model, the repetition and segment-contrast analyses are simple and falsifiable, and code is provided. However, the central perceptual claim rests on the EEG analysis, and that analysis currently leaves open a plausible acoustic-confound explanation. The free parameters of the pipeline (latent representation, frame rate, context length, GMM components, architecture) are not ablated, so the contribution is best read as a proof-of-concept rather than a systematic study.

major comments (3)
  1. [Section V-D] The EEG validation controls only for the energy envelope, but the paper's own Section V-C and Table I show that IC is strongly correlated with loudness (rl up to 0.70) and spectral flux (rf up to 0.57). The Hilbert envelope captures amplitude modulations but not spectral or timbral changes. The shuffled-IC baseline only breaks the temporal alignment between IC and the EEG signal; it does not control for acoustic features that are correlated with IC. As a result, the incremental EEG variance attributed to IC could be carried by uncontrolled loudness- or timbre-related acoustics, which would undermine the central conclusion that IC reflects musical surprisal. Please include spectral flux and loudness (and ideally other acoustic descriptors) as additional nuisance regressors, or otherwise show that the IC effect survives when these features compete for the same variance.
  2. [Table I and Section V-C] All complexity correlations are reported only as point estimates, with no confidence intervals, standard errors, or cross-validation, and no correction for the multiple comparisons across five metrics and three datasets. The statement that 'all correlations are significant' is therefore under-supported, especially for the smaller public datasets. Since these correlations are used both to support the complexity associations and to quantify the potential acoustic confounds in the EEG analysis, please provide interval estimates and describe the exact significance-testing procedure.
  3. [Section V-A] The repetition analysis is described as a 'one-tailed t-test for positivity,' but the reported mean difference for PR is negative (-1.30) and is described as a significant decrease. This is confusing: the sign convention for the computed difference (first segment minus second, or vice versa) and the alternative hypothesis of the t-test need to be stated explicitly. As written, a reader cannot tell whether the test is for a positive mean difference or for a negative mean difference.
minor comments (4)
  1. [Equation (2)] The text says the IC 'is unbounded,' but it may be worth clarifying that the negative log density of a continuous GMM can also take negative values when the density exceeds 1. This is a minor notational point but affects how readers interpret IC values.
  2. [Figure 1] The temporal correlation curves are plotted without any confidence bands or significance shading, so it is not clear whether the visible decrease over time is statistically reliable. Please add error bars or a significance annotation.
  3. [Section V-D] The EEG stimuli are monophonic sung music, while the model is trained on multi-stem popular music. The fine-tuning on vocal stems is mentioned, but the paper does not report how IC is computed and aligned for the EEG stimuli (e.g., whether the same 11 Hz frame rate and context length are used, and how the IC time series is resampled to the EEG sampling rate). A brief description of the alignment procedure would improve reproducibility.
  4. [Throughout] There are several typographical errors ('rhymic complexity' in Section II, 'suprisal' in the introduction) and the phrase 't-test for positivity' in Section V-A. A careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the IC is defined from an independently trained next-step density model and validated against external EEG data with a shuffled-IC baseline.

full rationale

The paper's derivation chain is self-contained with respect to its central claim. The IC is defined in Eq. (2) as the negative log density of a Gaussian mixture model parameterized by a causally trained Transformer, with no EEG or perceptual labels in the training objective. The EEG validation in Section V-D uses an external dataset (Cantisani et al., [47]), a linear temporal response model, and a shuffled-IC baseline to assess incremental predictive power; the model parameters are not fitted to neural responses. The repetition, segment-contrast, and feature-correlation analyses are empirical checks, not built into the definition of IC. The paper does cite prior work by its own authors (e.g., Music2Latent [19] and the EEG dataset [47]), but these are external artifacts and not load-bearing in a circular way: the method would stand or fall on the independent reproducibility of those components. The main scientific risk, that IC's EEG predictive power may be carried by acoustic features such as spectral flux or loudness beyond the Hilbert envelope, is a validity concern about confounds, not a circularity in which the prediction is equivalent to a fitted input by construction.

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

The central claim rests on the choice of latent representation, the autoregressive model's architecture and training data, the reliability of automatic metadata, and the validity of the EEG analysis. The trained neural network itself is a fitted model, but its parameters are not tuned to the EEG or complexity measures; the listed free parameters are the manually chosen design decisions that shape the IC. No new physical entities are introduced.

free parameters (5)
  • GMM components = 32
    Number of Gaussian components in the output density model, chosen by hand; affects the smoothness and shape of the estimated IC.
  • Latent dimension = 64
    Dimensionality of the Music2Latent representation; a larger or smaller dimension would change the sequence statistics and the resulting IC.
  • Frame rate = ~11 Hz
    Temporal resolution of the latent sequence; the choice of Music2Latent frame rate determines the granularity of prediction and surprisal.
  • Context length = 4600 frames
    Maximum sequence length the Transformer can condition on, corresponding to about 7 minutes of audio; a shorter context would change what counts as 'unexpected'.
  • Model architecture = 12-layer Transformer
    Capacity and inductive bias of the autoregressive model; a larger or smaller model would produce different density estimates and IC values.
assumptions (5)
  • domain assumption Music2Latent's latent representation preserves enough audio information to capture surprisal-relevant acoustic properties.
    The paper relies on the invertibility of Music2Latent (Section III-A) to justify that no preselected features are needed, but the lossy compression could discard or distort information relevant to harmonic or timbral surprisal.
  • domain assumption Deep12's automatic segment labels and boundaries are accurate enough for the repetition and contrast analyses.
    Experiments in Sections V-A and V-B depend on Deep12 metadata; errors in segment labeling would bias the reported IC differences.
  • domain assumption The mTRF linear encoding model adequately captures the relationship between IC and EEG responses.
    Section V-D uses regularized linear regression with multiple time lags; nonlinear or non-stationary neural encoding would not be captured.
  • domain assumption The EEG dataset and preprocessing (including interpolation of unvoiced segments) are appropriate and unbiased.
    The paper uses EEG data from [47] without re-analyzing the raw recordings; any preprocessing mismatch (e.g., how the vocal stimuli align with the model's IC) could affect the result.
  • domain assumption TIV dissonance, IOI entropy, onset density, and spectral flux are valid proxies for musical complexity.
    These features are adopted from prior literature (Sections V-C) to interpret IC correlations; if they do not measure what the paper claims, the complexity correlations are less meaningful.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Estimating Musical Surprisal in Audio." pith.science (2026). https://pith.science/paper/WUDE36R2

@misc{pith2026250107474,
  author       = {Pith},
  title        = {Pith review of: Estimating Musical Surprisal in Audio},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WUDE36R2}},
  note         = {Machine review of arXiv:2501.07474}
}
read the original abstract

In modeling musical surprisal expectancy with computational methods, it has been proposed to use the information content (IC) of one-step predictions from an autoregressive model as a proxy for surprisal in symbolic music. With an appropriately chosen model, the IC of musical events has been shown to correlate with human perception of surprise and complexity aspects, including tonal and rhythmic complexity. This work investigates whether an analogous methodology can be applied to music audio. We train an autoregressive Transformer model to predict compressed latent audio representations of a pretrained autoencoder network. We verify learning effects by estimating the decrease in IC with repetitions. We investigate the mean IC of musical segment types (e.g., A or B) and find that segment types appearing later in a piece have a higher IC than earlier ones on average. We investigate the IC's relation to audio and musical features and find it correlated with timbral variations and loudness and, to a lesser extent, dissonance, rhythmic complexity, and onset density related to audio and musical features. Finally, we investigate if the IC can predict EEG responses to songs and thus model humans' surprisal in music. We provide code for our method on github.com/sonycslparis/audioic.

Figures

Figures reproduced from arXiv: 2501.07474 by the authors.

Figure 1
Figure 1. Temporal development of correlations between IC and [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. z-transformed Pearson’s correlation gain for individual [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 41 canonical work pages

  1. [1]

    Meaning in music and information theory,

    L. B. Meyer, “Meaning in music and information theory,” The Journal of Aesthetics and Art Criticism , vol. 15, no. 4, pp. 412–424, 1957

    work page 1957

  2. [2]

    Multiple viewpoint systems for music prediction,

    D. Conklin and I. H. Witten, “Multiple viewpoint systems for music prediction,” Journal of New Music Research , vol. 24, no. 1, pp. 51–73, 1995

    work page 1995

  3. [3]

    The construction and evaluation of statistical models of melodic structure in music perception and composition,

    M. Pearce, “The construction and evaluation of statistical models of melodic structure in music perception and composition,” Ph.D. disser- tation, Department of Computing, City University, London, UK, 2005

    work page 2005

  4. [4]

    Auditory expectation: The information dynamics of music perception and cognition,

    M. T. Pearce and G. A. Wiggins, “Auditory expectation: The information dynamics of music perception and cognition,” Top. Cogn. Sci. , vol. 4, no. 4, pp. 625–652, 2012

    work page 2012

  5. [5]

    Information-theoretic modeling of perceived musical complexity,

    S. A. Sauv ´e and M. T. Pearce, “Information-theoretic modeling of perceived musical complexity,” Music Perception: An Interdisciplinary Journal, vol. 37, no. 2, pp. 165–178, 2019

    work page 2019

  6. [6]

    Exploring sampling techniques for generating melodies with a transformer language model,

    M. R. Bjare, S. Lattner, and G. Widmer, “Exploring sampling techniques for generating melodies with a transformer language model,” in ISMIR, 2023, pp. 810–816

    work page 2023

  7. [7]

    Predictability and uncertainty in the pleasure of music: a reward for learning?

    B. P. Gold, M. T. Pearce, E. Mas-Herrero, A. Dagher, and R. J. Zatorre, “Predictability and uncertainty in the pleasure of music: a reward for learning?” Journal of Neuroscience , vol. 39, no. 47, pp. 9397–9409, 2019

    work page 2019

  8. [8]

    Guided music synthesis with variable markov oracle,

    C.-i. Wang and S. Dubnov, “Guided music synthesis with variable markov oracle,” in AAAI, vol. 10, no. 5, 2014, pp. 55–62

    work page 2014

Show all 49 references
  1. [9]

    Developing and evaluating computational models of musical style,

    T. Collins, R. Laney, A. Willis, and P. H. Garthwaite, “Developing and evaluating computational models of musical style,” AI EDAM , vol. 30, no. 1, pp. 16–43, 2016

    2016

  2. [10]

    Controlling surprisal in music generation via information content curve matching,

    M. R. Bjare, S. Lattner, and G. Widmer, “Controlling surprisal in music generation via information content curve matching,” in ISMIR, 2024

    2024

  3. [11]

    Audio oracle: A new algorithm for fast learning of audio structures,

    S. Dubnov, G. Assayag, and A. Cont, “Audio oracle: A new algorithm for fast learning of audio structures,” in ICMC, 2007, pp. 224–227

    2007

  4. [12]

    Deep generative models of music expectation,

    N. L. Masclef and T. A. Keller, “Deep generative models of music expectation,” NeurIPS ML for Audio Workshop 2023 , 2023

    2023

  5. [13]

    Retrieving musical information from neural data: how cognitive features enrich acoustic ones

    E. Abrams, E. M. Vidal, C. Pelofi, and P. Ripoll ´es, “Retrieving musical information from neural data: how cognitive features enrich acoustic ones.” in ISMIR, 2022, pp. 160–168

    2022

  6. [14]

    Differentiable short-term models for efficient online learning and prediction in monophonic music,

    M. R. Bjare, S. Lattner, and G. Widmer, “Differentiable short-term models for efficient online learning and prediction in monophonic music,” Trans. Int. Soc. Music. Inf. Retr . , vol. 5, no. 1, p. 190, 2022

    2022

  7. [15]

    On the typicality of musical sequences,

    M. R. Bjare and S. Lattner, “On the typicality of musical sequences,” in ISMIR Late Breaking and Demo Papers , 2022

    2022

  8. [16]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in NeurIPS, 2017, pp. 5998–6008

    2017

  9. [17]

    Long-form music generation with latent diffusion,

    Z. Evans, J. D. Parker, C. Carr, Z. Zukowski, J. Taylor, and J. Pons, “Long-form music generation with latent diffusion,” in ISMIR, 2024

    2024

  10. [18]

    GIVT: generative infinite- vocabulary transformers,

    M. Tschannen, C. Eastwood, and F. Mentzer, “GIVT: generative infinite- vocabulary transformers,” CoRR, vol. abs/2312.02116, 2023

    2023 arXiv

  11. [19]

    Music2latent: Consistency autoencoders for latent audio compression,

    M. Pasini, S. Lattner, and G. Fazekas, “Music2latent: Consistency autoencoders for latent audio compression,” in ISMIR, 2024

    2024

  12. [20]

    Unsupervised statistical learning underpins computational, behavioural, and neural manifestations of musical expectation,

    M. T. Pearce, M. H. Ruiz, S. Kapasi, G. A. Wiggins, and J. Bhattacharya, “Unsupervised statistical learning underpins computational, behavioural, and neural manifestations of musical expectation,” NeuroImage, vol. 50, no. 1, pp. 302–313, 2010

    2010

  13. [21]

    Cortical encoding of melodic expectations in human temporal cortex,

    G. M. Di Liberto, C. Pelofi, R. Bianco, P. Patel, A. D. Mehta, J. L. Herrero, A. De Cheveign ´e, S. Shamma, and N. Mesgarani, “Cortical encoding of melodic expectations in human temporal cortex,” Elife, vol. 9, p. e51784, 2020

    2020

  14. [22]

    Predictive uncertainty in auditory sequence processing,

    N. C. Hansen and M. T. Pearce, “Predictive uncertainty in auditory sequence processing,” Frontiers in psychology , vol. 5, p. 1052, 2014

    2014

  15. [23]

    Pupil responses to pitch deviants reflect predictability of melodic sequences,

    R. Bianco, L. E. Ptasczynski, and D. Omigie, “Pupil responses to pitch deviants reflect predictability of melodic sequences,” Brain and Cognition, vol. 138, p. 103621, 2020

    2020

  16. [24]

    Statistical learning of melodic patterns influences the brain’s response to wrong notes,

    T. Moldwin, O. Schwartz, and E. S. Sussman, “Statistical learning of melodic patterns influences the brain’s response to wrong notes,”Journal of cognitive neuroscience , vol. 29, no. 12, pp. 2114–2122, 2017

    2017

  17. [25]

    Detecting change in stochastic sound sequences,

    B. Skerritt-Davis and M. Elhilali, “Detecting change in stochastic sound sequences,” PLoS Comput. Biol. , vol. 14, no. 5, 2018

    2018

  18. [26]

    A model for statistical regularity extraction from dynamic sounds,

    B. Skerritt-Davis and M. Elhilali, “A model for statistical regularity extraction from dynamic sounds,” Acta Acustica united with Acustica , vol. 105, no. 1, pp. 1–4, 2019

    2019

  19. [27]

    Consistency models,

    Y . Song, P. Dhariwal, M. Chen, and I. Sutskever, “Consistency models,” in ICML, vol. 202. PMLR, 2023, pp. 32 211–32 252

    2023

  20. [28]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” in NeurIPS, 2020, pp. 12 449–12 460

    2020

  21. [29]

    Contrastive learning of musical representations,

    J. Spijkervet and J. A. Burgoyne, “Contrastive learning of musical representations,” in ISMIR, 2021, pp. 673–681

    2021

  22. [30]

    Jukebox: A generative model for music,

    P. Dhariwal, H. Jun, C. Payne, J. W. Kim, A. Radford, and I. Sutskever, “Jukebox: A generative model for music,” CoRR, vol. abs/2005.00341, 2020

    2005 arXiv

  23. [31]

    SoundStream: An end-to-end neural audio codec,

    N. Zeghidour, A. Luebs, A. Omran, J. Skoglund, and M. Tagliasacchi, “SoundStream: An end-to-end neural audio codec,” IEEE ACM Trans. Audio Speech Lang. Process. , vol. 30, pp. 495–507, 2022

    2022

  24. [32]

    High fidelity neural audio compression,

    A. D ´efossez, J. Copet, G. Synnaeve, and Y . Adi, “High fidelity neural audio compression,” Trans. Mach. Learn. Res. , vol. 2023, 2023

    2023

  25. [33]

    High- fidelity audio compression with improved RVQGAN,

    R. Kumar, P. Seetharaman, A. Luebs, I. Kumar, and K. Kumar, “High- fidelity audio compression with improved RVQGAN,” inNeurIPS, 2023

    2023

  26. [34]

    Language models are unsupervised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever, “Language models are unsupervised multitask learners,” 2019

    2019

  27. [35]

    Medleydb 2.0: New data and a system for sustainable data collection,

    R. M. Bittner, J. Wilkins, H. Yip, and J. P. Bello, “Medleydb 2.0: New data and a system for sustainable data collection,” ISMIR Late Breaking and Demo Papers , vol. 36, 2016

    2016

  28. [36]

    The MUSDB18 corpus for music separation,

    Z. Rafii, A. Liutkus, F.-R. St ¨oter, S. I. Mimilakis, and R. Bittner, “The MUSDB18 corpus for music separation,” Dec. 2017. [Online]. Available: https://doi.org/10.5281/zenodo.1117372

    2017 doi

  29. [37]

    Kishi, N

    H. Kishi, N. Polouliakh, and T. Akama, “Deep12,” https://www.sonycsl.co.jp/tokyo/14621

  30. [38]

    Roformer: Enhanced transformer with rotary position embedding,

    J. Su, M. H. M. Ahmed, Y . Lu, S. Pan, W. Bo, and Y . Liu, “Roformer: Enhanced transformer with rotary position embedding,” Neurocomput- ing, vol. 568, p. 127063, 2024

    2024

  31. [39]

    FlashAttention: Fast and memory-efficient exact attention with io-awareness,

    T. Dao, D. Y . Fu, S. Ermon, A. Rudra, and C. R ´e, “FlashAttention: Fast and memory-efficient exact attention with io-awareness,” CoRR, vol. abs/2205.14135, 2022

    2022 arXiv

  32. [40]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in ICLR (Poster), 2015

    2015

  33. [41]

    Conchord: an application for generating musical harmony by navigating in a perceptually motivated tonal interval space,

    G. Bernardes, D. Cocharro, C. Guedes, and M. Davies, “Conchord: an application for generating musical harmony by navigating in a perceptually motivated tonal interval space,” inCMMR, 2015, pp. 71–86

    2015

  34. [42]

    A multi-level tonal interval space for modelling pitch relatedness and musical consonance,

    G. Bernardes, D. Cocharro, M. Caetano, C. Guedes, and M. E. Davies, “A multi-level tonal interval space for modelling pitch relatedness and musical consonance,” Journal of New Music Research , vol. 45, no. 4, pp. 281–294, 2016

    2016

  35. [43]

    A. A. Moles, Information theory and esthetic perception . The Univer- sity of Illinois Press, Urbana and London, 1966

    1966

  36. [44]

    Perceptual scaling of synthesized musical timbres: Common dimensions, specificities, and latent subject classes,

    S. McAdams, S. Winsberg, S. Donnadieu, G. De Soete, and J. Krim- phoff, “Perceptual scaling of synthesized musical timbres: Common dimensions, specificities, and latent subject classes,” Psychological re- search, vol. 58, pp. 177–192, 1995

    1995

  37. [45]

    Maximum filter vibrato suppression for onset detection,

    S. B ¨ock and G. Widmer, “Maximum filter vibrato suppression for onset detection,” in DAFx, vol. 7. Citeseer, 2013, p. 4

    2013

  38. [46]

    ITU- R BS.1770-4,

    ITU Radiocommunication Sector - Broadcasting service (sound), “ITU- R BS.1770-4,” https://www.itu.int/rec/R-REC-BS.1770/en

  39. [47]

    Inves- tigating the cortical tracking of speech and music with sung speech,

    G. Cantisani, A. Chalehchaleh, G. Di Liberto, and S. Shamma, “Inves- tigating the cortical tracking of speech and music with sung speech,” in INTERSPEECH. ISCA, 2023, pp. 5157–5161

    2023

  40. [48]

    The multivariate temporal response function (mtrf) toolbox: a matlab toolbox for relating neural signals to continuous stimuli,

    M. J. Crosse, G. M. Di Liberto, A. Bednar, and E. C. Lalor, “The multivariate temporal response function (mtrf) toolbox: a matlab toolbox for relating neural signals to continuous stimuli,” Frontiers in human neuroscience, vol. 10, p. 604, 2016

    2016

  41. [49]

    Neural signatures of musical and linguistic interactions during natural song listening,

    G. Cantisani, S. Shamma, and G. M. Di Liberto, “Neural signatures of musical and linguistic interactions during natural song listening,” Hal preprint, 2024

    2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.