Pith. sign in

REVIEW 4 major objections 3 minor 60 references

Prosodic Structure Beyond Lexical Content: A Study of Self-Supervised Learning

T0 review · 4 major / 3 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 The paper argues that self-supervised learning on pitch, energy, and voice activity alone reveals predictable prosodic structure at multiple timescales, independent of lexical content.

desk verdict The mask-timescale result is real and worth publishing, but the MPM-vs-CWT headline relies on a misconfigured classical baseline and needs a fix before acceptance. read the letter →

arxiv 2506.02584 v1 pith:4BZWNLZT submitted 2025-06-03 cs.CL cs.AIeess.AS

Sarenne Wallbridge , Christoph Minixhofer , Catherine Lai , Peter Bell This is my paper
classification cs.CLcs.AIeess.AS
keywords self-supervisedlearningprosodyprosodicstructuremaskedmodelemotionrecognitionprominencedetectionboundarysyllablesegmentation
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 asks whether prosody—intonation, loudness, and rhythm considered apart from the words being said—has predictable structure of its own. To test this, the authors train a Masked Prosody Model (MPM), a self-supervised transformer that learns to reconstruct corrupted sequences of pitch, energy, and voice activity, with no lexical input. Probing the learned representations shows they predict syllable counts, word prominence, prosodic boundaries, and emotion, outperforming both untransformed acoustic features and a classical wavelet-based prosody encoding on most tasks. The authors conclude that prosody exhibits systematicity independent of lexical content, and that the temporal scale of the training corruption controls what structure is captured: small masks help local tasks, large masks help emotion, and random masking yields a generally useful representation. If the claim is right, prosodic structure is a real, learnable object on its own timescales, not merely a by-product of word content.

What carries the argument

The central object is the Masked Prosody Model (MPM): a Conformer encoder with 16 blocks that takes three parallel sequences—F0 from the WORLD vocoder, RMS energy from mel spectrogram frames, and voice activity—at roughly 10ms resolution, normalizes each contour per utterance, quantizes each stream into codebooks of size c, masks random spans of length m (with a random-mask variant drawing m uniformly from 1 to 128) until about half the signal remains, and is trained to reconstruct the masked quantized symbols via categorical cross-entropy losses. Features from the 8th Conformer layer are extracted and fed to linear or Conformer probes, so the mechanism that carries the argument is the interaction between masking scale and the probe's ability to read off local versus global structure from the learned features.

What would settle it

Train the same MPM with identical masking but on pitch, energy, and voice activity sequences whose frames are randomly permuted within each utterance, so value distributions are preserved but temporal order is destroyed. If linear probes from the shuffled model match the original MPM's boundary and emotion scores, the apparent systematicity is an artifact of feature statistics rather than temporal structure; if scores collapse toward chance, temporal prosodic contours are what the representations encode.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a self-supervised model trained only on 10-millisecond contours of pitch (F0), RMS energy, and voice activity learns representations that carry usable information about prosodic labels at several timescales. Using linear probes, the random-mask MPM consistently beats untransformed pitch/energy/VAD features and the CWT hierarchical encoding; the largest relative gains appear for boundary detection and emotion recognition. Under the stronger Conformer probe, the prosody-only MPM is competitive with and sometimes slightly better than full-speech SSL models (wav2vec and HuBERT) at BURNC boundary and prominence detection, while full-speech models win on syllable segmentation and emotion. The paper takes this as evidence that prosody exhibits predictable structure independently of lexical content, that the SSL objective's mask timescale determines which structures are encoded, and that unconstrained learned structure is more valuable than the fixed hierarchical structure of CWT for abstract, longer-range labels.

Load-bearing premise

The comparison that drives the 'SSL beats classical structure' conclusion assumes the CWT baseline is a faithful implementation of the hierarchical prosody encoding from [21]; if the CWT features used here are misconfigured—their BURNC F1 scores of 0.39 and 0.02 sit far below the original method's 0.85 and 0.59—the conclusion rests on a weakened baseline.

Editorial extensions

If this is right

  • Prosody can be studied and modeled without any lexical or phonetic input; MPM-style features give downstream systems a route to prosodic information in delexicalized or low-resource settings.
  • The masking timescale is a control knob: short masks preserve local syllable-scale cues, long masks emphasize utterance-level emotion, and random masking is a simple generic choice that works across tasks.
  • For phrasal boundary and prominence detection, the three prosodic streams carry most of the needed information, since MPM roughly matches or exceeds full-speech SSL representations under the stronger probe.
  • Emotion recognition benefits from large-mask training, but full-speech SSL still outperforms prosody-only representations, showing that additional acoustic channels beyond pitch, energy, and voice activity matter even when lexical content is held constant.

Reading between the lines

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

  • A fairer test of the CWT comparison would reimplement the original rule-based CWT labelling on BURNC; if boundary F1 returns to roughly 0.59 and prominence to 0.85, the reported MPM advantage over constrained hierarchical structure would shrink, even though the existence of prosodic systematicity could still stand.
  • The random-span-masking recipe is likely to transfer to other continuous non-lexical contours—for instance, turn-taking or sarcasm prediction—where temporal span length is unknown in advance, because random masking avoids committing to a single scale.
  • Extending MPM with explicit duration features would test whether tempo, beyond pitch and loudness, adds an independent layer of prosodic predictability; the current design leaves that channel implicit in voice activity.
  • Because RAVDESS uses the same two sentences across emotions, the gap between full-speech and prosody-only SSL on emotion cannot be lexical; identifying which non-pitch, non-energy, non-VAD cues close that gap is a concrete next experiment.
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

4 major / 3 minor

Summary. The paper introduces the Masked Prosody Model (MPM), a self-supervised model trained to reconstruct masked sequences of pitch, energy, and voice-activity contours extracted from speech. The authors investigate how the masking strategy (mask length and codebook size) affects the utility of the learned representations, probing them on syllable segmentation (TIMIT), prominence and boundary detection (BURNC), and emotion recognition (RAVDESS) with linear and Conformer probes. They report that MPM representations outperform untransformed features and a continuous wavelet transform (CWT) baseline, and that a random-mask strategy balances local and global structure. The paper argues that prosody exhibits predictable structure independent of lexical content, and that SSL-learned structures are more valuable than constrained classical hierarchical encodings.

Significance. If the results hold, the paper makes a useful contribution to prosody representation learning by showing that SSL on low-level acoustic correlates of prosody can capture task-relevant structure and by analyzing the role of masking timescale. The release of code and pretrained weights supports reproducibility. The evaluation uses external benchmark labels, so the core comparisons are not circular. However, the central comparative claim against classical CWT encoding rests on a baseline that appears misconfigured, and the reported numbers lack uncertainty estimates, so the strength of the conclusions is currently disproportionate to the evidence.

major comments (4)
  1. [§3.3, Table 1a] The CWT baseline is not a faithful implementation of the classical hierarchical encoding from [21], and this undermines the claim that SSL-learned structures are more valuable than constrained classical structures. Section 2.2 describes [21] as using CWT across F0, energy, and duration features, but Section 3.3 states that CWT representations are generated from the same input features as MPMs—pitch, energy, and VAD—omitting duration. The resulting boundary detection F1 of 0.02 in Table 1a is far below the 0.59 reported for [21]'s own method in Section 3.2, and is near the floor for a binary task, indicating a configuration error or an incomplete feature set. The paper's conclusion in Section 5 that 'the additional structures learned through SSL are more valuable' relies on this weakened comparison. Please reproduce the CWT baseline with duration features and, ideally, the original rule-based classifier from [21], or clearly restrict the comparison to the specific input features used and temper the claim.
  2. [Table 1a/1b] The paper reports no confidence intervals, error bars, or significance tests, although several key comparisons are numerically small. For example, in Table 1a the difference between MPM-mask-16 and MPM-random on prominence is 0.58 vs 0.57, and the boundary difference between MPM-mask-4 and MPM-mask-16 is 0.27 vs 0.28. The claim that MPM 'consistently outperforms' untransformed and CWT features, and the task-dependent ranking of mask sizes, would be considerably strengthened by reporting variance across the 5-fold cross-validation used for BURNC, or by including standard errors for the other datasets. Without this, the reader cannot assess whether the observed ordering is reliable.
  3. [§3.3] Hyperparameters were selected using downstream tasks, which introduces a form of selection bias that should be disclosed and analyzed. The paper states that in preliminary experiments, mask sizes and codebook sizes were tested and the 'best performance for all tasks' determined c=128. If these preliminary experiments used the same downstream labels reported in Table 1, the reported results are not a purely self-supervised evaluation of the learned representations; they are partially tuned to those labels. Please describe the selection procedure, which tasks/splits were used, and whether the reported numbers are on held-out test folds or on the same folds used for selection.
  4. [Abstract, §4.1, Table 1a] The claim that MPM provides 'the most value for labels involving longer-term structures, like emotion recognition' should be qualified by the low absolute performance on RAVDESS. With eight emotion classes, chance accuracy is 12.5%; the linear-probe weighted accuracy of the random-mask MPM is 0.24 and the Conformer probe reaches 0.37. Although the relative gains over untransformed features are sizable (0.24 vs 0.10 linear), the representations capture only a small fraction of the label variance, and the near-chance absolute numbers suggest that the emotion-recognition result should be interpreted cautiously. Please report the chance level explicitly and discuss the practical significance of the relative gains.
minor comments (3)
  1. [Throughout] There are several typographical issues, including 'V AD' instead of 'VAD' in Table 1a, 'BURNCtasks' missing a space, and 'mask sizecodebook size' in the Figure 1 caption. These should be cleaned up.
  2. [§3.3] The description of the CWT implementation is underspecified: it does not state the number of wavelet scales, the specific wavelet family, or whether the features are used as raw coefficients or aggregated. This makes it difficult to assess whether the poor CWT performance in Table 1a is due to the transformation itself or to the probe setup.
  3. [§5] The future-work sentence 'more explicit inclusion of duration information' is an important caveat given the paper's emphasis on comparing against a classical duration-aware baseline; consider moving this limitation earlier and relating it directly to the CWT comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SSL training is label-free and downstream probes use external benchmark annotations, so the claimed prosodic-structure findings do not reduce to fitted inputs or self-citation.

full rationale

The central derivation (Sections 3.1-4.2) trains a Masked Prosody Model to reconstruct corrupted pitch, energy, and voice-activity sequences from LibriTTS with a categorical cross-entropy loss; no downstream label enters the pretraining objective. Utility is measured with linear and Conformer probes on external corpora (TIMIT, BURNC, RAVDESS) whose labels are independent of the pretraining data and of the model parameters. The comparison baselines (untransformed features, CWT, wav2vec, HuBERT, mel spectrogram) are evaluated on the same external labels, so MPM's relative gains are not an artifact of definitional equivalence. The only selection-on-evaluation concern is that codebook size and mask-size ranges were chosen in preliminary experiments using downstream performance (Section 3.3: 'In preliminary experiments, m∈{1,2,4,8,16,32,64,128} and c∈{4,8,16,32,64,128,512} were tested; the smallest mask sizes consistently performed poorly and the best performance for all tasks was achieved with c=128'), but this is hyperparameter selection, not a fitted parameter being renamed as a prediction; the final reported mask sizes are compared among themselves and against untrained baselines, and the random-mask strategy is motivated independently by span-masking work [51] before evaluation. The CWT baseline concern raised by the skeptic (Section 3.3 says CWT is generated from the same input features as MPMs, omitting duration features, while [21] used duration; boundary F1=0.02 vs 0.59 in [21]) is a baseline-fidelity/correctness risk, not circularity. No load-bearing self-citation chain appears: citations to the authors' earlier work (e.g., [3], [10], [53]) are background or comparison material, not uniqueness theorems or definitions of the target result. Thus the derivation is self-contained against external benchmarks and merits a non-circularity score of 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claims rest on the choice of acoustic features and the probe methodology. No new physical or theoretical entities are introduced. The main free parameters are the SSL training hyperparameters (codebook size, mask proportion), one of which was tuned on downstream tasks.

free parameters (3)
  • Codebook size c = 128
    Selected in preliminary experiments as giving the best downstream task performance for all tasks (Section 3.3). This is a hyperparameter tuned on the evaluation labels, not derived from first principles.
  • Mask target proportion = 50±5% of input signal remains
    Chosen by hand as the proportion of input signal to remain unmasked; it sets the difficulty of the SSL pretext task and is not swept.
  • Random mask range = 1 to 128 frames
    For the random masking strategy, m is sampled uniformly between 1 and 128; this range is chosen by hand as a compromise between local and global corruption.
assumptions (5)
  • domain assumption WORLD vocoder provides reliable F0 and voice activity estimates at 10ms resolution.
    The input features are extracted with WORLD [29]; if F0 estimation is unreliable, the learned representations would not reflect prosody.
  • domain assumption RMS energy from mel spectrogram frames adequately represents loudness.
    Energy is computed as RMS of each Mel Spectrogram frame (Section 3.1); this is a standard but lossy proxy for perceived loudness.
  • domain assumption Per-utterance normalization of feature contours preserves prosodic structure while removing speaker-specific absolute values.
    Features are normalized across the utterance to allow encoding of unseen speakers; this may discard absolute pitch/energy information relevant to emotion recognition.
  • domain assumption The downstream labels (syllable count, prominence, boundaries, emotion) are at least partially determined by prosodic cues at the assumed timescales.
    The task selection is motivated by literature, but if a label is not actually prosody-driven at that timescale, the interpretation of mask-size effects would be affected.
  • standard math Linear probing is a valid measure of the accessibility of information in a representation.
    The paper uses linear probes [33] to compare representations; this is a standard methodology but not guaranteed to reflect all usable information.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Prosodic Structure Beyond Lexical Content: A Study of Self-Supervised Learning." pith.science (2026). https://pith.science/paper/4BZWNLZT

@misc{pith2026250602584,
  author       = {Pith},
  title        = {Pith review of: Prosodic Structure Beyond Lexical Content: A Study of Self-Supervised Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4BZWNLZT}},
  note         = {Machine review of arXiv:2506.02584}
}
read the original abstract

People exploit the predictability of lexical structures during text comprehension. Though predictable structure is also present in speech, the degree to which prosody, e.g. intonation, tempo, and loudness, contributes to such structure independently of the lexical content is unclear. This study leverages self-supervised learning (SSL) to examine the temporal granularity of structures in the acoustic correlates of prosody. Representations from our proposed Masked Prosody Model can predict perceptual labels dependent on local information, such as word boundaries, but provide the most value for labels involving longer-term structures, like emotion recognition. Probing experiments across various perceptual labels show strong relative gains over untransformed pitch, energy, and voice activity features. Our results reveal the importance of SSL training objective timescale and highlight the value of complex SSL-encoded structures compared to more constrained classical structures.

Figures

Figures reproduced from arXiv: 2506.02584 by the authors.

Figure 1
Figure 1. Architecture of the Masked Prosody Model. To evaluate model representations, features from the 8th MPM Conformer layer are extracted as intermediary network representations are robust across tasks [32]. When aggregating at word or utterance levels, we compute the mean and maximum features for the target unit and concatenate them. Representa￾tions are evaluated using both linear probing [33] by training a linear clas… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 57 canonical work pages

  1. [21]

    wav2vec: Unsupervised pre-training for speech recognition,

    S. Schneider, A. Baevski, R. Collobert, and M. Auli, “wav2vec: Unsupervised pre-training for speech recognition,” arXiv:1904.05862, 2019

    arXiv 1904

  2. [1]

    Introduction Cognitive science theories often describe human language comprehension as a process of leveraging the predictable struc- ture of linguistic signals, i.e., theirsystematicity, to generate expectations about the upcoming signal [1]. Though empirical support stems primarily from written language comprehension, listeners also exploit systematicit...

    work page

  3. [2]

    Background 2.1. Self-supervised learning SSL mechanisms encode the structures of their training data by learning to reverse a corruption function—e.g., removing fu- ture, past, or intermediate context. Without a need for external labels, SSL can exploit large amounts of unlabeled data. The resulting representations are useful for downstream tasks with lim...

    arXiv 2025

  4. [3]

    Experiments Inspired by masked language models for text, we introduce a Masked Prosody Model (MPM) that learns to reconstruct cor- rupted sequences of pitch, loudness and voice activity during pretraining [12]. We investigate the effects of different corrup- tion strategies on the utility of resulting representations across a set of downstream tasks that ...

    work page

  5. [4]

    Results 4.1. Effect of Corruption Strategy In the MPM section of Table 1a, we use a linear probe to directly compare the performance achieved from models withc= 128 andm∈ {4,16,128}across the downstream tasks. We find that syllable segmentation benefits from smaller mask sizes; larger mask sizes may not encode sufficiently fine-grained lo- cal structures....

    work page

  6. [5]

    However, the extent to which this predictive capacity is contingent on the structure of lexical information or of its acoustic realisation is unclear

    and conversational turns [6]). However, the extent to which this predictive capacity is contingent on the structure of lexical information or of its acoustic realisation is unclear. To better understand how listeners generate expectations about upcom- ing speech, we investigate systematicity in non-lexical features of prosody—relative pitch (F0) and loudn...

    work page

  7. [6]

    How much does prosody help turn- taking? investigations using voice activity projection models,

    E. Ekstedt and G. Skantze, “How much does prosody help turn- taking? investigations using voice activity projection models,” in SIGDIAL, 2022

    work page 2022

  8. [7]

    Discussion & Conclusions We find that self-supervised methods can produce useful rep- resentations of the acoustic correlates of prosody, indicating that prosody exhibits predictable structure—systematicity— independently of lexical content. Our comparisons of corrup- tion strategies reveal structure across timescales: syllable seg- mentation benefits fro...

    work page

Show all 60 references
  1. [8]

    A probabilistic Earley parser as a psycholinguistic model,

    J. Hale, “A probabilistic Earley parser as a psycholinguistic model,” inNAACL, 2001

    2001

  2. [9]

    Metrical expectations from preceding prosody influence percep- tion of lexical stress

    M. Brown, A. P. Salverda, L. C. Dilley, and M. K. Tanenhaus, “Metrical expectations from preceding prosody influence percep- tion of lexical stress.”JEPHPP, 2015

    2015

  3. [10]

    Quantifying the perceptual value of lexical and non-lexical channels in speech,

    S. Wallbridge, P. Bell, and C. Lai, “Quantifying the perceptual value of lexical and non-lexical channels in speech,” inInter- speech, 2023

    2023

  4. [11]

    Intonation facilitates prediction of focus even in the presence of lexical tones,

    M. Ip and A. Cutler, “Intonation facilitates prediction of focus even in the presence of lexical tones,” inInterspeech, 2017

    2017

  5. [12]

    How long is the sentence? Prediction and prosody in the online processing of language,

    F. Grosjean, “How long is the sentence? Prediction and prosody in the online processing of language,” inLinguistics, 1983

    1983

  6. [13]

    The language of proteins: NLP, machine learning & protein sequences,

    D. Ofer, N. Brandes, and M. Linial, “The language of proteins: NLP, machine learning & protein sequences,”CSBJ, 2021

    2021

  7. [14]

    How prosody is both mandatory and optional,

    A. Cutler and J. McQueen, “How prosody is both mandatory and optional,” inAbove and beyond the segments, 2014

    2014

  8. [15]

    For Wav2Vec and HuBERT, we use the model checkpoints trained on≈900h of LibriSpeech, which is comparable to the amount of data seen by the MPM

    and Mel Spectrograms. For Wav2Vec and HuBERT, we use the model checkpoints trained on≈900h of LibriSpeech, which is comparable to the amount of data seen by the MPM. The linear and Conformer (2 Conformer blocks, 5M pa- rameters) classification probes are trained for 1000 steps...

  9. [16]

    What do you mean, you’re uncertain? The interpretation of cue words and rising intonation in dialogue,

    C. Lai, “What do you mean, you’re uncertain? The interpretation of cue words and rising intonation in dialogue,” inInterspeech, 2010

    2010

  10. [17]

    Delexicalised auditory priming of implicit prosody,

    J. Mills, “Delexicalised auditory priming of implicit prosody,” in Speech Prosody, 2020

    2020

  11. [18]

    Towards automatic detection of reported speech in dialogue using prosodic cues

    A. Cervone, C. Lai, S. Pareti, and P. Bell, “Towards automatic detection of reported speech in dialogue using prosodic cues.” in Interspeech, 2015

    2015

  12. [19]

    DeCAF: A deep convolutional activation fea- ture for generic visual recognition,

    J. Donahueet al., “DeCAF: A deep convolutional activation fea- ture for generic visual recognition,” inPMLR, 2014

    2014

  13. [20]

    BERT: Pre- training of deep bidirectional transformers for language under- standing,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre- training of deep bidirectional transformers for language under- standing,” inNAACL, 2019

    2019

  14. [22]

    HuBERT: Self-supervised speech repre- sentation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “HuBERT: Self-supervised speech repre- sentation learning by masked prediction of hidden units,”TASLP, 2021

    2021

  15. [23]

    ProsAudit, a prosodic benchmark for self- supervised speech models,

    M. de Seysselet al., “ProsAudit, a prosodic benchmark for self- supervised speech models,” inInterspeech, 2023

    2023

  16. [24]

    Towards end-to-end prosody transfer for expressive speech synthesis with Tacotron,

    R. Skerry-Ryanet al., “Towards end-to-end prosody transfer for expressive speech synthesis with Tacotron,” inICML, 2018

    2018

  17. [25]

    Disentangling prosody representations with unsupervised speech reconstruction,

    L. Qu, T. Li, C. Weber, T. Pekarek-Rosin, F. Ren, and S. Wermter, “Disentangling prosody representations with unsupervised speech reconstruction,”TASLP, 2023

    2023

  18. [26]

    Towards paralinguistic-only speech representations for end-to- end speech emotion recognition,

    G. Ioannides, M. Owen, A. Fletcher, V . Rozgic, and C. Wang, “Towards paralinguistic-only speech representations for end-to- end speech emotion recognition,” inInterspeech, 2023

    2023

  19. [27]

    Do prosody transfer models transfer prosody?

    A. T. Sigurgeirsson and S. King, “Do prosody transfer models transfer prosody?” inICASSP, 2023

    2023

  20. [28]

    Hierarchical rep- resentation and estimation of prosody using continuous wavelet transform,

    A. Suni, J. Simko, D. Aalto, and M. Vainio, “Hierarchical rep- resentation and estimation of prosody using continuous wavelet transform,”Comput. Speech Lang., 2017

    2017

  21. [29]

    Prosodic promi- nence and boundaries in sequence-to-sequence speech synthesis,

    A. Suni, S. Kakouros, M. Vainio, and J. Simko, “Prosodic promi- nence and boundaries in sequence-to-sequence speech synthesis,” inSpeech Prosody, 2020

    2020

  22. [30]

    E. O. Selkirk,Phonology and Syntax: The Relation between Sound and Structure. MIT Press, 1984

    1984

  23. [31]

    Beckman,Stress And Non-Stress Accent

    M. Beckman,Stress And Non-Stress Accent. De Gruyter, 1986

    1986

  24. [32]

    Automatic detection of sentence prominence in speech using predictability of word-level acoustic features,

    S. Kakouros and O. R ¨as¨anen, “Automatic detection of sentence prominence in speech using predictability of word-level acoustic features,” inInterspeech, 2015

    2015

  25. [33]

    Learn- ing de-identified representations of prosody from raw audio,

    J. Weston, R. Lenain, U. Meepegama, and E. Fristed, “Learn- ing de-identified representations of prosody from raw audio,” in ICML, 2021

    2021

  26. [34]

    PMI-Masking: Principled masking of correlated spans,

    Y . Levineet al., “PMI-Masking: Principled masking of correlated spans,” inICLR, 2020

    2020

  27. [35]

    Conformer: Convolution-augmented transformer for speech recognition,

    A. Gulatiet al., “Conformer: Convolution-augmented transformer for speech recognition,” inInterspeech, 2020

    2020

  28. [36]

    WORLD: a vocoder- based high-quality speech synthesis system for real-time applica- tions,

    M. Morise, F. Yokomori, and K. Ozawa, “WORLD: a vocoder- based high-quality speech synthesis system for real-time applica- tions,”IEICE Transactions, 2016

    2016

  29. [37]

    Classification of prosodic events using quantized contour modeling,

    A. Rosenberg, “Classification of prosodic events using quantized contour modeling,” inNAACL, 2010

    2010

  30. [38]

    Pitch contour shape matters in mem- ory,

    A. E. Kimball and J. Cole, “Pitch contour shape matters in mem- ory,” inICSP, 2016

    2016

  31. [39]

    How transferable are features in deep neural networks?

    J. Yosinski, J. Clune, Y . Bengio, and H. Lipson, “How transferable are features in deep neural networks?” inNeurIPS, 2014

    2014

  32. [40]

    Understanding intermediate layers using linear classifier probes,

    G. Alain and Y . Bengio, “Understanding intermediate layers using linear classifier probes,”ICLR Workshop, 2017

    2017

  33. [41]

    S. G. Parker,Quantifying the sonority hierarchy. University of Massachusetts Amherst, 2002

    2002

  34. [42]

    Young infants’ retention of information about syllables,

    P. W. Jusczyk, L. J. Kennedy, and A. M. Jusczyk, “Young infants’ retention of information about syllables,”Infant Behavior and De- velopment, 1995

    1995

  35. [43]

    The influence of different prosodic cues on word segmentation,

    T. Matzinger, N. Ritt, and W. T. Fitch, “The influence of different prosodic cues on word segmentation,”Front. in Psychology, 2021

    2021

  36. [44]

    Prosody in context: A review,

    J. Cole, “Prosody in context: A review,”Language, Cognition and Neuroscience, 2015

    2015

  37. [45]

    Segmenting into adequate units for automatic recognition of emotion-related episodes: a speech-based approach,

    A. Batliner, S. Steidl, D. Seppi, and B. Schuller, “Segmenting into adequate units for automatic recognition of emotion-related episodes: a speech-based approach,”AHCI, 2010

    2010

  38. [46]

    TIMIT acoustic phonetic continuous speech cor- pus,

    J. S. Garofolo, “TIMIT acoustic phonetic continuous speech cor- pus,”LDS, 1993

    1993

  39. [47]

    Convex weighting criteria for speaking rate estimation,

    Y . Jiao, V . Berisha, M. Tu, and J. Liss, “Convex weighting criteria for speaking rate estimation,”TASLP, 2015

    2015

  40. [48]

    Automatic recognition of suprasegmentals in speech,

    J. Yuan, N. Ryant, X. Cai, K. Church, and M. Liber- man, “Automatic recognition of suprasegmentals in speech,” arXiv:2108.01122, 2021

    2021 arXiv

  41. [49]

    Pre-linguistic segmenta- tion of speech into syllable-like units,

    O. R ¨as¨anen, G. Doyle, and M. C. Frank, “Pre-linguistic segmenta- tion of speech into syllable-like units,”Cognition, vol. 171, 2018

    2018

  42. [50]

    Contribution of prosody to the segmentation and storage of

    O. Bagou, C. Fougeron, and U. Frauenfelder, “Contribution of prosody to the segmentation and storage of” words” in the acqui- sition of a new mini-language,” inSpeech Prosody, 2002

    2002

  43. [51]

    The Boston University radio news corpus,

    M. Ostendorf, P. J. Price, and S. Shattuck-Hufnagel, “The Boston University radio news corpus,”LDS, 1995

    1995

  44. [52]

    ToBI: a standard for labeling English prosody,

    K. Silvermanet al., “ToBI: a standard for labeling English prosody,” inSLP, 1992

    1992

  45. [53]

    Automatic prosodic event detection using acoustic, lexical, and syntactic evidence,

    S. Ananthakrishnan and S. S. Narayanan, “Automatic prosodic event detection using acoustic, lexical, and syntactic evidence,” TASLP, 2007

    2007

  46. [54]

    Towards low-resource prosodic boundary detection

    B. Ludusan and E. Dupoux, “Towards low-resource prosodic boundary detection.” inSLTU, 2014

    2014

  47. [55]

    A saliency-based auditory atten- tion model with applications to unsupervised prominent syllable detection in speech

    O. Kalinli and S. S. Narayanan, “A saliency-based auditory atten- tion model with applications to unsupervised prominent syllable detection in speech.” inInterspeech, 2007

    2007

  48. [56]

    The Ryerson audio-visual database of emotional speech and song (RA VDESS): A dynamic, multimodal set of facial and vocal expressions in North American English,

    S. R. Livingstone and F. A. Russo, “The Ryerson audio-visual database of emotional speech and song (RA VDESS): A dynamic, multimodal set of facial and vocal expressions in North American English,”PLOS ONE, 2018

    2018

  49. [57]

    Prosodic cues for emotion: analysis with discrete characteriza- tion of intonation,

    H. Cao, ˇS. Be ˇnuˇs, R. C. Gur, R. Verma, and A. Nenkova, “Prosodic cues for emotion: analysis with discrete characteriza- tion of intonation,” inSpeech Prosody, 2014

    2014

  50. [58]

    SpanBERT: Improving pre-training by representing and predicting spans,

    M. Joshi, D. Chen, Y . Liu, D. S. Weld, L. Zettlemoyer, and O. Levy, “SpanBERT: Improving pre-training by representing and predicting spans,”TACL, 2020

    2020

  51. [59]

    LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech,

    H. Zenet al., “LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech,” inInterspeech, 2019

    2019

  52. [60]

    Recognizing emotions in spoken dialogue with hierarchically fused acoustic and lexical features,

    L. Tian, J. Moore, and C. Lai, “Recognizing emotions in spoken dialogue with hierarchically fused acoustic and lexical features,” inSLT, 2016

    2016

Pith tools

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