Pith. sign in

REVIEW 3 major objections 5 minor 44 references

Voice Quality Dimensions as Interpretable Primitives for Speaking Style for Atypical Speech and Affect

T0 review · 3 major / 5 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 Seven voice-quality scores cross languages without retraining.

desk verdict Solid within-domain VQD probe results, but the zero-shot transfer claim is overreaching because the out-of-domain labels are per-speaker severity, not per-utterance VQD ratings. read the letter →

arxiv 2505.21809 v1 pith:DQ4CNBBW submitted 2025-05-27 cs.SD cs.LGeess.AS

Jaya Narain , Vasudha Kowtha , Colin Lea , Lauren Tooley , Dianna Yee , Vikramjit Mitra , Zifang Huang , Miquel Espi Marques
show 3 more authors
Jon Huang Carlos Avendano Shirley Ren
This is my paper
classification cs.SDcs.LGeess.AS
keywords voicequalitydimensionsatypicalspeechdysarthriazero-shottransferembeddingslinearprobesaffectiveinterpretability
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 seven perceptual voice-quality dimensions—intelligibility, imprecise consonants, harsh voice, naturalness, monoloudness, monopitch, and breathiness—can act as a common, interpretable language for describing speaking style. The authors train simple linear probes on top of frozen audio embeddings using speech-language-pathologist ratings of atypical English speech, then apply the same probes without any retraining to Italian dysarthric speech, a different English dysarthric corpus, and acted emotional speech. They report that the probes generalize across all three, and that their per-dimension scores line up with clinical severity and with intuitive emotional profiles. If the claim holds, a single small set of voice-quality descriptors could replace task-specific models for intelligibility, dysarthria, affect, and related speaking-style tasks.

What carries the argument

The central object is the seven-dimensional perceptual voice-quality space: each utterance is represented by seven scalar ratings. The probe is a Lasso regression or logistic-regression model trained to predict one dimension from a frozen embedding; because the embedding is fixed and the probe is linear, the probe is a transparent, lightweight mapping from any utterance to a voice-quality vector. The freezing and linearity are what make zero-shot transfer interpretable: no new task heads, no retraining, and each dimension can be read directly as a score.

What would settle it

Run the frozen-embedding linear probes on a corpus with per-utterance speech-language-pathologist ratings in a language never seen during training, and compare predicted scores directly to those ratings; if the Spearman correlation is near zero for any dimension, such as breathiness in Italian, while the paper's per-speaker severity separation is high, the zero-shot transfer claim is falsified.

Watch

Extended reading notes

Core claim

The paper claims that seven perceptual voice-quality dimensions, predicted by linear probes over frozen audio embeddings, transfer zero-shot from English atypical speech to unseen languages and tasks. On the held-out split of the training corpus, the best probes reach AUC 0.91 for imprecise consonants and 0.82 for intelligibility; summed across dimensions, zero-shot severity classification reaches AUC 0.89 on a dysarthric English corpus and 0.78 on a dysarthric Italian corpus, outperforming an intelligibility-specific baseline even for intelligibility. On acted emotional speech, the same probes yield interpretable profiles, with angry speech scoring low on monoloudness, calm speech low on harshness, and sad speech high on monopitch. The authors read these results as evidence that a seven-dimensional voice-quality vector is a useful interpretable primitive for speaking-style tasks generally.

Load-bearing premise

The transfer claim rests on assuming that the coarse, per-speaker clinical severity scores in the two out-of-domain corpora are a valid proxy for the seven per-utterance voice-quality ratings; if severity is not monotonically related to each dimension, the reported zero-shot AUCs do not establish that the dimensions themselves transferred.

Editorial extensions

If this is right

  • Any utterance can be assigned a seven-dimensional voice-quality vector by a frozen embedding plus linear probe, so unlabeled speech corpora can gain clinical-style annotations automatically.
  • Voice-quality scores can explain what task-specific models are responding to, for example showing that an affect model shifts toward sadness as atypicality increases.
  • Speech technology data curation can use voice-quality profiles to find utterances or speakers with particular style confusions before training recognition systems.
  • Because probes generalize across speech elicitation categories, voice-quality labels can be pooled from commands, read sentences, and spontaneous speech without loss.
  • Cross-language differences in which dimensions drive severity, such as breathiness differentiating severity more for Italian than English, show that the dimensions transfer but their relative weight may need recalibration.

Reading between the lines

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

  • A direct per-utterance evaluation in a held-out language would be stronger than the per-speaker severity validation used here; if collected, it would either confirm or bound the zero-shot claim.
  • The same seven-dimensional probes could condition neural text-to-speech or voice conversion by setting target values for each dimension, producing, for example, a deliberately breathier or less monopitch voice, a control scheme the paper does not test.
  • Because one embedding family transfers best to Italian while an ASR-tuned embedding family does best on English pronunciation dimensions, the pretraining objective may trade language independence against phonetic precision, and comparing more embedding families across languages would test this.
  • The reported affect-model bias implies a fairness test: voice-quality-conditioned emotion classifiers may reduce misclassification of atypical speakers, and this is measurable with existing affect datasets.
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 / 5 minor

Summary. This paper trains linear probes for seven perceptual voice quality dimensions (intelligibility, imprecise consonants, harsh voice, naturalness, monoloudness, monopitch, breathiness) on frozen embeddings from HuBERT, HuBERT-ASR, CLAP, and RawNet3, using the public Speech Accessibility Project dataset. The authors report within-domain Spearman correlations and AUCs, generalization across speech elicitation categories, zero-shot AUCs on two atypical-speech datasets (Eds-vc and EasyCall), and an exploratory analysis on the RAVDESS affect dataset. The central claim is that a single set of seven interpretable VQD primitives transfers zero-shot across languages, recording protocols, and speaking styles without retraining.

Significance. If the zero-shot transfer claims were fully supported, the paper would offer practical interpretable primitives for atypical speech and affect, built from frozen embeddings with simple linear probes. The within-domain methodology is sound: the SAP data are public, splits are speaker-stratified, multiple embeddings are compared, and the focus on interpretable dimensions is timely. However, the zero-shot evaluation for atypical speech relies on coarse per-speaker severity labels with non-independent per-sample scoring, and the affect analysis has no ground-truth VQD labels; the strong generalization claim is therefore not yet established. These issues are addressable with revised evaluation and wording, so the manuscript is promising but needs substantial revision.

major comments (3)
  1. [Abstract and Section 2.4/Figure 5] The zero-shot severity evaluation uses per-speaker labels: Eds-vc has one SLP severity score per speaker (32 speakers, 35,661 samples) and EasyCall has one neurologist TOM score per participant (51 participants, 21,361 samples), yet the paper reports per-sample AUC. This violates sample independence because all utterances from a speaker share the same label and the frozen embeddings encode speaker identity, especially RawNet3, a speaker-ID model. The near-uniform high AUCs across all seven dimensions on Eds-vc (e.g., HuBERT: .85 intelligibility, .85 imprecise consonants, .88 naturalness, .84 monoloudness, .83 monopitch) and the highest AUC for the sum of all dimensions (.89) suggest that every probe captures a common speaker-level severity factor rather than seven distinct perceptual primitives. Please re-evaluate at the speaker level (e.g., average predictions per speaker before computing AUC) and report whether the seven dimensions provide discriminant information beyond the first principal component of the predictions.
  2. [Abstract and Section 2.4/Figure 5] The abstract states that zero-shot performance on affective speech was 'validated,' but the affect experiment is exploratory: it tabulates mean predicted VQD scores per categorical emotion with no ground-truth VQD annotations, no quantitative accuracy measure, and no confidence intervals. Since RAVDESS has no VQD labels, the experiment can only illustrate consistency with prior speech-science findings, not validate the probes. Please change the abstract and Section 3 wording to 'explored' or 'illustrated,' and present Figure 5 with error bars or per-speaker variation if it is retained.
  3. [Table 1] Several reported 95% confidence intervals are internally inconsistent. For example, monoloudness HuBERT Spearman is .61 with CI (.59, .53), where the lower bound exceeds the upper bound; breathiness HuBERT is .31 with CI (.39, .35), which does not contain the point estimate; and monopitch RawNet3 has CI (.67, .82) for a point estimate of .70, which is far wider than the other CIs. These errors prevent the reader from trusting the performance claims and need to be corrected, ideally with a description of the bootstrap procedure.
minor comments (5)
  1. [Section 2.2] The sentence beginning 'Leading and trailing silence was trimmed prior to embedding extraction for all samples a wav2vec2 forced alignment model' is missing a verb and a period; it should be rephrased.
  2. [Section 2.1] The phrase 'Flaccid Dyarthria' should be 'Flaccid Dysarthria.'
  3. [Tables 1 and 3] Dimension names are inconsistent across tables (e.g., 'Mono loud.', 'Mono-pitch', 'Impr. conson.', 'Breath-iness'); please standardize abbreviations.
  4. [Figures 4 and 5] The figures are referenced in the text but not displayed in the manuscript; if included in the final version, add axis labels, legends, and per-speaker or per-sample variability information.
  5. [Section 3] The statement that probes trained only on novel sentences had poor generalization would be more informative with a statistical comparison rather than the averaged Spearman values in Table 2.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: VQD probes are fitted to held-out SAP labels and evaluated against external severity datasets, with no prediction reduced to its own training target.

full rationale

The paper's derivation chain is self-contained. Probes are trained on SAP per-utterance SLP annotations (Section 2.1) using frozen embeddings, and all reported SAP results are on held-out speaker-stratified test splits; the zero-shot evaluations (Table 3, Figure 4) use Eds-vc and EasyCall severity labels that are not part of the training labels, so no prediction is re-derived from the labels that fitted it. The affect exploration (Section 2.4, Figure 5) is explicitly exploratory and does not feed back into model fitting. Citations to prior work are for datasets, embedding models, benchmarks, and unrelated affect/severity baselines; [25] includes a coauthor but is not load-bearing, and no uniqueness or ansatz is imported from a self-citation. The use of coarse per-speaker severity labels as an out-of-domain benchmark is a validity/measurement concern, not a circularity, because the fitted VQD labels and the benchmark labels are different; evaluating transfer on an external severity proxy cannot reduce the VQD prediction to its own training target.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper is empirical; it introduces no physical constants or new entities. Its load-bearing assumptions are about label validity: that SLP ratings are consistent, that external severity scores measure the same underlying dimensions, and that frozen embeddings support linear probes. The main unverified component is the internal CLAP model and the internal Eds-vc dataset, which are not publicly documented.

free parameters (2)
  • Lasso regularization parameter = not reported
    Selected on the validation set for each probe in Section 2.2; affects numeric scores but not the qualitative conclusion.
  • Classification binarization threshold = near 20% positive labels per dimension
    Chosen to stabilize task difficulty and comparability across dimensions (Section 2.2); AUC values depend on this label balance.
assumptions (4)
  • domain assumption Perceptual VQD ratings from SLPs are treated as ground truth on a 1-7 scale
    Section 2.1; all regression and classification targets are built from these ratings, whose reliability is not measured in this paper.
  • domain assumption Out-of-domain severity labels are comparable proxies for VQD ratings
    Section 2.3; Eds-vc per-speaker SLP severity and EasyCall TOM scores are used as labels for zero-shot AUC, but they are not per-utterance VQD annotations.
  • domain assumption Frozen SSL embeddings contain linearly separable VQD information
    Section 2.2 and Section 3; the entire method assumes linear probes over HuBERT/CLAP/RawNet3 embeddings can capture these dimensions.
  • standard math Standard statistical assumptions of Lasso and logistic regression
    Section 2.2; regularization and i.i.d. assumptions are standard and not the source of the central claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Voice Quality Dimensions as Interpretable Primitives for Speaking Style for Atypical Speech and Affect." pith.science (2026). https://pith.science/paper/DQ4CNBBW

@misc{pith2026250521809,
  author       = {Pith},
  title        = {Pith review of: Voice Quality Dimensions as Interpretable Primitives for Speaking Style for Atypical Speech and Affect},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DQ4CNBBW}},
  note         = {Machine review of arXiv:2505.21809}
}
read the original abstract

Perceptual voice quality dimensions describe key characteristics of atypical speech and other speech modulations. Here we develop and evaluate voice quality models for seven voice and speech dimensions (intelligibility, imprecise consonants, harsh voice, naturalness, monoloudness, monopitch, and breathiness). Probes were trained on the public Speech Accessibility (SAP) project dataset with 11,184 samples from 434 speakers, using embeddings from frozen pre-trained models as features. We found that our probes had both strong performance and strong generalization across speech elicitation categories in the SAP dataset. We further validated zero-shot performance on additional datasets, encompassing unseen languages and tasks: Italian atypical speech, English atypical speech, and affective speech. The strong zero-shot performance and the interpretability of results across an array of evaluations suggests the utility of using voice quality dimensions in speaking style-related tasks.

Figures

Figures reproduced from arXiv: 2505.21809 by the authors.

Figure 2
Figure 2. Distributions of annotations for imprecise consonants and naturalness for each speech category [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Regression probes, showing clear progression of pre￾dicted scores with rated severity VQD, using the validation set for selecting the regularization parameter value. We converted the annotated scores to a bi￾nary label for classification, aiming for near 20% positive labels (selected to stabilize task difficulty and enable better compar￾isons between model performance across dimensions). We also trained NN probes, b… view at source ↗
Figure 4
Figure 4. Zero-shot predictions from the HuBERT probe for each voice quality dimension on the Eds-vc dataset and on the EasyCall dataset, stratified by rated speech severity 2.3. Generalization We investigated generalization between speech categories (by training probes independently on each speech category), and generalization to new datasets - including to a language not seen during probe training. To evaluate generalizatio… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p003_5.png]
Figure 6
Figure 6. Figure 6: Affective predictions for more atypical speech shows less neutrality and higher sadness Unseen datasets The probes had strong zero-shot gener￾alization to predicting atypical speech severity on both eval￾uated datasets ( [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 38 canonical work pages

  1. [1]

    Introduction and Related Work V oice quality dimensions (VQDs) capture information about a speaker’s style and voice related to anatomy, vocal tract con- figuration, and learned components [1]. V oice quality can be impacted by disorders directly related to speech including struc- tural (e.g., voice nodules) and neurogenic conditions (e.g., spas- modic dy...

    work page

  2. [2]

    Voice Quality Dimensions as Interpretable Primitives for Speaking Style for Atypical Speech and Affect

    Methods 2.1. Training data and labels Probes were trained using the Speech Accessibility Project [29] a publicly available dataset of atypical speech. We used the sub- set of data with annotations from speech-language pathologists, which included 11,184 samples from 434 speakers (n=284 with arXiv:2505.21809v1 [cs.SD] 27 May 2025 Parkinson’s Disease, 78 wi...

    work page Pith review arXiv 2025

  3. [3]

    Figure 3 shows regression results for two VQDs

    Results and Discussion Embeddings and Speech Category Generalization.Ta- ble 1 lists Spearman correlations (similar trends were observed for the R2 and mean average error) between the regression probe prediction and ground truth annotation and the AUC for each classification probe, trained and evaluated on all speech categories. Figure 3 shows regression ...

    work page

  4. [4]

    Additionally, VQD models can also act as explanations or in- terpretations of decisions made by application focused models, as illustrated by the presented affect exploration

    Conclusions The generalizability and interpretability of the trained VQD models across datasets, languages, and tasks highlights the util- ity of VQDs for modeling non-semantic speech dimensions. Additionally, VQD models can also act as explanations or in- terpretations of decisions made by application focused models, as illustrated by the presented affec...

    work page

  5. [5]

    Consensus auditory-perceptual evaluation of voice: development of a standardized clinical proto- col,

    G. B. Kempster, B. R. Gerratt, K. V . Abbott, J. Barkmeier- Kraemer, and R. E. Hillman, “Consensus auditory-perceptual evaluation of voice: development of a standardized clinical proto- col,” 2009

    work page 2009

  6. [6]

    Auditory perceptual voice analysis,

    D. Weinstein, “Auditory perceptual voice analysis,” 2010. [Online]. Available: https://alahns.org

    work page 2010

  7. [7]

    V oice qualities,

    The National Center for V oice and Speech, “V oice qualities,”

    work page

  8. [8]

    Beyond binary gender: creaky voice, gender, and the variationist enterprise,

    K. Becker, L. Zimmanet al., “Beyond binary gender: creaky voice, gender, and the variationist enterprise,”Language Varia- tion and Change, vol. 34, no. 2, pp. 215–238, 2022

    work page 2022

Show all 44 references
  1. [9]

    Artificial intelligence in communication sciences and disorders: Introduction to the forum,

    J. R. Green, “Artificial intelligence in communication sciences and disorders: Introduction to the forum,” pp. 4157–4161, 2024

    2024

  2. [10]

    The mPower study, Parkinson disease mobile data collected using Re- searchKit,

    B. M. Bot, C. Suver, E. C. Neto, M. Kellen, A. Klein, C. Bare, M. Doerr, A. Pratap, J. Wilbanks, E. Dorseyet al., “The mPower study, Parkinson disease mobile data collected using Re- searchKit,”Scientific data, vol. 3, no. 1, pp. 1–9, 2016

    2016

  3. [11]

    Treatment outcomes for professional voice users,

    J. M. Wingate, W. S. Brown, R. Shrivastav, P. Davenport, and C. M. Sapienza, “Treatment outcomes for professional voice users,”Journal of voice, vol. 21, no. 4, pp. 433–449, 2007

    2007

  4. [12]

    Evidence-based clinical voice assessment: a systematic review,

    N. Roy, J. Barkmeier-Kraemer, T. Eadie, M. P. Sivasankar, D. Mehta, D. Paul, and R. Hillman, “Evidence-based clinical voice assessment: a systematic review,” 2013

    2013

  5. [13]

    Machine learning applied to GRBAS voice quality assessment,

    Z. Xie, C. Gadepalli, J. Farideh, B. M. Cheetham, and J. J. Homer, “Machine learning applied to GRBAS voice quality assessment,” Advances in Science, Technology and Engineering Systems Jour- nal, vol. 3, no. 6, pp. 329–338, 2018

    2018

  6. [14]

    Information conveyed by voice quality,

    J. Kreiman, “Information conveyed by voice quality,”The Journal of the Acoustical Society of America, vol. 155, no. 2, pp. 1264– 1271, 2024

    2024

  7. [15]

    Acoustic corre- lates of vocal quality,

    L. Eskenazi, D. G. Childers, and D. M. Hicks, “Acoustic corre- lates of vocal quality,”Journal of Speech, Language, and Hearing Research, vol. 33, no. 2, pp. 298–306, 1990

    1990

  8. [16]

    The role of voice quality in com- municating emotion, mood and attitude,

    C. Gobl and A. N. Chasaide, “The role of voice quality in com- municating emotion, mood and attitude,”Speech communication, vol. 40, no. 1-2, pp. 189–212, 2003

    2003

  9. [17]

    Dimensionality in voice quality,

    I. V . Bele, “Dimensionality in voice quality,”Journal of Voice, vol. 21, no. 3, pp. 257–272, 2007

    2007

  10. [18]

    Perceptual scaling of voice identity: common dimensions for different vowels and speakers,

    O. Baumann and P. Belin, “Perceptual scaling of voice identity: common dimensions for different vowels and speakers,”Psycho- logical Research PRPF, vol. 74, no. 1, pp. 110–120, 2010

    2010

  11. [19]

    Speech intelligibility classifiers from 550k disordered speech samples,

    S. Venugopalan, J. Tobin, S. J. Yang, K. Seaver, R. J. Cave, P.-P. Jiang, N. Zeghidour, R. Heywood, J. Green, and M. P. Brenner, “Speech intelligibility classifiers from 550k disordered speech samples,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and ...

    2023

  12. [20]

    A VEC 2019 workshop and challenge: state-of-mind, detecting depression with AI, and cross-cultural affect recogni- tion,

    F. Ringeval, B. Schuller, M. Valstar, N. Cummins, R. Cowie, L. Tavabi, M. Schmitt, S. Alisamir, S. Amiriparian, E.-M. Mess- neret al., “A VEC 2019 workshop and challenge: state-of-mind, detecting depression with AI, and cross-cultural affect recogni- tion,” inProceedings of th...

    2019

  13. [21]

    Objective measurements of voice quality,

    H. Dhamyal and R. Singh, “Objective measurements of voice quality,”arXiv preprint arXiv:2410.09578, 2024

    2024 arXiv

  14. [22]

    Dissociating COVID-19 from other respira- tory infections based on acoustic, motor coordination, and phone- mic patterns,

    T. Talkar, D. M. Low, A. J. Simpkin, S. Ghosh, D. T. O’Keeffe, and T. F. Quatieri, “Dissociating COVID-19 from other respira- tory infections based on acoustic, motor coordination, and phone- mic patterns,”Scientific Reports, vol. 13, no. 1, p. 1567, 2023

    2023

  15. [23]

    Shue,The voice source in speech production: Data, analysis and models

    Y . Shue,The voice source in speech production: Data, analysis and models. University of California, Los Angeles, 2010

    2010

  16. [24]

    Enabling off-the-shelf disfluency detection and categorization for pathological speech,

    A. Romana, M. Niu, M. Perez, A. Roberts, and E. M. Provost, “Enabling off-the-shelf disfluency detection and categorization for pathological speech,”Proc. Interspeech 2022, pp. 1916–1920, 2022

    2022

  17. [25]

    Seq2seq for automatic paraphasia detection in aphasic speech,

    M. Perez, D. Le, A. Romana, E. Jones, K. Licata, and E. M. Provost, “Seq2seq for automatic paraphasia detection in aphasic speech,”arXiv preprint arXiv:2312.10518, 2023

    2023 arXiv

  18. [26]

    Odyssey 2024-speech emotion recognition challenge: Dataset, baseline framework, and results,

    L. Goncalves, A. N. Salman, A. R. Naini, L. M. Velazquez, T. Thebaud, L. P. Garcia, N. Dehak, B. Sisman, and C. Busso, “Odyssey 2024-speech emotion recognition challenge: Dataset, baseline framework, and results,”Development, vol. 10, no. 9,290, pp. 4–54, 2024

    2024

  19. [27]

    Universal paralinguistic speech representations using self-supervised con- formers,

    J. Shor, A. Jansen, W. Han, D. Park, and Y . Zhang, “Universal paralinguistic speech representations using self-supervised con- formers,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 3169–3173

    2022

  20. [28]

    A rapid, non-invasive method for fatigue detection based on voice information,

    X. Gao, K. Ma, H. Yang, K. Wang, B. Fu, Y . Zhu, X. She, and B. Cui, “A rapid, non-invasive method for fatigue detection based on voice information,”Frontiers in Cell and Developmental Biol- ogy, vol. 10, p. 994001, 2022

    2022

  21. [29]

    Community-supported shared infrastructure in support of speech accessibility,

    M. Hasegawa-Johnson, X. Zheng, H. Kim, C. Mendes, M. Dickin- son, E. Hege, C. Zwilling, M. M. Channell, L. Mattie, H. Hodges et al., “Community-supported shared infrastructure in support of speech accessibility,”Journal of Speech, Language, and Hearing Research, vol. 67, no. 1...

    2024

  22. [30]

    A machine-learning based objective measure for ALS disease sever- ity,

    F. G. Vieira, S. Venugopalan, A. S. Premasiri, M. McNally, A. Jansen, K. McCloskey, M. P. Brenner, and S. Perrin, “A machine-learning based objective measure for ALS disease sever- ity,”NPJ digital medicine, vol. 5, no. 1, p. 45, 2022

    2022

  23. [31]

    Alzheimer’s dementia recognition through spontaneous speech,

    S. Luz, F. Haider, S. de la Fuente Garcia, D. Fromm, and B. MacWhinney, “Alzheimer’s dementia recognition through spontaneous speech,” p. 780169, 2021

    2021

  24. [32]

    Investigating salient representa- tions and label variance in dimensional speech emotion analysis,

    V . Mitra, J. Nie, and E. Azemi, “Investigating salient representa- tions and label variance in dimensional speech emotion analysis,” inICASSP 2024-2024 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 11 111–11 115

    2024

  25. [33]

    Wav2Vec2FABundle,

    Torchaudio contributors, “Wav2Vec2FABundle,” 2024. [Online]. Available: https://pytorch.org

    2024

  26. [34]

    Easycall corpus: a dysarthric speech dataset,

    R. Turrisi, A. Braccia, M. Emanuele, S. Giulietti, M. Pugliatti, M. Sensi, L. Fadiga, and L. Badino, “Easycall corpus: a dysarthric speech dataset,”arXiv preprint arXiv:2104.02542, 2021

    2021 arXiv

  27. [35]

    RA VDESS is an acted affect dataset with seven categorical emotions: calm, happy, sad, angry, fearful, disgust, surprised

    dataset. RA VDESS is an acted affect dataset with seven categorical emotions: calm, happy, sad, angry, fearful, disgust, surprised. We used probes trained only on the SAP dataset (which has no elicited affective content) to generate predictions for each VQD. We tabulated the a...

  28. [36]

    On the utility of self-supervised models for prosody- related tasks,

    G. Lin, C. Feng, W. Huang, Y . Tseng, T. Lin, C. Li, H. Lee, and N. G. Ward, “On the utility of self-supervised models for prosody- related tasks,” in2022 IEEE Spoken Language Technology Work- shop (SLT). IEEE, 2023, pp. 1104–1111

    2023

  29. [37]

    Automatic speech recognition of conversational speech in individuals with disordered speech,

    J. Tobin, P. Nelson, B. MacDonald, R. Heywood, R. Cave, K. Seaver, A. Desjardins, P.-P. Jiang, and J. R. Green, “Automatic speech recognition of conversational speech in individuals with disordered speech,”Journal of Speech, Language, and Hearing Research, vol. 67, no. 11, pp....

    2024

  30. [38]

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

    W. Hsu, B. Bolte, Y . H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,”IEEE/ACM trans- actions on audio, speech, and language processing, vol. 29, pp. 3451–3460, 2021

    2021

  31. [39]

    ESPnet-SPK: full pipeline speaker embedding toolkit with reproducible recipes, self- supervised front-ends, and off-the-shelf models,

    J. Jung, W. Zhang, J. Shi, Z. Aldeneh, T. Higuchi, B. Theobald, A. H. Abdelaziz, and S. Watanabe, “ESPnet-SPK: full pipeline speaker embedding toolkit with reproducible recipes, self- supervised front-ends, and off-the-shelf models,”Proc. Inter- speech 2024, 2024

    2024

  32. [42]

    The Ryerson Audio-Visual Database of Emotional Speech and Song (RA VDESS): A dy- namic, 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 dy- namic, multimodal set of facial and vocal expressions in North American English,”PloS one, vol. 13, no. 5, p. e0196391, 2018

    2018

  33. [43]

    A moan of pleasure should be breathy: the effect of voice quality on the meaning of human nonverbal vocalizations,

    A. Anikin, “A moan of pleasure should be breathy: the effect of voice quality on the meaning of human nonverbal vocalizations,” Phonetica, vol. 77, no. 5, pp. 327–349, 2020

    2020

  34. [44]

    Investigating the emergent audio classification ability of ASR Foundation Models,

    R. Ma, A. Liusie, M. J. Gales, and K. M. Knill, “Investigating the emergent audio classification ability of ASR Foundation Models,” arXiv preprint arXiv:2311.09363, 2023

    2023 arXiv

  35. [192]

    RawNet3 was in- cluded despite its smaller size to evaluate the use of a model trained for speaker ID, which could relate to VQDs as they may be tied to speaker characteristics

    [32], pre-trained speaker identification model HuBERT Large and HuBERT Large ASR were included to investigate the impact of ASR-targeted fine-tuning and to compare with the CLAP training strategy. RawNet3 was in- cluded despite its smaller size to evaluate the use of a model t...

  36. [2024]

    Available: https://ncvs.org

    [Online]. Available: https://ncvs.org

Pith tools

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