Pith. sign in

REVIEW 3 major objections 6 minor 35 references

Towards Machine Unlearning for Paralinguistic Speech Processing

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

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

Pith's one-line read This paper claims that weight-averaging the shard models of SISA's unlearning pipeline preserves more speech-task accuracy after user-data removal than SISA's vote-based aggregation.

desk verdict First MU-for-PSP paper; SISA++ is weight-averaged SISA with plausible utility gains, but the forgetting claim itself is unverified. read the letter →

arxiv 2506.02230 v1 pith:YW53ZXV4 submitted 2025-06-02 eess.AS cs.SD

Orchid Chetia Phukan , Girish , Mohd Mujtaba Akhtar , Shubham Singh , Swarup Ranjan Behera , Vandana Rajan , Muskaan Singh , Arun Balaji Buduru
show 1 more author
Rajesh Sharma
This is my paper · ORCID
classification eess.AScs.SD
keywords machineunlearningparalinguisticspeechprocessingemotionrecognitiondepressiondetectionSISA++weightaveragingTRILLssonrighttobeforgotten
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 brings machine unlearning—selectively removing the influence of a user's data from a trained model—to paralinguistic speech tasks, where voice samples carry emotion and mental-health information. It proposes SISA++, a variant of the SISA unlearning method that merges separately trained data-shard models by averaging their weights instead of combining their outputs by vote. The paper's claim is that this one change consistently preserves more downstream accuracy on speech emotion recognition (CREMA-D) and depression detection (E-DAIC) after simulated one- or two-user removal, across 4- and 8-shard setups, without adding inference cost. It also reports recipe-style guidance, identifying TRILLsson features with a transformer head as the combination that loses least performance after unlearning. If the claim holds, weight averaging is a cheap way to make right-to-be-forgotten requests less damaging to speech models.

What carries the argument

The central object is SISA++, the weight-averaging aggregation step added to the SISA unlearning pipeline. After shard sub-models are retrained following a deletion, SISA++ computes $M_A = \frac{1}{N}\sum_{i=1}^{N} M_i$, the element-wise average of the shard model weights, and uses that single model for all inference. This replaces SISA's majority-vote or output-averaging aggregation and is the mechanism the paper credits for preserving post-unlearning performance.

What would settle it

Run SISA++ and SISA on identical shard configurations and measure a membership inference attack or parameter distance to a retrained-from-scratch model for the removed users; if SISA++ does not reduce the deleted users' influence relative to SISA, the utility-preservation result does not demonstrate unlearning.

Watch

Extended reading notes

Core claim

The paper's central claim is that SISA++ outperforms SISA at retaining task performance after unlearning. SISA divides the training set into disjoint shards, trains a sub-model on each, and when deletion requests arrive retrains only the shards containing removed data, combining all sub-model outputs by majority vote (classification) or averaging (regression). SISA++ keeps the sharding and retraining procedure but replaces the output-level combination with a weight-averaged model: the final weights are the arithmetic mean of all shard sub-model weights. Across one-user and two-user removal simulations on CREMA-D and E-DAIC, the paper reports higher accuracy and F1 for emotion recognition and lower MAE and RMSE for depression detection with SISA++ than with SISA in the tested configurations, and it attributes this to the consolidating effect of weight averaging.

Load-bearing premise

The paper's conclusion stands on the assumption that test accuracy and error after deletion are enough to show unlearning worked; it never measures whether the supposedly deleted users' data still influences the model.

Editorial extensions

If this is right

  • If SISA++ works as claimed, an unlearning request can be served by retraining only the affected shard and re-averaging weights, with no increase in inference cost over a single model.
  • Post-unlearning utility on emotion recognition and depression detection is higher with weight averaging than with vote-based aggregation in the configurations tested.
  • The reported evaluations suggest TRILLsson features paired with a transformer downstream suffer the smallest performance drop after unlearning among the feature-architecture pairs tried.
  • Finer-grained sharding (8 shards) sometimes reduces degradation on its own, but the SISA++ advantage holds across both 4- and 8-shard settings.

Reading between the lines

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

  • A follow-up the paper leaves untested: whether weight averaging actually removes the influence of deleted users, or only preserves utility. A membership-inference attack on supposedly unlearned speakers would settle this.
  • The result hints at a broader principle for shard-based unlearning: parameter-space aggregation can be more lossless than output-space ensembling, which could be tested on image and text benchmarks.
  • A practical extension of the cookbook recipes would turn the feature-architecture ranking into a compliance checklist, but regulatory value depends on forgetting guarantees, not accuracy retention alone.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper introduces SISA++, a modification of SISA for machine unlearning in paralinguistic speech processing. SISA++ trains one model per data shard and merges the shard models by weight averaging (Algorithm 1) instead of aggregating their outputs at inference. Experiments on CREMA-D (speech emotion recognition) and E-DAIC (depression detection) compare SISA and SISA++ under one- and two-user removal with four-shard and eight-shard configurations, across five feature sets (MFCC, x-vector, XLS-R, WavLM, TRILLsson) and three downstream networks (SVM, CNN, Transformer). The paper reports that SISA++ retains accuracy/F1/MAE/RMSE better than SISA after unlearning, and it distills the findings into 'cookbook recipes' that recommend TRILLsson features with a transformer encoder for best post-unlearning utility retention.

Significance. If the utility-retention result is robust, SISA++ is an appealingly simple, zero-inference-cost modification: it replaces output aggregation with weight averaging and requires no additional training. The paper's strengths include its broad experimental sweep (five feature sets, three downstream models, two datasets, two shard counts, two deletion sizes), the public code and model repository link, and the actionable recommendations for practitioners. The contribution would be a useful empirical result for the paralinguistic speech community, provided the unlearning claim is actually verified; in its current form the evaluation is utility-only and does not establish that deleted users' influence is removed from the model.

major comments (3)
  1. [§3.2, Table 2] The paper's central claim that SISA++ demonstrates superior retention 'post-unlearning' is based entirely on utility metrics; there is no forgetting metric, no membership-inference attack, and no comparison with a retrain-from-scratch model. Because Algorithm 1 forms the final model by averaging all shard-model weights, the observed gains could in principle be the model-soup effect of weight averaging (Ref. [20]) rather than evidence that deleted users' influence has been removed. This distinction matters for the right-to-be-forgotten framing in Section 1; please add at least one verification of forgetting (e.g., membership-inference success before and after unlearning, or closeness to a model retrained without the deleted users), or explicitly reframe the contribution as utility-preserving shard retraining.
  2. [§3.2, Table 2] All comparisons in Table 2 are single-run point estimates without error bars, multiple seeds, or significance tests, yet the text states that SISA++ 'consistently outperforms' SISA. Differences in several cells are small (for example, E-DAIC, two-user removal, eight shards, TRILLsson+Transformer: MAE 4.45 for SISA versus 4.30 for SISA++), and without variance estimates that claim is not supported. Please report means and standard deviations over multiple runs, and use a paired significance test or confidence intervals where the comparison is made.
  3. [§2.2, §3.2] The evaluation protocol is underspecified. It is not stated how a removed user's utterances are assigned to shards (for example, whether all clips of a user stay in one shard), whether the test set contains utterances from removed users, how SISA's output aggregation is implemented for the regression task in Table 2, or whether SISA++ is evaluated with the weight-averaged model MA on the same test set. These details are needed to interpret the comparison and to reproduce the numbers; without them, the observed 4-shard versus 8-shard and SISA versus SISA++ differences are difficult to attribute to the proposed mechanism.
minor comments (6)
  1. [Table 1 caption] The caption contains typos: 'Evalution Scores' should be 'Evaluation Scores', and 'marco average F1 score' should be 'macro average F1 score'.
  2. [Contributions bullet list] The third bullet contains 'guideline swith'; it should read 'guidelines with'.
  3. [Introduction] The sentence 'one notable issue in this context is the is the principle of the right to be forgotten' has a duplicated phrase and should be corrected.
  4. [§2.3] The downstream networks are described as SVM, CNN, and Transformer, but the parameter counts refer to 'FCN models'; please clarify whether FCN is a fourth architecture or a typo for CNN.
  5. [§3.2] The final paragraph contains the raw LaTeX command 'textbfSISA++' instead of formatted bold text.
  6. [Conclusion] The phrase 'an novel extension' should be 'a novel extension'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the SISA++ vs SISA comparison is a straightforward empirical A/B test; no parameter is fitted to the reported outcome and no load-bearing self-citation appears.

full rationale

The central claim, 'SISA++ consistently outperforms SISA, demonstrating superior retention of performance post-unlearning' (Section 3.2), is supported by Table 2, which reports test-set metrics for both methods under identical sharding, feature, and downstream configurations. The proposed modification (Algorithm 1) is uniform weight averaging of shard models, introduced in Section 2.2; no averaging coefficient or other hyperparameter is optimized against the post-unlearning metrics, so the comparison cannot reduce to a fit by construction. The SISA baseline is external prior work (Bourtoule et al. [7]), and weight averaging is justified by external model-soup results (Wortsman et al. [20]; Vander Eeckt and Van Hamme [21]); no load-bearing self-citation chain is present. The 'cookbook recipes' in Sections 2.3 and 3.2 select the best feature/architecture pair from Tables 1-2; this is post hoc recommendation, not a prediction derived from its own input. The paper's main weakness, the absence of membership-inference or retrain-from-scratch comparisons to verify that removed users' influence is actually gone, is an evaluation/completeness limitation rather than a circular-reasoning defect. No circular step can be exhibited with specific equations or the paper's own reduction.

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

The central claim rests on standard supervised learning assumptions plus two domain assumptions: retraining only affected shards is sufficient for unlearning, and weight averaging of shard models is a valid aggregation. No new entities are introduced.

assumptions (4)
  • domain assumption Pretrained feature extractors (TRILLsson, XLS-R, WavLM, x-vector) are frozen and provide task-relevant fixed representations.
    The entire experiment uses these frozen embeddings as inputs to downstream classifiers; if any encoder leaks target labels or is not actually task-relevant, the comparisons shift. Invoked in Section 2.3.
  • domain assumption Retraining only affected shards after data removal is a valid unlearning procedure.
    SISA's core premise, adopted by SISA++, is that partition-based retraining removes the influence of deleted points. No certified or empirical forgetting check is performed. Invoked in Section 2.1 and Section 3.2.
  • domain assumption Weight averaging of models trained on disjoint shards is a valid inference-time aggregation.
    SISA++ relies on model soups [20] and weight averaging [21] to merge shard models, but those methods were designed for fine-tuned variants of the same architecture, not necessarily for models trained on disjoint data. Invoked in Section 2.2.
  • domain assumption The standard train/test splits for CREMA-D and E-DAIC are appropriate and the evaluation metrics measure the intended capability.
    CREMA-D uses an 80:20 split and E-DAIC uses the official split; the paper does not discuss speaker overlap or data leakage between splits. Invoked in Section 3.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Machine Unlearning for Paralinguistic Speech Processing." pith.science (2026). https://pith.science/paper/YW53ZXV4

@misc{pith2026250602230,
  author       = {Pith},
  title        = {Pith review of: Towards Machine Unlearning for Paralinguistic Speech Processing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YW53ZXV4}},
  note         = {Machine review of arXiv:2506.02230}
}
read the original abstract

In this work, we pioneer the study of Machine Unlearning (MU) for Paralinguistic Speech Processing (PSP). We focus on two key PSP tasks: Speech Emotion Recognition (SER) and Depression Detection (DD). To this end, we propose, SISA++, a novel extension to previous state-of-the-art (SOTA) MU method, SISA by merging models trained on different shards with weight-averaging. With such modifications, we show that SISA++ preserves performance more in comparison to SISA after unlearning in benchmark SER (CREMA-D) and DD (E-DAIC) datasets. Also, to guide future research for easier adoption of MU for PSP, we present ``cookbook recipes'' - actionable recommendations for selecting optimal feature representations and downstream architectures that can mitigate performance degradation after the unlearning process.

Figures

Figures reproduced from arXiv: 2506.02230 by the authors.

Figure 1
Figure 1. Workflow of SISA++; Original training dataset D and its multiple shards (D1, D2, . . . , DN ), which are further divided into slices (D1,1, D1,2, . . . , DN,R); Constituent model for each shard (M1, M2, . . . , MN ); Orange symbol indicates the data point to be unlearned; (MA) represents the final model after weight averaging predictions. Algorithm 1 SISA++ Require: A set of models {M1,M2, . . . ,MN }, each trained … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 28 canonical work pages

  1. [20]

    Privacy risks in speech emotion recognition: A systematic study on gender inference at- tack,

    B. Alsenani, T. Guha, and A. Vinciarelli, “Privacy risks in speech emotion recognition: A systematic study on gender inference at- tack,” inInterspeech 2023, 2023, pp. 651–655

    work page 2023

  2. [1]

    Towards Machine Unlearning for Paralinguistic Speech Processing

    Introduction The widespread use of ML models in applications ranging from personalized recommendations to health diagnostics has brought privacy issues to the forefront. As these models rely on vast amounts of personal data to optimize their performance, en- suring the ethical handling of such data has become a critical challenge. With such comes the risk...

    work page Pith review arXiv 2025

  3. [2]

    cookbook recipes

    Methodology In this section, we discuss preliminaries on SISA, the proposed novel extension, SISA++ and lastly followed by the“cookbook recipes”. 2.1. Preliminary on SISA SISA [ 7] is a SOTA data-driven technique for MU. It em- ploys a data partitioning strategy with structured model train- ing to enable efficient unlearning. The initial dataset D= {(xi, ...

    work page

  4. [3]

    Experiments 3.1. Dataset Crowd-sourced Emotional Multimodal Actors Dataset (CREMA-D)[ 30] is widely used SER dataset containing 7,442 utterances from 91 speakers (48 male, 43 female) across six emo- tions: Anger, Happiness, Sadness, Fear, Disgust, and Neutral. We use 80:20 split ratio for training and testing our models. E-DAIC[ 31] is a benchmark DD data...

    work page

  5. [4]

    cookbook recipes

    Conclusion In this study, we pioneered MU for PSP and proposed SISA++, an novel extension to previous SOTA MU method SISA. It merges models trained on different shards via weight averaging, outperforming SISA in preserving performance post-unlearning. To aid future research, we provide“cookbook recipes”and as a part of this we recommend the use of transfo...

    work page

  6. [5]

    Membership inference attacks against machine learning models,

    R. Shokri, M. Stronati, C. Song, and V . Shmatikov, “Membership inference attacks against machine learning models,” in2017 IEEE symposium on security and privacy (SP). IEEE, 2017, pp. 3–18

    2017

  7. [6]

    Secure and efficient federated learning by combining homomorphic en- cryption and gradient pruning in speech emotion recognition,

    S. Mohammadi, S. Sinaei, A. Balador, and F. Flammini, “Secure and efficient federated learning by combining homomorphic en- cryption and gradient pruning in speech emotion recognition,” in International Conference on Information Security Practice and Experience. Springer, 2023, pp. 1–16

    work page 2023

  8. [7]

    Privacy-preserving speech emotion recognition through semi-supervised federated learning,

    V . Tsouvalas, T. Ozcelebi, and N. Meratnia, “Privacy-preserving speech emotion recognition through semi-supervised federated learning,” in2022 IEEE International Conference on Pervasive Computing and Communications Workshops and other Affiliated Events (PerCom Workshops). IEEE, 2022, pp. 359–364

    work page 2022

Show all 35 references
  1. [8]

    Differential privacy: A survey of results,

    C. Dwork, “Differential privacy: A survey of results,” inInterna- tional conference on theory and applications of models of compu- tation. Springer, 2008, pp. 1–19

    2008

  2. [9]

    The eu proposal for a general data protection regu- lation and the roots of the ‘right to be forgotten’,

    A. Mantelero, “The eu proposal for a general data protection regu- lation and the roots of the ‘right to be forgotten’,”Computer Law & Security Review, vol. 29, no. 3, pp. 229–235, 2013

    2013

  3. [10]

    Towards probabilistic verification of machine unlearning,

    D. M. Sommer, L. Song, S. Wagh, and P. Mittal, “Towards probabilistic verification of machine unlearning,”arXiv preprint arXiv:2003.04247, 2020

    2003 arXiv

  4. [11]

    Machine un- learning,

    L. Bourtoule, V . Chandrasekaran, C. A. Choquette-Choo, H. Jia, A. Travers, B. Zhang, D. Lie, and N. Papernot, “Machine un- learning,” in2021 IEEE Symposium on Security and Privacy (SP). IEEE, 2021, pp. 141–159

    2021

  5. [12]

    A survey of machine unlearning,

    T. T. Nguyen, T. T. Huynh, P. L. Nguyen, A. W.-C. Liew, H. Yin, and Q. V . H. Nguyen, “A survey of machine unlearning,”arXiv preprint arXiv:2209.02299, 2022

    2022 arXiv

  6. [13]

    Erm-ktp: Knowledge-level machine unlearning via knowledge transfer,

    S. Lin, X. Zhang, C. Chen, X. Chen, and W. Susilo, “Erm-ktp: Knowledge-level machine unlearning via knowledge transfer,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 20 147–20 155

    2023

  7. [14]

    Machine unlearning for image- to-image generative models,

    G. Li, H. Hsu, R. Marculescuet al., “Machine unlearning for image- to-image generative models,”arXiv preprint arXiv:2402.00351, 2024

    2024 arXiv

  8. [15]

    Understanding black-box predictions via influence functions,

    P. W. Koh and P. Liang, “Understanding black-box predictions via influence functions,” inInternational conference on machine learning. PMLR, 2017, pp. 1885–1894

    2017

  9. [16]

    Backdoor attacks via machine unlearning,

    Z. Liu, T. Wang, M. Huai, and C. Miao, “Backdoor attacks via machine unlearning,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 13, 2024, pp. 14 115–14 123

    2024

  10. [17]

    Fast model debias with machine unlearn- ing,

    R. Chen, J. Yang, H. Xiong, J. Bai, T. Hu, J. Hao, Y . Feng, J. T. Zhou, J. Wu, and Z. Liu, “Fast model debias with machine unlearn- ing,”Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  11. [18]

    Making rec- ommender systems forget: Learning and unlearning for erasable recommendation,

    Y . Li, C. Chen, X. Zheng, J. Liu, and J. Wang, “Making rec- ommender systems forget: Learning and unlearning for erasable recommendation,”Knowledge-Based Systems, vol. 283, p. 111124, 2024

    2024

  12. [19]

    KGA: A general machine unlearning framework based on knowledge gap alignment,

    L. Wang, T. Chen, W. Yuan, X. Zeng, K.-F. Wong, and H. Yin, “KGA: A general machine unlearning framework based on knowledge gap alignment,” inProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), A. Rogers, J. Boyd-Gra...

    2023

  13. [21]

    Assessing privacy risks of attribute inference attacks against speech-based depression detection system,

    B. Alsenani, A. Esposito, A. Vinciarelli, and T. Guha, “Assessing privacy risks of attribute inference attacks against speech-based depression detection system,” 2024

    2024

  14. [22]

    Self-reported symptoms of depression and ptsd are associated with reduced vowel space in screening interviews,

    S. Scherer, G. M. Lucas, J. Gratch, A. S. Rizzo, and L.-P. Morency, “Self-reported symptoms of depression and ptsd are associated with reduced vowel space in screening interviews,”IEEE Transactions on Affective Computing, vol. 7, no. 1, pp. 59–73, 2015

    2015

  15. [23]

    Ethics and law in research on algo- rithmic and data-driven technology in mental health care: scoping review,

    P. Gooding and T. Kariotis, “Ethics and law in research on algo- rithmic and data-driven technology in mental health care: scoping review,”JMIR Mental Health, vol. 8, no. 6, p. e24668, 2021

    2021

  16. [24]

    Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time,

    M. Wortsman, G. Ilharco, S. Y . Gadre, R. Roelofs, R. Gontijo- Lopes, A. S. Morcos, H. Namkoong, A. Farhadi, Y . Carmon, S. Kornblithet al., “Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time,” inInternational conf...

    2022

  17. [25]

    Weight averaging: A sim- ple yet effective method to overcome catastrophic forgetting in automatic speech recognition,

    S. Vander Eeckt and H. Van Hamme, “Weight averaging: A sim- ple yet effective method to overcome catastrophic forgetting in automatic speech recognition,” inICASSP 2023 - 2023 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  18. [26]

    Trillsson: Distilled universal paralin- guistic speech representations,

    J. Shor and S. Venugopalan, “Trillsson: Distilled universal paralin- guistic speech representations,”arXiv preprint arXiv:2203.00236, 2022

    2022 arXiv

  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]

    XLS-R: Self-supervised Cross-lingual Speech Rep- resentation Learning at Scale,

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. von Platen, Y . Saraf, J. Pino, A. Baevski, A. Conneau, and M. Auli, “XLS-R: Self-supervised Cross-lingual Speech Rep- resentation Learning at Scale,” inProc. Interspeech 2022, 2022, pp. 2278–2282

    2022

  21. [29]

    Wavlm: Large-scale self-supervised pre-training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiaoet al., “Wavlm: Large-scale self-supervised pre-training for full stack speech processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505– 1518, 2022

    2022

  22. [30]

    X-vectors: Robust dnn embeddings for speaker recognition,

    D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudanpur, “X-vectors: Robust dnn embeddings for speaker recognition,” in 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, pp. 5329–5333

    2018

  23. [31]

    X- vectors meet emotions: A study on dependencies between emotion and speaker recognition,

    R. Pappagari, T. Wang, J. Villalba, N. Chen, and N. Dehak, “X- vectors meet emotions: A study on dependencies between emotion and speaker recognition,” inICASSP 2020 - 2020 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020, pp. 7169–7173

    2020

  24. [32]

    Auto- matic assessment of the degree of clinical depression from speech using x-vectors,

    J. V . Egas-López, G. Kiss, D. Sztahó, and G. Gosztolya, “Auto- matic assessment of the degree of clinical depression from speech using x-vectors,” inICASSP 2022 - 2022 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 8502–8506

    2022

  25. [33]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,”Advances in Neural Infor- mation Processing Systems, 2017

    2017

  26. [34]

    Crema-d: Crowd-sourced emotional multimodal actors dataset,

    H. Cao, D. G. Cooper, M. K. Keutmann, R. C. Gur, A. Nenkova, and R. Verma, “Crema-d: Crowd-sourced emotional multimodal actors dataset,”IEEE transactions on affective computing, vol. 5, no. 4, pp. 377–390, 2014

    2014

  27. [35]

    Avec 2018 workshop and challenge: Bipolar disorder and cross-cultural affect recognition,

    F. Ringeval, B. Schuller, M. Valstar, R. Cowie, H. Kaya, M. Schmitt, S. Amiriparian, N. Cummins, D. Lalanne, A. Michaud et al., “Avec 2018 workshop and challenge: Bipolar disorder and cross-cultural affect recognition,” inProceedings of the 2018 on audio/visual emotion challen...

    2018

Pith tools

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