EMO-Debias: Benchmarking Gender Debiasing Techniques in Multi-Label Speech Emotion Recognition

Huang-Cheng Chou; Hung-yi Lee; Yi-Cheng Lin; Yu-Hsuan Li Liang

REVIEW 3 major objections 4 minor 56 references

EMO-Debias: Benchmarking Gender Debiasing Techniques in Multi-Label Speech Emotion Recognition

T0 review · 3 major / 4 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 Under high gender imbalance, GADRO with gender labels and LfF without them reduce all four fairness gaps in multi-label speech emotion recognition while keeping accuracy nearly intact.

desk verdict Useful first benchmark of debiasing methods for multi-label SER, but the headline ranking of GADRO and LfF is not actually established by single-run numbers. read the letter →

arxiv 2506.04652 v1 pith:QRVRGPVF submitted 2025-06-05 eess.AS cs.CL

Yi-Cheng Lin , Huang-Cheng Chou , Yu-Hsuan Li Liang , Hung-yi Lee This is my paper

classification eess.AScs.CL

keywords speechemotionrecognitiongenderbiasmulti-labelclassificationfairnessmetricsdebiasingdistributionallyrobustoptimizationbiasedlearnersself-supervisedrepresentations

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 debiasing techniques developed for single-label tasks still work when speech emotion recognition is treated as multi-label, where one utterance can carry several emotions at once. It adapts twelve existing methods and adds one new regularizer, then compares all thirteen against a plain empirical-risk-minimization baseline on MSP-Podcast and CREMA-D with WavLM and XLSR encoders, under gender ratios from balanced to 1:40. Its main claim is that the choice of method should depend on whether gender labels are available: with labels, GADRO—a group distributionally robust objective with a small-group regularizer—consistently cuts all four fairness gaps with minimal accuracy loss, and without labels, LfF—a biased-learner scheme—is the most robust pick. The paper also finds that adversarial debiasing, a common single-label remedy, often trades one fairness gap for another in this setting, because removing gender information can break correlations among co-occurring emotions. If correct, the result gives practitioners a concrete selection rule and shows that imbalance in the training distribution is itself a major driver of unfairness.

What carries the argument

The machinery is a controlled evaluation design plus three mechanisms. The evaluation varies male-to-female utterance ratios from 1:1 to 1:40 in training and development sets while leaving the test set unchanged, so fairness is measured under explicit distributional stress. The mechanisms that carry the main result are GADRO, the group distributionally robust objective $\max_{g\in G} E(\mathrm{CE}(y,\hat y)) + \lambda\sqrt{n_g}$; LfF's relative-difficulty weighting $W(x)=\mathrm{CE}(y,\hat y_B)/(\mathrm{CE}(y,\hat y_B)+\mathrm{CE}(y,\hat y_D))$; and the authors' new Gap Regularization, which adds a penalty proportional to TPR gap plus FPR gap. All emotion losses are computed against distributional labels rather than one-hot targets, which is the adaptation that lets single-label debiasing methods work in the multi-label setting. For the no-supervision accuracy–fairness trade-off, the adapted class-wise low-variance regularizer is the key mechanism.

What would settle it

Retrain ERM, GADRO, and LfF at the 1:20 imbalance on MSP-Podcast with WavLM for, say, ten seeds and compute the spread of TPR gap, F1, and ACC. If the four-fairness-metric advantage of GADRO and LfF over ERM is smaller than the run-to-run spread, the robustness ranking is not supported by the current single-run measurements.

Watch

Extended reading notes

Core claim

The paper's central claim is an evidence-based selection rule for debiasing multi-label speech emotion recognition under high gender imbalance. On both WavLM and XLSR backbones and on both MSP-Podcast and CREMA-D, GADRO, which minimizes the worst-group loss with a $\lambda\sqrt{n_g}$ penalty that prevents overfitting to small gender groups, and LfF, which trains a deliberately biased model with generalized cross-entropy and uses the relative difficulty of each sample to reweight the debiased model, are singled out as meeting a dual requirement: they reduce TPR gap, FPR gap, F1 gap, and DP gap relative to plain training while keeping F1 and accuracy drops minimal. The paper presents this dual requirement—fairness gains across two datasets and two backbones plus negligible performance loss—as making GADRO and LfF the most robust debiasing methods under high gender imbalance. It also shows that accuracy falls and TPR gap grows as the training gender ratio moves from 1:1 to 1:40, and that single-label adversarial debiasing does not transfer cleanly to multi-label emotion recognition, because forcing gender information out can disrupt the subtle relationships among co-occurring emotions.

Load-bearing premise

The rankings are treated as meaningful even though every configuration is trained once, with no random seeds, confidence intervals, or significance tests; several deciding differences between methods are 0.01 or 0.02 in F1 or fairness gaps.

Editorial extensions

If this is right

A practitioner with gender labels should prefer GADRO over other bias-supervised methods under heavy imbalance, since it is the only method the paper identifies as robustly reducing all four fairness gaps on both backbones with minimal F1 and accuracy loss.
A practitioner without gender labels should prefer LfF over plain training and over other unsupervised methods, because it is the most robust no-supervision debiaser by the same dual criterion.
Adversarial debiasing should be used with caution in multi-label SER: ADV and MADV can cut DP gap but increase TPR gap, FPR gap, and F1 gap, often sharply, so they fail the no-trade-off test.
Gender imbalance in the training data is itself a driver of unfairness: as the ratio moves from 1:1 to 1:40, TPR gap rises while Hamming accuracy falls on both datasets.
Among bias-supervised methods, reweighting offers the best accuracy-preserving option, while downsampling reaches the lowest gaps only by sacrificing F1; without supervision, the adapted low-variance regularizer gives the best accuracy–fairness trade-off.

Reading between the lines

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

Editorial extension: a natural next experiment is to replace explicit gender labels with unsupervised speaker-identity or cluster estimates and retrain GADRO, which would test whether the best supervised method can be made to work in the no-supervision regime the paper shows is harder.
Editorial extension: the multi-label adaptation of generalized cross-entropy and relative-difficulty weighting is a transferable mechanism, so applying LfF to other multi-label fairness problems would test whether the result is specific to speech emotion recognition or generalizes to multi-label structure.
Editorial extension: because the benchmark simulates imbalance by subsampling rather than collecting naturally imbalanced speech, the reported monotonic fairness degradation is a lower-bound estimate; real datasets with correlated speaker demographics could show larger or different effects.
Editorial caveat: the phrase 'most robust' is stronger than the single-run evidence supports; a multi-seed replication would turn the selection rule from a recommendation into a quantified estimate.

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.

Desk Editor's Note

read the letter

Two things worth knowing. First, this is the first systematic benchmark of gender-debiasing methods for multi-label SER, and it fills a real gap. Second, the headline recommendation—GADRO and LfF are the most robust—rests on single-run comparisons where the decisive margins are often 0.01–0.04 and sometimes ties. That is the paper's real vulnerability, and the stress-test note gets it right.

What is genuinely new: twelve single-label debiasing methods are adapted to distributional multi-label emotion targets, plus one proposed gap-regularization loss (GR). The benchmark covers five method families, two datasets (acted and naturalistic), two SSL backbones, and multiple gender-imbalance ratios. That is a lot of work per method, and the comparison protocol is mostly sensible. The negative result for adversarial methods is useful: ADV and MADV reduce DPgap but badly worsen TPRgap and FPRgap in multi-label SER, which is a non-obvious finding. The imbalance-ratio analysis in Table III is also clean.

Now the soft spots, in order of importance.

The central selection rule in Section VI-C is not statistically supported. Every model is trained once. No seeds, no confidence intervals, no significance tests. In Tables IV and V, the gaps between ERM and the recommended methods are frequently 0.01 or 0.02, and in several cells they are exact ties. For example, LfF versus ERM on CREMA-D with WavLM ties on FPRgap, F1gap, and DPgap; GADRO ties FPRgap. Section VI-C says eight methods reduced all fairness gaps, but Table IV shows equality on some metrics. That is overclaiming independent of variance. Since GADRO and LfF involve auxiliary optimization that tends to be seed-sensitive, a 0.01 advantage could easily flip. The recommendation should be softened or backed by repeated runs.

Two smaller issues. The test set is filtered to utterances with a dominant emotion, which narrows the multi-label claim. GR's loss directly contains TPRgap and FPRgap, so its fairness gains are partly by construction; the paper should say so rather than present GR as a neutral benchmark entry. Code is also not released, which makes the single-run results harder to audit.

Who this is for: anyone choosing a debiasing method for multi-label SER, and people working on fairness evaluation methodology. It deserves serious peer review, but the authors should be asked for seeded runs, or at least for the top claims to be verified across several seeds, and for the language about ties to be reworded. With those changes, this would be a solid reference point.

Referee Report

3 major / 4 minor

Summary. The paper introduces EMO-Debias, a benchmark comparing 13 gender-debiasing methods for multi-label speech emotion recognition (SER). Twelve methods are adapted from prior work and one, Gap Regularization (GR), is proposed. Experiments use MSP-Podcast and CREMA-D with WavLM and XLSR frozen encoders under controlled gender-imbalance ratios (1:1 through 1:40), and report macro-F1, Hamming accuracy, and four fairness gaps (TPRgap, FPRgap, F1gap, DPgap). The central claim, stated in Section VI-C, is that GADRO (with bias supervision) and LfF (without bias supervision) consistently outperform the ERM baseline on all four fairness metrics for both backbones and both datasets, with only minimal accuracy/F1 loss, making them the most robust debiasing methods. The paper also reports that adversarial methods struggle in multi-label SER and that pre-processing methods such as RW and DS perform well on fairness at some accuracy cost.

Significance. If the central ranking were statistically established, this benchmark would be a useful resource for practitioners selecting debiasing methods for multi-label SER. The paper covers a broad set of method families, two SSL backbones, two dataset types, and a controlled imbalance protocol, and the fairness metrics are clearly defined. The promise of released code is another strength. However, the current evidence is not sufficient to support the headline recommendation: all models are trained once, several decisive comparisons differ by only 0.01–0.02, and the paper's own tables contain ties that contradict the 'outperform on all four metrics' wording. The proposed GR method directly optimizes two of the four reported fairness metrics, so its placement in the benchmark needs a separate, non-circular evaluation. The benchmark is potentially valuable, but the claims need revision and the experiments need multiple seeds before the ranking can be accepted.

major comments (3)

[Section VI-C, Tables IV and V] The claim that 'GADRO (with bias supervision) and LfF (without bias supervision) consistently outperform ERM on all four fairness metrics for both WavLM and XLSR' is not supported by the reported point estimates. In Table IV (WavLM, CREMA-D), LfF versus ERM gives FPRgap 0.13 vs 0.13, F1gap 0.12 vs 0.12, and DPgap 0.10 vs 0.10, with only TPRgap lower by 0.01; GADRO versus ERM gives FPRgap 0.13 vs 0.13. In Table V (XLSR, CREMA-D), LfF versus ERM gives TPRgap 0.25 vs 0.25, F1gap 0.12 vs 0.12, and DPgap 0.09 vs 0.09, with only FPRgap lower by 0.01. These are ties, not strict reductions, so the conjunction 'all four fairness metrics' is false even before considering statistical uncertainty. The text should be revised to state the exact metric-wise comparison and to handle ties explicitly.
[Section VI, Tables III–V] Every model is trained once, with no seeds, confidence intervals, or significance tests reported. Many of the differences that separate methods are 0.01 or 0.02 in F1 or fairness gaps; for example, the LfF versus ERM TPRgap advantage on MSP-Podcast is 0.18 vs 0.19 in Table IV and 0.14 vs 0.18 in Table V. Since GADRO and LfF involve auxiliary optimization (group-robust weighting and a second biased model) that is typically more seed-sensitive than plain ERM, a single run cannot establish 'consistently outperform' or 'most robust.' Please add at least three to five seeds and report means with standard deviations or confidence intervals, and use an appropriate paired test or overlap criterion before drawing ranking conclusions.
[Section V-E1 and Table II] The proposed GR method includes TPRgap and FPRgap directly in its training objective, so its improved values on those two metrics are expected by construction. The paper does not specify how the gap terms are computed during training (e.g., per batch, on the training or development set) or whether the reported fairness metrics are computed on held-out data. This overlap between the training objective and the evaluation metric makes GR's benchmark ranking not directly comparable with methods that optimize only the classification loss. A separate evaluation protocol should be described, and the discussion should not present GR's fairness gains as evidence of generalizable debiasing without a non-circular test.

minor comments (4)

[Section V-A, Equation (3)] The notation '||h^T_Ai h_Aj||_2 / BD {i≠j}' is unclear: the expression in the denominator appears to be the batch size, but 'BD' is not defined, and the norm of a matrix should be a Frobenius norm rather than an L2 vector norm. Please rewrite this equation with a defined denominator and consistent norm notation.
[Sections VI-C, VI-E, and VII] The paper gives conflicting conclusions about the best method without bias supervision: Section VI-C says LfF is most robust, Section VI-E says LVR and SiH achieve the best trade-off, and Section VII says the modified LVR achieves the best trade-off. These statements should be reconciled with a single, clearly stated criterion.
[Section VI-C] The statement that 'eight out of the thirteen experimental results ... successfully reduced bias across all fairness metrics' should be replaced with a precise metric-by-metric comparison, since several methods are equal to ERM on some metrics rather than strictly reduced; for example, GR on CREMA-D in Table IV has FPRgap 0.13, F1gap 0.12, and DPgap 0.10, all equal to ERM.
[Tables IV and V] Bold and underline are used to mark the best and second-best values in each column, but without confidence intervals or significance tests, these markers give a false impression of statistical ranking. It would be clearer to report mean ± std for multiple runs and to avoid bolding differences that are within noise.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor overlap between GR's training objective and its reported fairness metrics; otherwise the benchmark comparison is self-contained and not circular.

self definitional [Section V-E (Gap Regularization), Table II; discussed in Section VI-C]
"GR [Ours] CE (y, ˆy) + λ GR(TPRgap+FPRgap) ... "We design an auxiliary loss to penalize discrepancies in multi-label fairness metrics in section IV-D, the TPR gap and FPR gap, as shown in Table II." ... "In Table IV, eight out of the thirteen experimental results (specifically, GR, DS, RW, GADRO, LfF, LVR, SiH, and DisEnt) successfully reduced bias across all fairness metrics in both databases""

GR's objective is defined as CE plus a penalty on (TPRgap+FPRgap), which are exactly the fairness metrics used to declare its success. Thus the reported decrease in TPRgap/FPRgap for GR relative to ERM is a direct result of optimizing for those quantities during training, rather than an independent empirical finding. This overlap affects only the GR row; the ranking of GADRO, LfF, and the other adapted methods is not defined in terms of the measured fairness gaps.

full rationale

The paper's central claim is a comparative benchmark: GADRO and LfF outperform ERM on measured fairness and accuracy tables, with the compared methods being standard prior work and ERM an external baseline. That claim does not reduce to an equation or a fitted parameter; it is an empirical observation, albeit one that would be stronger with multiple seeds and confidence intervals. The only place where the evaluation metric enters the training objective is the proposed GR method, whose TPRgap/FPRgap improvements are expected from its own loss definition. This is a localized, mild circularity and does not invalidate the ranking of the other twelve methods. Self-citations to EMO-SUPERB and the authors' prior work are used for dataset partitioning, framework, and related-work context, and are not load-bearing for the debiasing comparison. Therefore the overall circularity is low.

Assumptions & free parameters 12 free parameters · 5 assumptions · 0 invented entities

The central empirical claims rest on a handful of tuned hyperparameters and on assumptions that binary gender, dominant-emotion filtering, and simulated imbalance provide a valid fairness evaluation. No new entities are introduced; the proposed GR is an auxiliary loss term.

free parameters (12)

lambda_GR = 4
Weight of the proposed gap-regularization loss; hand-set in Section V-E1.
lambda_GD = 4 (CREMA-D), 20 (MSP-Podcast)
Group adjustment strength for GADRO, chosen per dataset in Section V-C; tuned values differ by dataset.
lambda_adv = 3.2
Adversarial loss weight in ADV and MADV, stated in Section V-A2.
lambda_diff = 0.2
Orthogonality loss weight for MADV, stated in Section V-A2.
gamma_BLIND = 0.7
Down-weighting exponent in BLIND, stated in Section V-D4.
lambda_B = 1
Weight for the bias or success detector loss in BLIND, stated in Section V-D4.
q_GCE = 0.7
GCE exponent for biased learners, following prior work as noted in Section V-D1.
alpha_LfF = 0.7
EMA decay for relative difficulty in LfF, Section V-D1.
r_SiH = 0.7
Focal emphasis exponent for SiH, Section V-D2.
omega_LVR = 0.3
EMA momentum for class centers in LVR, Section V-E2.
lambda_LVR = 0.1
Weight for low-variance regularization, Section V-E2.
learning_rate = 1e-4 WavLM, 5e-4 XLSR
Learning rates chosen by search, Section IV-C.

assumptions (5)

domain assumption Gender is a binary attribute and the only protected group is male versus female.
Section VIII acknowledges this limitation; all fairness metrics and debiasing objectives are computed on two groups.
domain assumption An utterance's ground truth is the distribution of annotator votes, and evaluation can be restricted to utterances where one emotion exceeds 0.5 while still representing multi-label SER.
Section IV-B: distributional labels from rating frequencies and the dominant-emotion filter define the task and the test set.
domain assumption Frozen WavLM and XLSR representations contain enough emotion and gender information that the linear head and debiasing modules can meaningfully change fairness.
Section IV-A: the SSL encoder is frozen and only a linear head plus method-specific components are trained.
domain assumption Biased learners (LfF, SiH, DisEnt) can identify gender-correlated shortcuts without explicit gender labels.
Section V-D: these methods rely on a GCE-trained biased model capturing bias-aligned examples; this is imported from the cited prior work.
domain assumption The simulated per-emotion gender skew, with majority gender varying by emotion, approximates real-world skewed data distributions.
Section III and IV-B and Table I: imbalance ratios are constructed per emotion; Section VIII admits they may not fully reflect real-world distributions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of EMO-Debias: Benchmarking Gender Debiasing Techniques in Multi-Label Speech Emotion Recognition." pith.science (2026). https://pith.science/paper/QRVRGPVF

@misc{pith2026250604652,
  author       = {Pith},
  title        = {Pith review of: EMO-Debias: Benchmarking Gender Debiasing Techniques in Multi-Label Speech Emotion Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QRVRGPVF}},
  note         = {Machine review of arXiv:2506.04652}
}

read the original abstract

Speech emotion recognition (SER) systems often exhibit gender bias. However, the effectiveness and robustness of existing debiasing methods in such multi-label scenarios remain underexplored. To address this gap, we present EMO-Debias, a large-scale comparison of 13 debiasing methods applied to multi-label SER. Our study encompasses techniques from pre-processing, regularization, adversarial learning, biased learners, and distributionally robust optimization. Experiments conducted on acted and naturalistic emotion datasets, using WavLM and XLSR representations, evaluate each method under conditions of gender imbalance. Our analysis quantifies the trade-offs between fairness and accuracy, identifying which approaches consistently reduce gender performance gaps without compromising overall model performance. The findings provide actionable insights for selecting effective debiasing strategies and highlight the impact of dataset distributions.

Figures

Figures reproduced from arXiv: 2506.04652 by the authors.

**Figure 1.** Hamming accuracy (ACC) and TPRgap under various gender-biased data distribution conditions. The X-axis represents the Ratio. performance. We use hyperparameter λLV R = 0.1 and ω = 0.3, following previous work [52]. VI. EXPERIMENTAL RESULTS Since gender annotations may not always be available in realworld applications, we analyze two scenarios: one with demographic information (denoted as bias supervision, BS) and o… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 56 canonical work pages

[1]

Geneva: World Health Organization, 2021

World Health Organization, Comprehensive Mental Health Action Plan 2013–2030 . Geneva: World Health Organization, 2021

work page 2013
[2]

Automatic Speech Emotion Recognition Using Machine Learning,

L. Kerkeni et al. , “Automatic Speech Emotion Recognition Using Machine Learning,” in Social Media and Machine Learning , 2019

work page 2019
[3]

Speech Emotion Recognition using Supervised Deep Recurrent System for Mental Health Monitoring,

N. Elsayed et al., “Speech Emotion Recognition using Supervised Deep Recurrent System for Mental Health Monitoring,” in 2022 IEEE 8th W orld F orum on Internet of Things (WF-IoT) , 2022

work page 2022
[4]

Emotion V ariation Detection in Discrete English Speech: A Wavelet Transform Use Case in Mental Health Monito ring,

A. Adeleye et al. , “Emotion V ariation Detection in Discrete English Speech: A Wavelet Transform Use Case in Mental Health Monito ring,” in Proceedings of the 2024 Australasian Computer Science W eek , 2024

work page 2024
[5]

Emo-bias: A Large Scale Evaluation of Social Bias on Speech Emotion Recognition,

Y .-C. Lin et al., “Emo-bias: A Large Scale Evaluation of Social Bias on Speech Emotion Recognition,” in Interspeech 2024 , 2024

work page 2024
[6]

Gender De-Biasing in Speech Emotion Recognition,

C. Gorrostieta, R. Lotﬁan, K. Taylor, R. Brutti, and J. Ka ne, “Gender De-Biasing in Speech Emotion Recognition,” in Interspeech 2019, 2019

work page 2019
[7]

On the social bias of speech self-supervised models,

Y .-C. Lin et al. , “On the social bias of speech self-supervised models,” in Interspeech 2024 , 2024

work page 2024
[8]

Mitigating subgrou p dispari- ties in multi-label speech emotion recognition: A pseudo-l abeling and unsupervised learning approach,

Y .-C. Lin, H.-C. Chou, and H. yi Lee, “Mitigating subgrou p dispari- ties in multi-label speech emotion recognition: A pseudo-l abeling and unsupervised learning approach,” 2025

work page 2025

Show all 56 references

[9]

Explor- ing data augmentation in bias mitigation against non-nativ e-accented speech,

Y . Zhang, A. Herygers, T. Patel, Z. Y ue, and O. Scharenbor g, “Explor- ing data augmentation in bias mitigation against non-nativ e-accented speech,” 2023

2023
[10]

Towards comprehensive subgroup performance a nalysis in speech models,

A. Koudounas, E. Pastor, G. Attanasio, V . Mazzia, M. Gio llo, T. Gueu- dre, E. Reale, L. Cagliero, S. Cumani, L. de Alfaro, E. Barali s, and D. Amberti, “Towards comprehensive subgroup performance a nalysis in speech models,” IEEE/ACM Transactions on Audio, Speech, and Langua...

2024
[11]

User-Level Differe ntial Privacy against Attribute Inference Attack of Speech Emotion Recog nition on Federated Learning,

T. Feng, R. Peri, and S. Narayanan, “User-Level Differe ntial Privacy against Attribute Inference Attack of Speech Emotion Recog nition on Federated Learning,” in Interspeech 2022 , 2022

2022
[12]

Achieving Fair Speech Emoti on Recogni- tion via Perceptual Fairness,

W.-S. Chien and C.-C. Lee, “Achieving Fair Speech Emoti on Recogni- tion via Perceptual Fairness,” in ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICA SSP), 2023

2023
[13]

Balancing S peaker-Rater Fairness for Gender-Neutral Speech Emotion Recognition,

W.-S. Chien, S. G. Upadhyay, and C.-C. Lee, “Balancing S peaker-Rater Fairness for Gender-Neutral Speech Emotion Recognition,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Spe ech and Signal Processing (ICASSP) , 2024

2024
[14]

Is It Still F air? Investi- gating Gender Fairness in Cross-Corpus Speech Emotion Reco gnition,

S. G. Upadhyay, W.-S. Chien, and C.-C. Lee, “Is It Still F air? Investi- gating Gender Fairness in Cross-Corpus Speech Emotion Reco gnition,” in ICASSP 2025 - 2025 IEEE International Conference on Acousti cs, Speech and Signal Processing (ICASSP) , 2025

2025
[15]

Exploiting Co-occ urrence Frequency of Emotions in Perceptual Evaluations To Train A S peech Emotion Classiﬁer,

H.-C. Chou, C.-C. Lee, and C. Busso, “Exploiting Co-occ urrence Frequency of Emotions in Perceptual Evaluations To Train A S peech Emotion Classiﬁer,” in Interspeech 2022 , 2022, pp. 161–165

2022
[16]

Open-Emotion: A Reproducible EMO-Superb For Speech Emotion Recognition Systems,

H. Wu et al., “Open-Emotion: A Reproducible EMO-Superb For Speech Emotion Recognition Systems,” in 2024 IEEE Spoken Language Tech- nology W orkshop (SLT), 2024

2024
[17]

Building Naturalistic Emotion ally Balanced Speech Corpus by Retrieving Emotional Speech from Existing Podcast Recordings,

R. Lotﬁan and C. Busso, “Building Naturalistic Emotion ally Balanced Speech Corpus by Retrieving Emotional Speech from Existing Podcast Recordings,” IEEE Transactions on Affective Computing , 2019

2019
[18]

CREMA-D: Crowd-Sourced Emotional Multimodal Ac tors Dataset,

H. Cao, D. G. Cooper, M. K. Keutmann, R. C. Gur, A. Nenkova , and R. V erma, “CREMA-D: Crowd-Sourced Emotional Multimodal Ac tors Dataset,” IEEE Transactions on Affective Computing , 2014

2014
[19]

Emotion R ecognition Systems Must Embrace Ambiguity,

J. Wu, T. Dang, V . Sethu, and E. Ambikairajah, “Emotion R ecognition Systems Must Embrace Ambiguity,” in 2024 12th International Confer- ence on Affective Computing and Intelligent Interaction W o rkshops and Demos (ACIIW) , 2024, pp. 166–170

2024
[20]

Embracing Ambiguity And Subjectivity Using The All-Inclusive Aggregation Rule For Evaluating Multi-L abel Speech Emotion Recognition Systems,

H.-C. Chou et al. , “Embracing Ambiguity And Subjectivity Using The All-Inclusive Aggregation Rule For Evaluating Multi-L abel Speech Emotion Recognition Systems,” in 2024 IEEE Spoken Language Tech- nology W orkshop (SLT), 2024

2024
[21]

Multi-Label Emotion Recognition of Korean Speech Data Using Deep Fusion Models,

S. Park et al., “Multi-Label Emotion Recognition of Korean Speech Data Using Deep Fusion Models,” Applied Sciences , vol. 14, no. 17, 2024

2024
[22]

Self-report captures 27 dis tinct categories of emotion bridged by continuous gradients,

A. S. Cowen and D. Keltner, “Self-report captures 27 dis tinct categories of emotion bridged by continuous gradients,” Proceedings of the Na- tional Academy of Sciences , vol. 114, no. 38, pp. E7900–E7909, 2017

2017
[23]

Semantic Space Theory: A Computational Approach t o Emotion,

——, “Semantic Space Theory: A Computational Approach t o Emotion,” Trends in Cognitive Sciences , 2021

2021
[24]

An Int er-Speaker Fairness-Aware Speech Emotion Regression Framework,

H.-H. Chou, W.-S. Chien, Y .-T. Wu, and C.-C. Lee, “An Int er-Speaker Fairness-Aware Speech Emotion Regression Framework,” in Interspeech 2024, 2024

2024
[25]

Emo-superb: An in-depth look at sp eech emotion recognition,

H. Wu, H.-C. Chou, K.-W. Chang, L. Goncalves, J. Du, J.-S . R. Jang, C.-C. Lee, and H.-Y . Lee, “Emo-superb: An in-depth look at sp eech emotion recognition,” 2024

2024
[26]

A tiny whisper-ser: Unifying automatic sp eech recognition and multi-label speech emotion recognition tasks,

H.-C. Chou, “A tiny whisper-ser: Unifying automatic sp eech recognition and multi-label speech emotion recognition tasks,” in 2024 Asia Paciﬁc Signal and Information Processing Association Annual Summ it and Conference (APSIPA ASC) , 2024, pp. 1–6

2024
[27]

Stimulus Modality Mat ters: Impact of Perceptual Evaluations from Different Modalities on Spe ech Emo- tion Recognition System Performance,

H.-C. Chou, H. Wu, and C.-C. Lee, “Stimulus Modality Mat ters: Impact of Perceptual Evaluations from Different Modalities on Spe ech Emo- tion Recognition System Performance,” in ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal P rocessing (ICAS...

2025
[28]

No Sampl e Left Behind: Towards a Comprehensive Evaluation of Speech Emotion Recog nition Systems,

P . Riera, L. Ferrer, A. Gravano, and L. Gauder, “No Sampl e Left Behind: Towards a Comprehensive Evaluation of Speech Emotion Recog nition Systems,” in Proc. SMM19, W orkshop on Speech, Music and Mind 2019 , 2019

2019
[29]

Minority Views Matter: Evaluating Speech Emo tion Classiﬁers with Human Subjective Annotations by an All-Inc lusive Aggregation Rule,

H.-C. Chou, L. Goncalves, S.-G. Leem, A. N. Salman, C.-C . Lee, and C. Busso, “Minority Views Matter: Evaluating Speech Emo tion Classiﬁers with Human Subjective Annotations by an All-Inc lusive Aggregation Rule,” IEEE Transactions on Affective Computing , 2024

2024
[30]

Common voice : A massively-multilingual speech corpus,

R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. H enretty, R. Morais, L. Saunders, F. Tyers, and G. Weber, “Common voice : A massively-multilingual speech corpus,” in Proceedings of the Twelfth Language Resources and Evaluation Conference, N. Calzolari, F. B´ echet, P...

2020
[31]

Darpa timit acoustic-phonetic continous speech corpus cd-rom. nist speech disc 1-1.1,

J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, an d D. S. Pallett, “Darpa timit acoustic-phonetic continous speech corpus cd-rom. nist speech disc 1-1.1,” NASA STI/Recon technical report n , vol. 93, p. 27403, 1993

1993
[32]

A Large-Scale Evaluation of Speech Foundation Models,

S.-w. Y ang et al. , “A Large-Scale Evaluation of Speech Foundation Models,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024

2024
[33]

Decoupled Weight Decay Re gularization,

I. Loshchilov and F. Hutter, “Decoupled Weight Decay Re gularization,” in International Conference on Learning Representations , 2019

2019
[34]

Msp-podcast ser c hallenge 2024: L’antenne du ventoux multimodal self-supervised lea rning for speech emotion recognition,

J. Duret, Y . Est` eve, and M. Rouvier, “Msp-podcast ser c hallenge 2024: L’antenne du ventoux multimodal self-supervised lea rning for speech emotion recognition,” in The Speaker and Language Recognition W orkshop (Odyssey 2024), 2024, pp. 309–314

2024
[35]

Improv ing speech emotion recognition in under-resourced languages via spee ch-to-speech translation with bootstrapping data selection,

H.-C. Lin, Y .-C. Lin, H.-C. Chou, and H.-y. Lee, “Improv ing speech emotion recognition in under-resourced languages via spee ch-to-speech translation with bootstrapping data selection,” in ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Sig nal Pro-...

2025
[36]

Equality of Opportunity in Supervised Learning,

M. Hardt, E. Price, E. Price, and N. Srebro, “Equality of Opportunity in Supervised Learning,” in Advances in Neural Information Processing Systems, D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, Eds., vol. 29. Curran Associates, Inc., 2016

2016
[37]

Diverse Adversaries fo r Mitigating Bias in Training,

X. Han, T. Baldwin, and T. Cohn, “Diverse Adversaries fo r Mitigating Bias in Training,” in Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main V olume, 2021

2021
[38]

Who gets the beneﬁt of the doubt? racial bias in machine learning algorithms applied to secondary school math education,

H. Jeong, M. D. Wu, N. Dasgupta, M. M´ edard, and F. Calmon , “Who gets the beneﬁt of the doubt? racial bias in machine learning algorithms applied to secondary school math education,” Math AI for Education: Bridging the Gap Between Research and Smart Education , 2022

2022
[39]

Soft-prompt tuning for large language models to evaluate bias,

J.-J. Tian, D. Emerson, S. Z. Miyandoab, D. Pandya, L. Se yyed- Kalantari, and F. K. Khattak, “Soft-prompt tuning for large language models to evaluate bias,” 2024

2024
[40]

Debiasing with Su fﬁcient Pro- jection: A General Theoretical Framework for V ector Repres entations,

E. Shi, L. Ding, L. Kong, and B. Jiang, “Debiasing with Su fﬁcient Pro- jection: A General Theoretical Framework for V ector Repres entations,” in Proceedings of the 2024 Conference of the North American Cha pter of the Association for Computational Linguistics: Human Lang uag...

2024
[41]

Clas s-Balanced Loss Based on Effective Number of Samples,

Y . Cui, M. Jia, T.-Y . Lin, Y . Song, and S. Belongie, “Clas s-Balanced Loss Based on Effective Number of Samples,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni tion (CVPR), June 2019

2019
[42]

Towards robust and privacy-preserving text representat ions,

Y . Li et al., “Towards robust and privacy-preserving text representat ions,” in Proceedings of the 56th Annual Meeting of the Association fo r Computational Linguistics (V olume 2: Short Papers) , 2018

2018
[43]

Adversarial Removal of Demo graphic Attributes from Text Data,

Y . Elazar and Y . Goldberg, “Adversarial Removal of Demo graphic Attributes from Text Data,” in Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , 2018

2018
[44]

Diverse adversaries fo r mitigating bias in training,

X. Han, T. Baldwin, and T. Cohn, “Diverse adversaries fo r mitigating bias in training,” in Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main V olume, P . Merlo, J. Tiedemann, and R. Tsarfaty, Eds. Online: Associ ...

2021
[45]

Data preprocessing techniq ues for classi- ﬁcation without discrimination,

F. Kamiran and T. Calders, “Data preprocessing techniq ues for classi- ﬁcation without discrimination,” Knowledge and Information Systems , 2012

2012
[46]

BLIND: Bias removal with no de mograph- ics,

H. Orgad and Y . Belinkov, “BLIND: Bias removal with no de mograph- ics,” in Proceedings of the 61st Annual Meeting of the Association fo r Computational Linguistics (V olume 1: Long Papers) , 2023

2023
[47]

Robust Solutions of Optimization Problems Affected by Uncertain Probabilities,

A. Ben-Tal et al., “Robust Solutions of Optimization Problems Affected by Uncertain Probabilities,” Management Science, 2013

2013
[48]

Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generaliza- tion

S. Sagawa et al. , “Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generaliza- tion.” in International Conference on Learning Representations , 2020

2020
[49]

Learning from failure: training debiased classiﬁer from biased classiﬁer,

J. Nam et al. , “Learning from failure: training debiased classiﬁer from biased classiﬁer,” in Proceedings of the 34th International Conference on Neural Information Processing Systems , 2020

2020
[50]

Signal Is Harder To Learn Than Bias: Debiasing with Focal Loss,

M. V andenhirtz et al., “Signal Is Harder To Learn Than Bias: Debiasing with Focal Loss,” in ICLR 2023 W orkshop on Domain Generalization (DG), 2023

2023
[51]

Learning Debiased Representation via Disentangled Feature Augmentation,

J. Lee et al. , “Learning Debiased Representation via Disentangled Feature Augmentation,” in Advances in Neural Information Processing Systems, M. Ranzato et al., Eds., vol. 34. Curran Associates, Inc., 2021, pp. 25 123–25 133

2021
[52]

Unlabeled Debiasing in Downstream Tasks via Class-wise Low V ariance Regularization,

S. Masoudian et al. , “Unlabeled Debiasing in Downstream Tasks via Class-wise Low V ariance Regularization,” in Proceedings of the 2024 Conference on Empirical Methods in Natural Language Proces sing, Miami, Florida, USA, 2024

2024
[53]

Generalized Cross Entropy Los s for Training Deep Neural Networks with Noisy Labels,

Z. Zhang and M. Sabuncu, “Generalized Cross Entropy Los s for Training Deep Neural Networks with Noisy Labels,” in Advances in Neural Information Processing Systems , 2018

2018
[54]

Beyond the binary: Li mitations and possibilities of gender-related speech technology res earch,

A. Sanchez, A. Ross, and N. Markl, “Beyond the binary: Li mitations and possibilities of gender-related speech technology res earch,” in 2024 IEEE Spoken Language Technology W orkshop (SLT), 2024, pp. 526–532

2024
[55]

Listen and speak fairly: a study on semantic gender bias in speech integrated large language models,

Y .-C. Lin et al. , “Listen and speak fairly: a study on semantic gender bias in speech integrated large language models,” in 2024 IEEE Spoken Language Technology W orkshop (SLT), 2024

2024
[56]

Spoken stereoset: on evaluating social bias towar d speaker in speech large language models,

——, “Spoken stereoset: on evaluating social bias towar d speaker in speech large language models,” in 2024 IEEE Spoken Language Technology W orkshop (SLT), 2024

2024

Pith tools

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

[1] [1]

Geneva: World Health Organization, 2021

World Health Organization, Comprehensive Mental Health Action Plan 2013–2030 . Geneva: World Health Organization, 2021

work page 2013

[2] [2]

Automatic Speech Emotion Recognition Using Machine Learning,

L. Kerkeni et al. , “Automatic Speech Emotion Recognition Using Machine Learning,” in Social Media and Machine Learning , 2019

work page 2019

[3] [3]

Speech Emotion Recognition using Supervised Deep Recurrent System for Mental Health Monitoring,

N. Elsayed et al., “Speech Emotion Recognition using Supervised Deep Recurrent System for Mental Health Monitoring,” in 2022 IEEE 8th W orld F orum on Internet of Things (WF-IoT) , 2022

work page 2022

[4] [4]

Emotion V ariation Detection in Discrete English Speech: A Wavelet Transform Use Case in Mental Health Monito ring,

A. Adeleye et al. , “Emotion V ariation Detection in Discrete English Speech: A Wavelet Transform Use Case in Mental Health Monito ring,” in Proceedings of the 2024 Australasian Computer Science W eek , 2024

work page 2024

[5] [5]

Emo-bias: A Large Scale Evaluation of Social Bias on Speech Emotion Recognition,

Y .-C. Lin et al., “Emo-bias: A Large Scale Evaluation of Social Bias on Speech Emotion Recognition,” in Interspeech 2024 , 2024

work page 2024

[6] [6]

Gender De-Biasing in Speech Emotion Recognition,

C. Gorrostieta, R. Lotﬁan, K. Taylor, R. Brutti, and J. Ka ne, “Gender De-Biasing in Speech Emotion Recognition,” in Interspeech 2019, 2019

work page 2019

[7] [7]

On the social bias of speech self-supervised models,

Y .-C. Lin et al. , “On the social bias of speech self-supervised models,” in Interspeech 2024 , 2024

work page 2024

[8] [8]

Mitigating subgrou p dispari- ties in multi-label speech emotion recognition: A pseudo-l abeling and unsupervised learning approach,

Y .-C. Lin, H.-C. Chou, and H. yi Lee, “Mitigating subgrou p dispari- ties in multi-label speech emotion recognition: A pseudo-l abeling and unsupervised learning approach,” 2025

work page 2025