Pith. sign in

REVIEW 5 major objections 6 minor 47 references

Mitigating Subgroup Disparities in Multi-Label Speech Emotion Recognition: A Pseudo-Labeling and Unsupervised Learning Approach

T0 review · 5 major / 6 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 Speech emotion bias can be cut by 28% without demographic labels.

desk verdict A timely label-free debiasing study for multi-label SER whose headline claim hinges on an unspecified fairness evaluation group. read the letter →

arxiv 2505.14449 v3 pith:YHWKXVB4 submitted 2025-05-20 eess.AS cs.CLcs.SD

Yi-Cheng Lin , Huang-Cheng Chou , Hung-yi Lee This is my paper · ORCID
classification eess.AScs.CLcs.SD
keywords speechemotionrecognitionsubgroupdisparityfairnesspseudo-labelingunsupervisedclusteringmulti-labelclassificationdemographicinferenceCREMA-D
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 subgroup disparities in multi-label speech emotion recognition can be reduced without collecting sensitive demographic labels. It answers yes: an Implicit Demography Inference module infers group membership either from pseudo-labels produced by a pretrained gender detector or from k-means clusters of ECAPA-TDNN speech embeddings, and feeds those inferred groups into standard debiasing losses. On CREMA-D, pseudo-labeling improves the equal-opportunity gap ($\mathrm{TPR_{gap}}$) by 28.78% and the demographic-parity gap ($\mathrm{DP_{gap}}$) by 29.13%, with macro-F1 falling 4.38% relative to the unconstrained baseline. The unsupervised variant, with 16 clusters, improves $\mathrm{TPR_{gap}}$ by 20.05% and $\mathrm{DP_{gap}}$ by 4.61%, and transfers to race and age subgroups. A sympathetic reader should care because demographic labels are often unavailable or privacy-protected, so debiasing that works purely from speech-derived proxies could be deployed where explicit demographic data is not.

What carries the argument

The load-bearing object is the Implicit Demography Inference (IDI) module, which produces proxy group labels in one of two ways: pseudo-labeling, using a pretrained age/gender transformer that reaches 94.4% accuracy on CREMA-D, or unsupervised clustering, where k-means is applied to ECAPA-TDNN utterance embeddings with cluster counts 2, 4, 8, 16, and 32. These proxy labels replace true demographic labels inside four debiasing losses: reweighting, downsampling, group DRO, and group-aware DRO. The t-SNE visualisation of the embeddings, which separates male and female speakers, is the evidence offered that the clusters carry demographic information.

What would settle it

Replace k-means clusters with random group labels of the same sizes and rerun the debiasing training; if TPR-gap improvements of similar magnitude persist, the effect does not come from demographic structure in the embeddings.

Watch

Extended reading notes

Core claim

The paper's central claim is that demographic supervision is not strictly necessary for reducing subgroup disparities in multi-label SER. Using pseudo-labels from a pretrained age-gender classifier, the proposed IDI pipeline lowers $\mathrm{TPR_{gap}}$ from 0.278 to 0.198 and $\mathrm{DP_{gap}}$ from 0.103 to 0.073 relative to ERM, a 28.78% and 29.13% improvement, with macro-F1 falling from 0.651 to 0.623. With k-means clustering at $k=16$ on ECAPA-TDNN embeddings, $\mathrm{TPR_{gap}}$ drops to 0.222 and $\mathrm{DP_{gap}}$ to 0.098, improvements of 20.05% and 4.61%. The same unsupervised clusters also reduce the TPR gap for race (16.67% at $k=16$) and age (12.75% at $k=16$), which the paper reads as evidence that label-free demography inference generalizes beyond gender. The authors present these results as a more usable trade-off than prior unsupervised debiasing methods (LfF, DisEnt) and than probabilistic pseudo-labeling (PGDRO), which they find does not reduce gaps relative to ERM.

Load-bearing premise

The entire approach rests on the assumption that the inferred groups—pseudo-labels or k-means clusters—actually correspond to the demographic attribute being debiased; if the clusters encode emotion, speaker identity, or recording conditions instead, the reported race and age improvements would be coincidental.

Editorial extensions

If this is right

  • SER systems can be made fairer in settings where collecting gender, race, or age labels is ethically or legally difficult.
  • The unsupervised variant offers a path to debias for attributes whose labels were never collected, such as race and age, provided the clusters align with those attributes.
  • Because IDI is agnostic to the debiasing loss, it can be combined with newer debiasing objectives as they are introduced.
  • The multi-label formulation evaluates on the full test set rather than a cleaned subset, so the fairness gains are not obtained by discarding ambiguous utterances.

Reading between the lines

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

  • The cluster-alignment premise needs a direct test: the paper validates k=16 clusters against gender via t-SNE but never quantifies cluster purity against the true race and age labels, so the race and age gains could be an artifact of clusters tracking gender, emotion, or recording conditions.
  • The Random baseline's fairness scores (TPR gap 0.060, DP gap 0.015) are the best of any model, which warns that low accuracy can trivially satisfy the gap metrics; the reported gains should be read against the accuracy-fairness frontier, not gap reduction alone.
  • A natural stress test is to degrade the accuracy of the pretrained gender classifier and measure how quickly the pseudo-labeling fairness gains decay.
  • A testable extension is to apply IDI to naturalistic, non-acted emotion corpora where demographic imbalance arises organically, to check whether cluster-based proxies survive outside CREMA-D.
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

5 major / 6 minor

Summary. The paper proposes an Implicit Demography Inference (IDI) module for multi-label speech emotion recognition (SER) that replaces explicit demographic labels with inferred group labels, obtained either from a pretrained gender/age classifier (pseudo-labeling) or from k-means clustering of ECAPA-TDNN embeddings. These inferred labels are fed into existing debiasing techniques (Reweighting, Downsampling, GDRO, GADRO), and the method is evaluated on CREMA-D under the EMO-SUPERB multi-label protocol with simulated imbalanced training data. The authors report fairness gains measured by TPR gap and demographic parity gap, with modest SER accuracy costs, and they claim generalization to race and age subgroups.

Significance. If the central claim holds, the paper is a useful contribution because it addresses the practical problem of debiasing SER without demographic annotations, evaluates in a multi-label setup, and compares against several external baselines (LfF, DisEnt, PGDRO) as well as supervised debiasing with true labels. The use of the original test set rather than a filtered subset is a strength, and the inclusion of a simulated biased-training setup follows good evaluation practice. However, the current evidence is not yet convincing because the fairness metrics may be computed over inferred groups rather than true protected attributes, and the unsupervised claims for race and age lack a validation that the clusters align with those attributes.

major comments (5)
  1. [Sec. 4.3 and Tables 1-2] The paper never states what the group set G is when computing TPRgap and DPgap for the IDI rows. Equations (4)-(5) define the metrics for a generic set G, but in Tables 1 and 2 it is not specified whether G is the true gender/race/age label, the pseudo-label, or the k-means cluster assignment. If the metrics are evaluated on the same inferred groups used during training, then any method that equalizes behavior across those groups will reduce TPRgap/DPgap by construction, and the reported fairness gains would not demonstrate reduced disparity with respect to true demographic attributes. Please specify explicitly, for every row in both tables, which group partition is used in the evaluation, and if the IDI rows are evaluated on inferred groups, also report the metrics computed on the true test-set demographic labels.
  2. [Secs. 5.2-5.3 and Fig. 4] The unsupervised method's race and age results are not supported by any evidence that the k-means clusters correspond to race or age. Fig. 4 shows t-SNE separation between male and female for k=16, but Table 2 reports TPRgap gains for race and age without any analogous cluster-to-attribute check. If the clusters encode emotion, speaker identity, or recording conditions, the reported race and age fairness improvements could be coincidental. Please add a quantitative cluster-membership analysis against the true race and age labels (e.g., adjusted Rand index or contingency tables) or substantially temper the claims in Sec. 5.3.
  3. [Table 1 and Sec. 5.1] The Random baseline achieves the best fairness scores of any model (TPRgap 0.060, DPgap 0.015, versus ERM's 0.278 and 0.103), and this is never discussed. A classifier with no useful accuracy can trivially produce small group gaps, so the reported fairness gains need to be interpreted in light of the accuracy/fairness trade-off. The paper should either explain why the Random baseline's near-zero gaps do not undermine the interpretation of the IDI gains, or include a control that holds accuracy constant when comparing fairness.
  4. [Table 1 and Sec. 5.2] The headline unsupervised gains come from averaging over four debiasing methods, but the individual unsupervised IDI rows do not consistently beat their supervised counterparts. For example, DS-IDI (unsupervised) has TPRgap 0.201 versus 0.094 for the supervised DS row, and GADRO-IDI (unsupervised) has TPRgap 0.223 versus 0.185 for supervised GADRO. The averaged 'Overall' row masks that every unsupervised IDI method is strictly worse on TPRgap than its supervised counterpart. Please show per-method results and discuss this variability, rather than relying only on the average gain.
  5. [Abstract and Sec. 7] The abstract claims 'less than a 2% decrease in SER accuracy' for pseudo-labeling, while the conclusion and Table 1 report a 4.38% decrease in macro-F1 for the same method. The 2% figure appears to refer to Hamming accuracy, which drops by 1.37%. This selective metric framing should be corrected: the paper should report both accuracy metrics consistently in the abstract and conclusion, or clearly label which metric is being used.
minor comments (6)
  1. [Sec. 3.1 and Fig. 4 caption] The references to 'Appendix Table ??' are unresolved; they should point to Tables A1 and A2 for the K-means and t-SNE parameters.
  2. [Eq. (1)] Equation (1) appears to have a typo: the reweighting term should presumably be n_c / n_{c,g}, but the displayed expression 'nc nc,g' is missing the division symbol.
  3. [Table 1 caption] The symbols in the Table 1 caption are garbled: 'G #represents the pseudo-labeling approach, and# refers to bias unsupervised methods' is unreadable. Please define the symbols clearly, e.g., checkmark for true labels, 'G#' for pseudo-labeling, and '#' for unsupervised clustering.
  4. [Table 2 caption] The caption says 'G denotes the number of subgroups within the demographic information,' but the table rows contain only the number of clusters K; G=3 appears in the column headers. Please clarify the notation so the reader can map each column to the corresponding protected attribute.
  5. [Figs. 2, 3, 4] Several figure captions contain corrupted text fragments (e.g., sequences beginning with '/uni0000...') that should be removed; the captions also do not explain what the bars and colors represent in sufficient detail.
  6. [Sec. 4.3] The definition of TPRgap is described verbally but not written as an equation; a formal definition would help readers reproduce the metric, especially since the paper uses the same metric for different group definitions.

Circularity Check

0 steps flagged · score 0.0 of 10

No demonstrated circularity: the IDI fairness results are empirical comparisons against external baselines and true-demographic test evaluation; self-citations are not load-bearing.

full rationale

The paper is an empirical study, not a derivation chain in which an output is defined as an input. The IDI module infers group labels for training-time debiasing: pseudo-gender labels from a pre-trained 94.4%-accurate detector, or k-means clusters of ECAPA-TDNN embeddings. The central fairness claims are measured by TPRgap and DPgap. Section 4.3 defines these metrics over 'protected attributes (e.g., gender, race)', and Table 2 is explicitly organized by true race/age subgroups ('Race (G=3) Age (G=3)') using CREMA-D demographic metadata, indicating that evaluation groups are the true protected attributes rather than the inferred training groups. The experiments compare against external baselines (LfF, DisEnt, PGDRO) and against supervised debiasing with ground-truth gender labels, so the claimed improvements are anchored independently of the paper's own construction. Although the paper self-cites heavily (e.g., Refs. [4, 6, 11, 18, 33, 34, 35]), none of these citations is load-bearing: Ref. [18] is described as related work on unsupervised clustering, not as a theorem or uniqueness result that forces the present outcome. The main genuine weaknesses are a missing explicit statement that fairness for IDI rows is computed on true demographic labels, and the lack in Section 5.3 of a cluster-to-race/age alignment check before transferring gender-derived clusters to race and age. These are evaluation-specification and validity concerns, not demonstrated circularity: the reported gains could be wrong or confounded, but they are not equivalent to their inputs by construction. The Random baseline's near-zero gaps are likewise a benchmarking-interpretation issue, not a circularity.

Assumptions & free parameters 4 free parameters · 6 assumptions · 1 invented entities

The headline results rest on hand-chosen settings: the cluster count k=16 (selected from {2,4,8,16,32} after inspecting test results), the GADRO regularization constant lambda=4, and the simulated 1:20 gender imbalance ratio. No continuous parameter is fitted to the fairness data in the mathematical sense, but the reported gains are the best of several configurations. The domain assumptions carry the interpretation: CREMA-D metadata must be accurate, k-means clusters must correspond to demographic structure, the pretrained gender detector must be a reliable proxy, the imbalance simulation must reflect deployment, and TPRgap and DPgap must be meaningful even when accuracy drops.

free parameters (4)
  • number of clusters k (unsupervised IDI) = 16 (searched: 2, 4, 8, 16, 32)
    Headline unsupervised gains (TPRgap +20.05%, DPgap +4.61%) are reported at k=16, chosen after inspecting test-set fairness and accuracy across five cluster counts (Sec 5.2, Table 1).
  • GADRO regularization constant lambda_GD = 4
    Set by hand in Sec 3.2 for the GADRO loss; the identical GDRO-IDI and GADRO-IDI pseudo-label rows in Table 1 suggest the term had no measurable effect.
  • simulated gender imbalance ratio = 1:20
    Ad-hoc simulation of biased training data to follow natural demographic trends (Sec 4.2, Table A3); the fairness gains are measured relative to this constructed imbalance.
  • label binarization threshold = 1/6 (1/|Y|)
    Taken from EMO-SUPERB [12] rather than fitted here, but it defines the multi-label targets and therefore the TPRgap and DPgap values (Sec 4.3).
assumptions (6)
  • domain assumption CREMA-D demographic metadata (gender, age, race) are accurate enough to define evaluation subgroups
    All fairness metrics and the 1:20 imbalance simulation rely on these metadata (Sec 4.2, Tables A3 and A4).
  • domain assumption k-means clusters of ECAPA-TDNN embeddings correspond to demographic group structure
    The unsupervised IDI method's group labels are these clusters; evidence is a t-SNE plot for gender only (Fig 4), with no correspondence analysis for race or age before claiming race and age mitigation (Secs 5.2 and 5.3).
  • domain assumption Pseudo-labels from the wav2vec2 age-gender model are reliable proxies for true gender (94.4% accuracy on CREMA-D)
    The pseudo-labeling method and its claimed 28.78% TPRgap gain rest on this accuracy figure (Sec 3.1).
  • domain assumption The EMO-SUPERB split and the 1:20 imbalance simulation reflect realistic deployment conditions
    Claims of real-world relevance rely on this simulation (Sec 4.2, citing Ferrer et al.).
  • domain assumption TPRgap and DPgap as defined in Eqs (4) and (5) are meaningful fairness measures that should be minimized even at the cost of accuracy
    The paper optimizes these and reports gains without discussing that the Random baseline achieves the best values in Table 1.
  • standard math Standard ML machinery (k-means, t-SNE, GDRO, adversarial and contrastive baselines) works as specified
    No modifications to the underlying algorithms; they are taken from the cited work.
invented entities (1)
  • Implicit Demography Inference (IDI) module
    purpose: Infers demographic group proxies from speech without explicit labels, via pseudo-labeling or k-means clustering, to feed group-based debiasing
    A methodological construct, not a physical entity. Its only evidence is the paper's own experiments on CREMA-D; no external falsifiable handle (such as a benchmark prediction) is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mitigating Subgroup Disparities in Multi-Label Speech Emotion Recognition: A Pseudo-Labeling and Unsupervised Learning Approach." pith.science (2026). https://pith.science/paper/YHWKXVB4

@misc{pith2026250514449,
  author       = {Pith},
  title        = {Pith review of: Mitigating Subgroup Disparities in Multi-Label Speech Emotion Recognition: A Pseudo-Labeling and Unsupervised Learning Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YHWKXVB4}},
  note         = {Machine review of arXiv:2505.14449}
}
read the original abstract

While subgroup disparities and performance bias are increasingly studied in computational research, fairness in categorical Speech Emotion Recognition (SER) remains underexplored. Existing methods often rely on explicit demographic labels, which are difficult to obtain due to privacy concerns. To address this limitation, we introduce an Implicit Demography Inference (IDI) module that leverages pseudo-labeling from a pre-trained model and unsupervised learning using k-means clustering to mitigate bias in SER. Our experiments show that pseudo-labeling IDI reduces subgroup disparities, improving fairness metrics by over 28% with less than a 2% decrease in SER accuracy. Also, the unsupervised IDI yields more than a 4.6% improvement in fairness metrics with a drop of less than 3.6% in SER performance. Further analyses reveal that the unsupervised IDI consistently mitigates race and age disparities, demonstrating its potential when explicit demographic information is unavailable.

Figures

Figures reproduced from arXiv: 2505.14449 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Illustration of the averaged results across four models trained with the proposed unsupervised method, using different cluster sizes, in terms of Accuracy and TPR. Male Female [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Illustration of samples in Fold 5 using t-SNE (de￾tails about parameters are in Appendix Table ??), with ECAPA￾TDNN embeddings and a cluster size of 16. Circles and tri￾angles represent male and female speakers, respectively, while different colors indicate different clusters. 5.3. Potential of Unsupervised Clustering We assess the applicability of unsupervised clustering to ad￾ditional demographic attributes, as it… view at source ↗
Figures from the paper (1 more)
Figure 2
Figure 2. Figure 2: The left subfigure presents the TPRgap, while the right one depicts the Hamming accuracy. The bars illustrate the per￾formance of the proposed unsupervised debiasing method, eval￾uated across four different techniques with varying cluster sizes. 2 4 8 16 32 0.700 0.750…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 42 canonical work pages

  1. [1]

    Introduction Speech Emotion Recognition (SER) is essential in hu- man–computer interaction, enhancing user experiences across various applications [1]. As SER becomes increasingly inte- gral to applications ranging from virtual assistants to therapeu- tic tools, ensuring that these systems operate fairly and without bias is more critical than ever [2, 3]....

    work page

  2. [2]

    Debias by Supervised Learning Supervised learning approaches to mitigating gender bias in SER typically rely on explicitly annotated gender labels dur- ing training

    Related Works 2.1. Debias by Supervised Learning Supervised learning approaches to mitigating gender bias in SER typically rely on explicitly annotated gender labels dur- ing training. Previous studies have implemented fairness con- straints or modified training strategies to reduce performance disparities between gender subgroups. Chien et al. [7] employ...

    arXiv 2025

  3. [3]

    Central to our approach is the Implicit Demography Inference (IDI) Module, which infers group labels from speech data by combining pseudo-labeling and unsuper- vised clustering

    Methodology We propose a novel framework to mitigate subgroup dispari- ties in gender in multi-label SER without relying on explicit de- mographic annotations. Central to our approach is the Implicit Demography Inference (IDI) Module, which infers group labels from speech data by combining pseudo-labeling and unsuper- vised clustering. These inferred grou...

    work page

  4. [4]

    Experiments Setup We select the SSL-based SER model with WavLM base plus 2

    work page

  5. [5]

    Young” (20–35 years),“Middle

    Results and Analyses Table 1 presents the averaged results on performance and gen- der bias across five folds. We evaluate the effectiveness of our proposed methods and provide detailed analyses in the follow- ing sections. 5.1. Performance and Effects of Pseudo-labeling Our results show that pseudo-labeling reduces subgroup dis- parities in multi-label S...

    work page

  6. [6]

    First, our experiments are based on the CREMA-D dataset, which comprises acted emotional expressions; thus, the generalizability to naturalistic settings remains to be vali- dated

    Limitations While our approach demonstrates robust improvements in mit- igating subgroup disparities, several limitations warrant dis- cussion. First, our experiments are based on the CREMA-D dataset, which comprises acted emotional expressions; thus, the generalizability to naturalistic settings remains to be vali- dated. Second, our methods depend heavi...

    work page

  7. [7]

    Conclusion and Future Work This work empirically examines various debiasing methods in supervised, pseudo-label, and unsupervised learning to mitigate subgroup disparities without explicit demographic information. We leverage pseudo-labels generated by a pre-trained model and introduce an unsupervised clustering approach that inte- grates with existing de...

    work page

  8. [8]

    Acknowledgement We thank Taiwan’s National Center for High-Performance Computing (NCHC) at the National Applied Research Labora- tories (NARLabs) for providing the computational and storage resources

    work page

Show all 47 references
  1. [9]

    Prioritizing data acquisition for end-to-end speech model improvement,

    A. Koudounas, E. Pastor, G. Attanasio, L. de Alfaro, and E. Bar- alis, “Prioritizing data acquisition for end-to-end speech model improvement,” in ICASSP 2024 - 2024 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP) , 2024, pp. 7000–7004

    2024

  2. [10]

    Mitigating bias against non-native accents,

    Y . Zhang, Y . Zhang, B. Halpern, T. Patel, and O. Scharenborg, “Mitigating bias against non-native accents,” inInterspeech 2022, 2022, pp. 3168–3172

    2022

  3. [11]

    Speech emotion recognition for hu- man–computer interaction,

    D. Thiripurasundari, K. Bhangale, V . Aashritha, S. Mondreti, and M. Kothandaraman, “Speech emotion recognition for hu- man–computer interaction,” Int. J. Speech Technol., vol. 27, no. 3, p. 817–830, Aug. 2024

    2024

  4. [12]

    Discrimination, Bias, Fairness, and Trustworthy AI,

    D. Varona and J. L. Su ´arez, “Discrimination, Bias, Fairness, and Trustworthy AI,” Applied Sciences, vol. 12, no. 12, 2022

    2022

  5. [13]

    Fairness Without Demographic Data: A Survey of Approaches,

    C. Ashurst and A. Weller, “Fairness Without Demographic Data: A Survey of Approaches,” inProceedings of the 3rd ACM Confer- ence on Equity and Access in Algorithms, Mechanisms, and Opti- mization, ser. EAAMO ’23. New York, NY , USA: Association for Computing Machinery, 2023

    2023

  6. [14]

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

    Y .-C. Lin, H. Wu, H.-C. Chou, C.-C. Lee, and H. yi Lee, “Emo- bias: A Large Scale Evaluation of Social Bias on Speech Emotion Recognition,” in Interspeech 2024, 2024, pp. 4633–4637

    2024

  7. [15]

    Gen- der De-Biasing in Speech Emotion Recognition,

    C. Gorrostieta, R. Lotfian, K. Taylor, R. Brutti, and J. Kane, “Gen- der De-Biasing in Speech Emotion Recognition,” in Interspeech 2019, 2019, pp. 2823–2827

    2019

  8. [16]

    On the social bias of speech self-supervised models,

    Y .-C. Lin, T.-Q. Lin, H.-C. Lin, A. T. Liu, and H. yi Lee, “On the social bias of speech self-supervised models,” in Interspeech 2024, 2024, pp. 4638–4642

    2024

  9. [17]

    Achieving Fair Speech Emotion Recognition via Perceptual Fairness,

    W.-S. Chien and C.-C. Lee, “Achieving Fair Speech Emotion Recognition via Perceptual Fairness,” in ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  10. [18]

    Balancing Speaker- Rater Fairness for Gender-Neutral Speech Emotion Recognition,

    W.-S. Chien, S. G. Upadhyay, and C.-C. Lee, “Balancing Speaker- Rater Fairness for Gender-Neutral Speech Emotion Recognition,” in ICASSP 2024 - 2024 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP) , 2024, pp. 11 861– 11 865

    2024

  11. [19]

    Good practices for evaluation of machine learning systems,

    L. Ferrer, O. Scharenborg, and T. B ¨ackstr¨om, “Good practices for evaluation of machine learning systems,” 2024

    2024

  12. [20]

    Speech-based Age and Gender Prediction with Transformers,

    F. Burkhardt, J. Wagner, H. Wierstorf, F. Eyben, and B. Schuller, “Speech-based Age and Gender Prediction with Transformers,” in Speech Communication; 15th ITG Conference, 2023, pp. 46–50

    2023

  13. [21]

    The cluster assignments derived from this process are treated as group labels, reflecting latent structures within the data that may correspond to demographic differ- ences

    (details about parameters are in Appendix Table ??) to these embeddings. The cluster assignments derived from this process are treated as group labels, reflecting latent structures within the data that may correspond to demographic differ- ences. We set the cluster sizes to 2,...

  14. [22]

    Embracing Ambiguity And Subjectivity Using The All-Inclusive Aggregation Rule For Eval- uating Multi-Label Speech Emotion Recognition Systems,

    H.-C. Chou, H. Wu, L. Goncalves, S.-G. Leem, A. Salman, C. Busso, H.-Y . Lee, and C.-C. Lee, “Embracing Ambiguity And Subjectivity Using The All-Inclusive Aggregation Rule For Eval- uating Multi-Label Speech Emotion Recognition Systems,” in 2024 IEEE Spoken Language Technology...

    2024

  15. [23]

    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 Technology Workshop (SLT), 2024, pp. 510–517

    2024

  16. [24]

    Semantic Space Theory: A Com- putational Approach to Emotion,

    A. S. Cowen and D. Keltner, “Semantic Space Theory: A Com- putational Approach to Emotion,” Trends in Cognitive Sciences , vol. 25, no. 2, pp. 124–136, 2021

    2021

  17. [25]

    We use the class- balanced cross-entropy loss [26] as base SER loss LSER

    feature extractor and two linear layers as our primary back- bone for the following experiments, because it ranks second on 2https://huggingface.co/s3prl/converted ckpts/resolve/main/wavlm base plus.pt the SER leaderboard [12] while maintaining a lower parameter count compared...

  18. [26]

    ECAPA- TDNN: Emphasized Channel Attention, Propagation and Ag- gregation in TDNN Based Speaker Verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “ECAPA- TDNN: Emphasized Channel Attention, Propagation and Ag- gregation in TDNN Based Speaker Verification,” in Interspeech 2020, 2020, pp. 3830–3834

    2020

  19. [27]

    CREMA-D: Crowd-Sourced Emotional Multi- modal Actors Dataset,

    H. Cao, D. G. Cooper, M. K. Keutmann, R. C. Gur, A. Nenkova, and R. Verma, “CREMA-D: Crowd-Sourced Emotional Multi- modal Actors Dataset,” IEEE Transactions on Affective Comput- ing, vol. 5, no. 4, pp. 377–390, 2014

    2014

  20. [28]

    Is It Still Fair? Investigating Gender Fairness in Cross-Corpus Speech Emotion Recognition,

    S. G. Upadhyay, W.-S. Chien, and C.-C. Lee, “Is It Still Fair? Investigating Gender Fairness in Cross-Corpus Speech Emotion Recognition,” in ICASSP 2025 - 2025 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP) , 2025

    2025

  21. [29]

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

    S. Park, B. Jeon, S. Lee, and J. Yoon, “Multi-Label Emotion Recognition of Korean Speech Data Using Deep Fusion Models,” Applied Sciences, vol. 14, no. 17, 2024

    2024

  22. [30]

    Stimulus Modality Matters: Impact of Perceptual Evaluations from Different Modalities on Speech Emotion Recognition System Performance,

    H.-C. Chou, H. Wu, and C.-C. Lee, “Stimulus Modality Matters: Impact of Perceptual Evaluations from Different Modalities on Speech Emotion Recognition System Performance,” in ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025

    2025

  23. [31]

    Least squares quantization in PCM,

    S. Lloyd, “Least squares quantization in PCM,” IEEE Transac- tions on Information Theory, vol. 28, no. 2, pp. 129–137, 1982

    1982

  24. [32]

    Data preprocessing techniques for classification without discrimination,

    F. Kamiran and T. Calders, “Data preprocessing techniques for classification without discrimination,” Knowledge and Informa- tion Systems, vol. 33, no. 1, pp. 1–33, 2012

    2012

  25. [33]

    Robust Solutions of Optimization Problems Affected by Uncertain Probabilities,

    A. Ben-Tal, D. den Hertog, A. De Waegenaere, B. Melenberg, and G. Rennen, “Robust Solutions of Optimization Problems Affected by Uncertain Probabilities,” Management Science, vol. 59, no. 2, pp. 341–357, 2013

    2013

  26. [34]

    Dis- tributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization

    S. Sagawa*, P. W. Koh*, T. B. Hashimoto, and P. Liang, “Dis- tributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization.” in International Conference on Learning Representations, 2020

    2020

  27. [35]

    WavLM: Large-Scale Self-Supervised Pre- Training for Full Stack Speech Processing,

    S. Chen et al. , “WavLM: Large-Scale Self-Supervised Pre- Training for Full Stack Speech Processing,” IEEE Journal of Se- lected Topics in Signal Processing, 2022

    2022

  28. [36]

    Class- Balanced Loss Based on Effective Number of Samples,

    Y . Cui, M. Jia, T.-Y . Lin, Y . Song, and S. Belongie, “Class- Balanced Loss Based on Effective Number of Samples,” in Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  29. [37]

    Learning from fail- ure: training debiased classifier from biased classifier,

    J. Nam, H. Cha, S. Ahn, J. Lee, and J. Shin, “Learning from fail- ure: training debiased classifier from biased classifier,” in Pro- ceedings of the 34th International Conference on Neural Infor- mation Processing Systems, Red Hook, NY , USA, 2020

    2020

  30. [38]

    Learning Debiased Representation via Disentangled Feature Augmentation,

    J. Lee, E. Kim, J. Lee, J. Lee, and J. Choo, “Learning Debiased Representation via Disentangled Feature Augmentation,” in Ad- vances in Neural Information Processing Systems, 2021

    2021

  31. [39]

    Distributionally Robust Optimization with Probabilistic Group,

    S. S. Ghosal and Y . Li, “Distributionally Robust Optimization with Probabilistic Group,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 10, pp. 11 809–11 817, Jun. 2023

    2023

  32. [40]

    Diverse adversaries for miti- gating bias in training,

    X. Han, T. Baldwin, and T. Cohn, “Diverse adversaries for miti- gating bias in training,” in Proceedings of the 16th Conference of the European Chapter of the Association for Computational Lin- guistics: Main Volume, 2021

    2021

  33. [41]

    Bias in bios: A case study of semantic representation bias in a high- stakes setting,

    M. De-Arteaga, A. Romanov, H. Wallach, J. Chayes, C. Borgs, A. Chouldechova, S. Geyik, K. Kenthapadi, and A. T. Kalai, “Bias in bios: A case study of semantic representation bias in a high- stakes setting,” inProceedings of the Conference on Fairness, Ac- countability, and Tra...

    2019

  34. [42]

    Visualizing Data using t-SNE,

    L. van der Maaten and G. Hinton, “Visualizing Data using t-SNE,” Journal of Machine Learning Research, vol. 9, no. 86, pp. 2579– 2605, 2008

    2008

  35. [43]

    Jointly Learning From Unimodal and Multimodal- Rated Labels in Audio-Visual Emotion Recognition,

    L. Goncalves, H.-C. Chou, A. N. Salman, C.-C. Lee, and C. Busso, “Jointly Learning From Unimodal and Multimodal- Rated Labels in Audio-Visual Emotion Recognition,”IEEE Open Journal of Signal Processing, vol. 6, pp. 165–174, 2025

    2025

  36. [44]

    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 Workshop (SLT), 2024

    2024

  37. [45]

    Improving speech emotion recognition in under-resourced languages via speech-to-speech translation with bootstrapping data selection,

    H.-C. Lin, Y .-C. Lin, H.-C. Chou, and H. yi Lee, “Improving speech emotion recognition in under-resourced languages via speech-to-speech translation with bootstrapping data selection,”

  38. [47]

    Don’t speak too fast: The impact of data bias on self-supervised speech models,

    Y . Meng, Y .-H. Chou, A. T. Liu, and H.-y. Lee, “Don’t speak too fast: The impact of data bias on self-supervised speech models,” in ICASSP 2022 - 2022 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP), 2022. Supplementary Material The Supple...

    2022

  39. [2025]

    Available: https://arxiv.org/abs/2409.10985

    [Online]. Available: https://arxiv.org/abs/2409.10985

    arXiv

Pith tools

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