Pith. sign in

REVIEW 4 major objections 6 minor 35 references

Sounding Like a Winner? Prosodic Differences in Post-Match Interviews

T0 review · 4 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 Post-match tennis interview audio carries enough acoustic-prosodic signal to distinguish winners from losers, with self-supervised speech models outperforming handcrafted features and pitch variability emerging as a key marker of victory.

desk verdict New task, flawed evidence: the win/lose audio classification is contaminated by lexical content and the prosodic analysis rests on a chance-level classifier. read the letter →

arxiv 2506.02283 v1 pith:WRG4OQCL submitted 2025-06-02 cs.CL eess.AS

Sofoklis Kakouros , Haoyu Chen This is my paper
classification cs.CLeess.AS
keywords speechemotionrecognitionprosodyself-supervisedlearningWav2Vec2.0HuBERTpost-matchinterviewspitchvariabilitytennis
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 the sound of a tennis player's voice in a post-match interview betrays whether they won or lost. The authors extract both handcrafted prosodic features and self-supervised speech representations (HuBERT, Wav2Vec 2.0, exHuBERT) from the iMiGUE dataset of press-conference recordings, then train classifiers to predict match outcome. They report that the self-supervised representations separate winners from losers far better than traditional acoustic features, with HuBERT reaching 65.9% accuracy, and that prosodic analysis points to pitch variability—especially the variability of rising pitch contours—and overall loudness as the cues most associated with victory. If that holds, ordinary post-match audio becomes a measurable signal of competitive emotional state without any explicit emotion labels.

What carries the argument

The machinery is a pairing of self-supervised speech representations with a supervised classifier and a Shapley feature-attribution analysis. The SSL representations (HuBERT, Wav2Vec 2.0, exHuBERT) are mean-pooled, utterance-level embeddings from the final transformer layer, which the paper treats as rich encoders of emotional-prosodic state. The prosodic analysis then uses SHAP values—a feature-attribution method that quantifies each input feature's marginal contribution to a model's prediction—applied to a classifier trained on the 88-dimensional eGeMAPS feature set, to rank which acoustic features (pitch slope variability, loudness, spectral tilt) most influence the win/lose decision.

What would settle it

Compare the raw distributions of $F_0$ slope variability, average rising-slope steepness, and equivalent sound level between winner and loser interviews on a held-out set of speakers using a standard statistical test such as the Mann–Whitney U test; if the distributions do not differ significantly, the prosodic-differentiation claim is falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that winning and losing leave measurable traces in the prosody of post-match interview speech, and that these traces are more accessible to self-supervised speech models than to standard acoustic feature sets. The authors claim that winners speak with more dynamic pitch movements, greater pitch-slope variability, and higher overall intensity, while losers sound more monotonous; and that classifiers built on HuBERT and Wav2Vec 2.0 embeddings distinguish the two conditions with 65.9% and 63.9% accuracy, respectively, whereas a classifier on eGeMAPS prosodic features reaches only 44.4%. The paper presents this as evidence that SSL representations capture subtle speech patterns linked to emotional states in a natural, unlabeled setting.

Load-bearing premise

The prosodic-difference claims rest on feature-importance scores from a classifier that performed at 44.4% accuracy on an imbalanced test set, so those feature rankings may reflect noise rather than real acoustic differences between winners and losers.

Editorial extensions

If this is right

  • Audio-only analysis of post-match interviews can predict win/lose outcome with about two-thirds accuracy using SSL embeddings.
  • Handcrafted prosodic features used alone are not sufficient for this classification, suggesting the need for richer representations.
  • Winners and losers differ systematically in pitch dynamics and loudness, with winners showing more expressive, varied pitch and higher intensity.
  • The win/lose signal is accessible without transcripts or semantic content, since the models use only acoustic information.

Reading between the lines

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

  • The paper's SHAP analysis ranks features from a classifier trained on handcrafted features; an analogous attribution applied directly to the SSL representations could reveal which learned dimensions carry the outcome signal.
  • Because all recordings are tennis press conferences, extending the method to other competitive speech—other sports, esports, or political debates—would test whether the pitch-dynamics signature of winning generalizes.
  • Controlling for speaker identity, interview duration, and question content in a matched-pair design would separate the emotional-prosodic signal from situational artifacts.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper analyzes audio from post-match tennis press conferences in the iMiGUE dataset, asking whether winning and losing athletes can be distinguished from speech alone. It extracts handcrafted eGeMAPS/openSMILE prosodic features and final-layer mean-pooled embeddings from Wav2Vec 2.0, HuBERT, and exHuBERT, then trains a small MLP for binary win/lose classification. The reported accuracies are 65.9% (HuBERT), 63.9% (Wav2Vec 2.0), 60.7% (exHuBERT), and 44.4% (openSMILE). SHAP values from the openSMILE classifier are used to claim that pitch variability and rising-slope features distinguish winners from losers. The abstract concludes that SSL representations capture emotional speech patterns and that prosodic cues such as pitch variability remain strong indicators of victory.

Significance. If the claims were established, the paper would contribute a novel naturalistic benchmark for emotion inference and an interesting comparison of SSL versus handcrafted prosodic features. The authors deserve credit for being transparent about two facts that are in tension with their own conclusions: they report that transcript-only classification with LLaMA exceeds 90% accuracy (§2.1), and their openSMILE classifier only reaches 44.4% accuracy (Table 1). That transparency is a strength, but it makes the paper's central claims untenable as written. Because the SSL accuracies are below the majority-class baseline and the prosodic interpretation is built on a chance-level classifier, the evidence does not support the abstract's claims. The task and pipeline are worth reporting, but the current analysis needs substantial additional experiments and controls.

major comments (4)
  1. [Table 1 and §3.2] The paper never reports a majority-class baseline, and this is load-bearing because the data are roughly 80% win class (§3.2). A trivial classifier that always predicts 'win' would achieve about 80% accuracy, yet the best SSL model reports only 65.9% accuracy and the openSMILE model 44.4%. The claim in the abstract that 'SSL representations effectively differentiate between winning and losing outcomes' is therefore not supported; the reported models are far below even the constant predictor. The authors should report balanced accuracy, macro-F1, area under the ROC curve, and a statistical comparison against the majority baseline, and should adjust their conclusions accordingly.
  2. [§5.1 and Figure 2] The SHAP analysis is conducted on the openSMILE classifier, which achieves 44.4% accuracy in Table 1—indistinguishable from chance for a two-class problem. SHAP values describe the behavior of that classifier, not the true relationship between acoustic features and match outcome. Consequently, the statements in §5.1 and the conclusions that pitch variability (e.g., F0 stddevRisingSlope), rising pitch trends, and loudness are strong indicators of victory are not supported by any valid evidence. The authors should replace this interpretation with a direct descriptive comparison of feature distributions across win/lose groups, including effect sizes and confidence intervals, before drawing any prosodic conclusions.
  3. [§2.1, §3.1, and §5.2] The central SSL classification result is compromised by an uncontrolled lexical-content confound. The authors note in §2.1 that transcript-only classification with LLaMA-3.1-8B exceeds 90% accuracy, meaning the words spoken largely determine the outcome. Final-layer Wav2Vec 2.0 and HuBERT embeddings are known to encode phonetic and lexical information, and mean-pooled utterance-level embeddings can leak a weak but usable content signal. The observed 60–66% accuracy is exactly the range one would expect from such leakage, so the result does not establish that 'subtle speech patterns linked to emotional states' are captured. To support the audio-only claim, the paper needs a content-masked control condition, for example delexicalized resynthesis (e.g., syllable shuffling or vocoder-based prosody preservation), a scrambled-segment baseline, or a comparison against text-only embeddings with matched content.
  4. [§4.1 and §5.1] The prosodic analysis appears to be performed on the same test split using a model trained with SMOTE and a 70/20/10 split, but no information is given about the number of test recordings, class counts in the test set, or variance across training runs. Given that the test set is likely only about 36 recordings and the minority (lose) class is roughly 20% of the data, the effective number of positive test samples may be very small. Any classification or SHAP result on such a small test set requires confidence intervals or repeated cross-validation; otherwise the reported differences, and especially the chance-level openSMILE performance, cannot be interpreted reliably.
minor comments (6)
  1. [§2.2] The text contains a typo in 'V AD' (should be 'VAD'); please correct the spacing and ensure terminology is consistent throughout.
  2. [Table 1] Please add a majority-class baseline row and the number of test samples to the table; this would make the chance-level performance immediately visible and help readers interpret the F1 and recall values.
  3. [Figure 2] The caption does not say what the plotted SHAP values represent (e.g., mean absolute SHAP over the test set, per-class SHAP, or individual prediction contributions), nor what the color scale indicates. Please clarify the units and the aggregation method.
  4. [§5.1] The qualitative description of F0 stddevRisingSlope is internally hard to follow: the text says winners 'exhibit high and low values' while losers have 'uniformly high values,' yet it concludes that winners are more dynamic. Without numeric summaries or a figure showing actual distributions, this contrast is not verifiable.
  5. [Abstract and §6] The abstract and conclusion repeat the claim that pitch variability is a strong victory indicator even though the openSMILE classifier that motivated the SHAP analysis performs at chance. Please revise these statements to match what the evidence can support.
  6. [§1] The phrase 'an unprecedented natural-scenario task' is overstated; 'a novel task' would suffice. Also, the paper is submitted to a venue with an 'Index Terms' line in the abstract; please confirm that the submission format is appropriate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the win/lose target is defined independently of the speech features, and the cited self-work is background only.

full rationale

The paper's central derivation chain is not circular. The classification target (win vs. lose) comes from the iMiGUE match-outcome annotations and is external to the speech signal; SSL embeddings and openSMILE features are computed from the audio independently of the labels, and no fitted parameter is used to define the outcome. The reported accuracies (65.9% HuBERT, 63.9% Wav2Vec 2.0, etc.) are genuine test-set evaluations, not a fit renamed as a prediction. The SHAP-based prosodic analysis is statistically weak because the openSMILE classifier reaches only 44.4% accuracy, and the unaddressed lexical-content confound undermines the emotional-prosody interpretation, but these are validity and robustness problems, not circular reductions: SHAP values are post-hoc explanations of a trained model, and the model's chance-level performance does not make its feature attributions equivalent by construction to the input labels. The iMiGUE 'positive/negative emotion' labels are indeed defined from match outcomes, so calling the task 'emotion inference' is a framing choice with construct-validity limitations; however, the actual experiments classify win/lose outcomes, so no result is derived from a definition that presupposes the result. The background self-citations (Refs. [2,3,17]) are not load-bearing: they support general statements about SSL for emotion recognition and prosody, and no uniqueness theorem or ansatz is imported from prior work by the same authors. The paper is therefore self-contained for its stated classification experiments; no step reduces to its own input.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several unproved premises: the pretrained SSL models encode the emotional and prosodic information needed for the task; the diarization and alignment tools correctly isolate the athlete's speech; the iMiGUE labels are faithful win or lose outcomes; and SHAP attributions from a chance-level classifier are interpretable. None of these are validated, and the last one is contradicted by the classifier's own performance. The paper introduces no new invented entities, and its only hand-chosen free parameters are the SMOTE balancing procedure, the longest-speaker heuristic, and the mean-pooling strategy.

free parameters (3)
  • SMOTE oversampling target = unspecified (likely 1:1)
    SMOTE is used to balance the 80/20 win-lose classes (Section 3.2), but the oversampling ratio and whether it is applied before or after the split are not reported. If applied before splitting, synthetic samples could leak into the test set.
  • Athlete identification rule = speaker with longest total duration
    The athlete's segments are identified by assuming the longest-talking speaker is the player (Section 2.2). This hand-chosen rule is not validated and could misattribute journalist speech, contaminating all features.
  • SSL embedding pooling strategy = mean over frames and over segments
    Per-segment SSL embeddings are mean-pooled to a single vector per recording (Section 3.1). This discards temporal prosodic dynamics and is an arbitrary modeling choice that the paper does not justify.
assumptions (4)
  • domain assumption Pretrained SSL models (Wav2Vec 2.0, HuBERT, exHuBERT) encode task-relevant prosodic and emotional information.
    The paper relies on these models' final-layer representations without fine-tuning on the target task, assuming the representations carry enough signal for win-lose classification (Section 3.1).
  • domain assumption The speaker diarization and alignment tools (Pyannote, Whisper, MFA) produce sufficiently accurate speaker labels and word boundaries.
    All downstream features are computed on segments selected by these tools; errors in diarization would contaminate the athlete's speech features (Section 2.2).
  • domain assumption The iMiGUE dataset's emotion labels (positive or negative) correspond directly to win or lose match outcomes and are free of annotation errors.
    The paper uses these labels as ground truth for classification and prosodic comparison (Section 2.1).
  • ad hoc to paper SHAP values from a trained classifier reflect true feature-outcome relationships even when the classifier's accuracy is at chance.
    The prosodic analysis (Section 5.1) computes SHAP on the openSMILE classifier, which achieves 44.4% accuracy (Table 1); the paper implicitly assumes the attributions are still meaningful.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sounding Like a Winner? Prosodic Differences in Post-Match Interviews." pith.science (2026). https://pith.science/paper/WRG4OQCL

@misc{pith2026250602283,
  author       = {Pith},
  title        = {Pith review of: Sounding Like a Winner? Prosodic Differences in Post-Match Interviews},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WRG4OQCL}},
  note         = {Machine review of arXiv:2506.02283}
}
read the original abstract

This study examines the prosodic characteristics associated with winning and losing in post-match tennis interviews. Additionally, this research explores the potential to classify match outcomes solely based on post-match interview recordings using prosodic features and self-supervised learning (SSL) representations. By analyzing prosodic elements such as pitch and intensity, alongside SSL models like Wav2Vec 2.0 and HuBERT, the aim is to determine whether an athlete has won or lost their match. Traditional acoustic features and deep speech representations are extracted from the data, and machine learning classifiers are employed to distinguish between winning and losing players. Results indicate that SSL representations effectively differentiate between winning and losing outcomes, capturing subtle speech patterns linked to emotional states. At the same time, prosodic cues -- such as pitch variability -- remain strong indicators of victory.

Figures

Figures reproduced from arXiv: 2506.02283 by the authors.

Figure 1
Figure 1. Overview of the processing pipeline for audio extraction, feature extraction, and classification ditional features were extracted using state-of-the-art self￾supervised learning (SSL) representations. SSL has driven advancements across various speech technology domains, in￾cluding prosody modeling. Compared to standard acoustic features like MFCCs, SSL representations have significantly higher dimensionality, provid… view at source ↗
Figure 2
Figure 2. SHAP OpenSMILE feature importance values. 5.2. Win/Lose classification [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 26 canonical work pages

  1. [1]

    Introduction Although significant progress has been made in speech emotion recognition (SER) [1, 2, 3] and understanding [4], most research has focused on artificially designed tasks and subjective emo- tion annotations [5]. These tasks typically rely on predefined emotion categories (such as happiness, sadness, anger, etc.), and the emotion annotations a...

    arXiv 2025

  2. [2]

    Materials and data processing 2.1. iMiGUE dataset The iMiGUE (Identity-free Micro-Gesture Understanding and Emotion Analysis) dataset is a specialized video dataset de- signed for studying micro-gestures and their relationship to emotional states while ensuring privacy [24]. Unlike conven- tional emotion recognition datasets that focus on facial expres- s...

    work page

  3. [3]

    Feature extraction A standard set of acoustic features was extracted using eGeMAPS [18], which includes 88 distinct features and func- tionals for each interviewee segment

    Method 3.1. Feature extraction A standard set of acoustic features was extracted using eGeMAPS [18], which includes 88 distinct features and func- tionals for each interviewee segment. These features encompass the means and standard deviations of fundamental frequency (F0), loudness, spectral tilt, and mel-frequency cepstral coef- ficients (MFCCs). For ea...

    work page

  4. [4]

    Experiments The dataset was divided into training (70%), validation (20%), and test (10%) sets, ensuring that each subset contained distinct speakers with no overlap to prevent data leakage and improve generalization. 4.1. Prosodic analysis For the prosodic analysis, a neural network was trained for 50 epochs using 88-dimensional openSMILE features, which...

    work page

  5. [5]

    Results and analysis 5.1. Prosodic analysis Our findings show (see Figure 2) that winners tend to have more dynamic pitch movements, varied spectral features, and higher intensity, whereas players who lost their match exhibit more monotonous speech patterns with less variation in frequency and intensity. This pattern is particularly evident in features re...

    work page

  6. [6]

    The acoustic-prosodic analysis reveals that winners exhibit more dynamic pitch movements, varied spectral fea- tures, and greater intensity

    Conclusions This paper presents a prosodic analysis and win/lose classifi- cation performance on a novel dataset of post-match interview recordings. The acoustic-prosodic analysis reveals that winners exhibit more dynamic pitch movements, varied spectral fea- tures, and greater intensity. In contrast, players who lost their match tend to have flatter, mor...

    work page

  7. [7]

    The au- thors wish to acknowledge CSC – IT Center for Science, Fin- land, for providing the computational resources

    Acknowledgements This work was supported by the University of Oulu and the Re- search Council of Finland, PROFI7 352788 project. The au- thors wish to acknowledge CSC – IT Center for Science, Fin- land, for providing the computational resources

    work page

  8. [8]

    Multitask transformer for cross-corpus speech emotion recognition,

    C.-S. Ahn, R. Rana, C. Busso, and J. C. Rajapakse, “Multitask transformer for cross-corpus speech emotion recognition,”IEEE Transactions on Affective Computing, 2025

    work page 2025

Show all 35 references
  1. [9]

    Speech- based emotion recognition with self-supervised models using attentive channel-wise correlations and label smoothing,

    S. Kakouros, T. Stafylakis, L. Mo ˇsner, and L. Burget, “Speech- based emotion recognition with self-supervised models using attentive channel-wise correlations and label smoothing,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (...

    2023

  2. [10]

    Extracting speaker and emotion information from self-supervised speech models via channel-wise correla- tions,

    T. Stafylakis, L. Mo ˇsner, S. Kakouros, O. Plchot, L. Burget, and J. ´Cernock`y, “Extracting speaker and emotion information from self-supervised speech models via channel-wise correla- tions,” in2022 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2023, pp. 1136–1143

    2023

  3. [11]

    Analyzing the influence of different speech data corpora and speech features on speech emotion recog- nition: A review,

    T. Rathi and M. Tripathy, “Analyzing the influence of different speech data corpora and speech features on speech emotion recog- nition: A review,”Speech Communication, p. 103102, 2024

    2024

  4. [12]

    IEMOCAP: Interactive emotional dyadic motion capture database,

    C. Busso, M. Bulutet al., “IEMOCAP: Interactive emotional dyadic motion capture database,”Language resources and eval- uation, vol. 42, no. 4, pp. 335–359, 2008

    2008

  5. [13]

    The ryerson audio-visual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english,

    S. R. Livingstone and F. A. Russo, “The ryerson audio-visual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english,”PloS one, vol. 13, no. 5, p. e0196391, 2018

    2018

  6. [14]

    Emotional speech recogni- tion: Resources, features, and methods,

    D. Ververidis and C. Kotropoulos, “Emotional speech recogni- tion: Resources, features, and methods,”Speech communication, vol. 48, no. 9, pp. 1162–1181, 2006

    2006

  7. [15]

    Emotion recognition from speech using wav2vec 2.0 embeddings,

    L. Pepino, P. Riera, and L. Ferrer, “Emotion recognition from speech using wav2vec 2.0 embeddings,”arXiv preprint arXiv:2104.03502, 2021

    2021 arXiv

  8. [16]

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

    W.-N. Hsu, B. Bolteet al., “HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 29, pp. 3451–3460, 2021

    2021

  9. [17]

    wav2vec 2.0: A framework for self- supervised learning of speech representations,

    A. Baevski, Y . Zhouet al., “wav2vec 2.0: A framework for self- supervised learning of speech representations,”Advances in Neu- ral Information Processing Systems, vol. 33, pp. 12 449–12 460, 2020

    2020

  10. [18]

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

    S. Chen, C. Wanget al., “WavLM: Large-scale self-supervised pre-training for full stack speech processing,”arXiv preprint arXiv:2110.13900, 2021

    2021

  11. [19]

    SUPERB: Speech processing uni- versal performance benchmark,

    S.-w. Yang, P.-H. Chiet al., “SUPERB: Speech processing uni- versal performance benchmark,” inProceedings of Interspeech, 2021

    2021

  12. [20]

    Emotion identification from raw speech signals using dnns

    M. Sarma, P. Ghahremaniet al., “Emotion identification from raw speech signals using dnns.” inInterspeech, 2018, pp. 3097–3101

    2018

  13. [21]

    Leveraging inter-rater agreement for audio-visual emotion recognition,

    Y . Kim and E. M. Provost, “Leveraging inter-rater agreement for audio-visual emotion recognition,” in2015 International Confer- ence on Affective Computing and Intelligent Interaction (ACII). IEEE, 2015, pp. 553–559

    2015

  14. [22]

    Interpreting ambiguous emo- tional expressions,

    E. Mower, A. Metallinouet al., “Interpreting ambiguous emo- tional expressions,” in2009 3rd International Conference on Af- fective Computing and Intelligent Interaction and Workshops. IEEE, 2009, pp. 1–8

    2009

  15. [23]

    A speech emotion recognition framework for better discrimination of confusions,

    J. Liu and H. Wang, “A speech emotion recognition framework for better discrimination of confusions,” inInterspeech, 2021, pp. 4483–4487

    2021

  16. [24]

    The effect of noise on emotion perception in an unknown language,

    O. Scharenborg, S. Kakouros, and J. Koemans, “The effect of noise on emotion perception in an unknown language,” inInter- national Conference on Speech Prosody. International Speech Communication Association (ISCA), 2018, pp. 364–368

    2018

  17. [25]

    The geneva minimalistic acoustic parameter set (gemaps) for voice research and affective computing,

    F. Eyben, K. R. Scherer, B. W. Schuller, J. Sundberg, E. Andr ´e, C. Busso, L. Y . Devillers, J. Epps, P. Laukka, S. S. Narayanan et al., “The geneva minimalistic acoustic parameter set (gemaps) for voice research and affective computing,”IEEE transactions on affective computi...

    2015

  18. [26]

    Recognising realistic emotions and affect in speech: State of the art and lessons learnt from the first challenge,

    B. Schuller, A. Batliner, S. Steidl, and D. Seppi, “Recognising realistic emotions and affect in speech: State of the art and lessons learnt from the first challenge,”Speech communication, vol. 53, no. 9-10, pp. 1062–1087, 2011

    2011

  19. [27]

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

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

    2003

  20. [28]

    Human vo- cal attractiveness as signaled by body size projection,

    Y . Xu, A. Lee, W.-L. Wu, X. Liu, and P. Birkholz, “Human vo- cal attractiveness as signaled by body size projection,”PloS one, vol. 8, no. 4, p. e62397, 2013

    2013

  21. [29]

    Combining acous- tic and language information for emotion recognition

    C. M. Lee, S. S. Narayanan, and R. Pieraccini, “Combining acous- tic and language information for emotion recognition.” inINTER- SPEECH. Citeseer, 2002, pp. 873–876

    2002

  22. [30]

    Speech technology for healthcare: Opportunities, challenges, and state of the art,

    S. Latif, J. Qadir, A. Qayyum, M. Usama, and S. Younis, “Speech technology for healthcare: Opportunities, challenges, and state of the art,”IEEE Reviews in Biomedical Engineering, vol. 14, pp. 342–356, 2020

    2020

  23. [31]

    imigue: An identity-free video dataset for micro-gesture understanding and emotion analysis,

    X. Liu, H. Shi, H. Chen, Z. Yu, X. Li, and G. Zhao, “imigue: An identity-free video dataset for micro-gesture understanding and emotion analysis,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 10 631– 10 642

    2021

  24. [32]

    Ex- hubert: Enhancing hubert through block extension and fine-tuning on 37 emotion datasets,

    S. Amiriparian, F. Packa ´n, M. Gerczuk, and B. W. Schuller, “Ex- hubert: Enhancing hubert through block extension and fine-tuning on 37 emotion datasets,”arXiv preprint arXiv:2406.10275, 2024

    2024 arXiv

  25. [33]

    Smote: synthetic minority over-sampling technique,

    N. V . Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “Smote: synthetic minority over-sampling technique,”Journal of artificial intelligence research, vol. 16, pp. 321–357, 2002

    2002

  26. [34]

    A unified approach to interpreting model predic- tions,

    S. Lundberg, “A unified approach to interpreting model predic- tions,”arXiv preprint arXiv:1705.07874, 2017

    2017 arXiv

  27. [35]

    Explaining prediction models and individual predictions with feature contributions,

    E. ˇStrumbelj and I. Kononenko, “Explaining prediction models and individual predictions with feature contributions,”Knowledge and information systems, vol. 41, pp. 647–665, 2014

    2014

Pith tools

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