Pith. sign in

REVIEW 4 major objections 5 minor 29 references

Developing a Top-tier Framework in Naturalistic Conditions Challenge for Categorized Emotion Prediction: From Speech Foundation Models and Learning Objective to Data Augmentation and Engineering Choices

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A simple, reproducible speech emotion system outperforms more than 95% of challenge submissions on its own, and a three-system ensemble reaches a top-3 rank.

desk verdict A solid, reproducible shared-task SER system that deserves refereeing, but the abstract oversells the single-system rank and the ablations lack variance estimates. read the letter →

arxiv 2505.22133 v2 pith:N2O36RNB submitted 2025-05-28 cs.SD eess.AS

Tiantian Feng , Thanathai Lertpetchpun , Dani Byrd , Shrikanth Narayanan This is my paper
classification cs.SDeess.AS
keywords speechemotionrecognitionnaturalisticfoundationmodelssoftlabelingdataaugmentationclassimbalancemultimodalfusionreproduciblesystems
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 claims that a deliberately simple, reproducible system can reach top-tier performance in naturalistic speech emotion recognition, a task complicated by imbalanced labels and subjective annotation. The proposed SAILER system concatenates embeddings from a speech foundation model and a text foundation model, treats each sample's emotion as a distribution over labels rather than a single one-hot class, and adds two data augmentations called annotation dropout and audio mixing to handle class imbalance. The authors report that a single SAILER system without any ensembling exceeds 0.4 macro-F1 and outperforms more than 95% of the submitted systems, and that a three-system ensemble reaches a top-3 ranking. The practical upshot is that competitive results do not require bulky multi-model ensembles or heavy hyperparameter search, making the recipe easy to replicate and extend.

What carries the argument

The load-bearing components are five simple choices. First, a multimodal backbone that concatenates the pooled outputs of a speech foundation model and a text foundation model, where the text stream doubles as a source of semantic context. Second, soft labeling: each sample's multiple annotations are aggregated into an emotion distribution and trained with KL divergence, which recovers the roughly 19% of training samples that would be discarded under majority-vote hard labeling. Third, annotation dropout, which randomly drops 20% of a sample's annotations to make the model robust to noise in the estimated distribution. Fourth, audio mixing, which concatenates or overlaps a majority-class speech sample with a minority-class sample, averaging the two emotion distributions. Fifth, engineering choices: inverse-frequency distribution reweighting, a minority-class average-precision validation metric, and multi-task prediction of secondary emotions and arousal, valence, and dominance labels, with a weighted average over encoder layers feeding a 3-layer pointwise convolution and a two-layer MLP.

What would settle it

Repeat the Table 3 experiments with ten different random seeds each; if the mean macro-F1 of the no-augmentation baseline and the combined-augmentation condition differ by less than the run-to-run standard deviation, the claimed benefit of the augmentations collapses.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a simple multimodal recipe is enough to be competitive at the top of a hard, naturalistic emotion recognition benchmark. The system combines Whisper-Large V3 speech representations and RoBERTa-Large text representations by weighted pooling, concatenation, a pointwise convolution, and a two-layer MLP, trained with KL divergence against soft emotion distributions instead of cross-entropy against hard labels. Training with all annotations and two simple augmentations, annotation dropout and audio mixing, moves macro-F1 on the development set from 0.401 to 0.406, and adding multi-task prediction of secondary emotions and arousal, valence, and dominance attributes reaches 0.411. The final three-system ensemble scores 0.431 macro-F1 on the test set, a top-3 result among 166 submissions, and the paper stresses that all models use a fixed learning rate, 15 epochs, and no hyperparameter search.

Load-bearing premise

The load-bearing premise is that the small macro-F1 differences between configurations, measured once with a fixed seed, are genuine signal rather than random run-to-run variation, because the design recommendations rest on those differences.

Editorial extensions

If this is right

  • A single system without ensembling can outperform more than 95% of the submitted systems on this naturalistic emotion recognition challenge, so large ensembles are not required to reach the top of the leaderboard.
  • Soft labeling lets the model train on every annotated sample, including the approximately 19% of the training set that majority-vote hard labeling discards, and this contributes to the macro-F1 gain.
  • Both annotation dropout and audio mixing improve minority-class average precision, indicating that these augmentations are effective responses to the imbalance between majority and minority emotion classes.
  • Multi-task learning that also predicts secondary emotions and arousal, valence, and dominance labels raises overall macro-F1 from 0.406 to 0.411, though the paper notes it does not improve minority-class precision.
  • A three-system ensemble reaches a macro-F1 above 0.41 and a top-3 test ranking, and the authors suggest that ensembles beyond three systems are unnecessary for this level of performance.

Reading between the lines

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

  • The macro-F1 gaps that drive the augmentation and multi-task recommendations (0.401 versus 0.406, and 0.406 versus 0.411) come from single runs with a fixed seed, so the specific ordering of these choices is not statistically assured; the headline claim about single-system competitiveness does not depend on those small differences.
  • Because the recipe leans on off-the-shelf Whisper and RoBERTa features, transfer to domains very different from podcast speech and text, such as child speech or high-noise clinical recordings, is likely to be weaker than the challenge results suggest.
  • Annotation dropout and audio mixing are generic enough to test on other imbalanced subjective-label tasks, such as sentiment, intent, or pain-expression annotation, wherever soft label distributions are available.
  • The multi-task attribute branch helps overall macro-F1 but slightly hurts minority precision, implying that the auxiliary-task weighting is not yet tuned; a reader might recover the minority-class loss by reweighting the attribute losses.
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 / 5 minor

Summary. The paper presents SAILER, a multimodal speech emotion recognition system for the INTERSPEECH 2025 Emotion Recognition Challenge (Task 1). The system combines a speech foundation model (WavLM Large or Whisper-Large V3) with a text model (RoBERTa-Large), uses soft-label distribution modeling with KL-divergence loss, and introduces two data augmentations (annotation dropout and audio mixing) plus engineering choices such as distribution re-weighting, minority-class validation metrics, and multi-task prediction of secondary emotions and attribute labels. The authors report development-set ablations and leaderboard results, claiming that a single system exceeds 0.4 macro-F1, that an ensemble of three systems achieves a top-3 rank, and that the system outperforms more than 95% of challenge submissions. The code and best single system are publicly released.

Significance. If the reported results hold, the paper provides a practically valuable recipe for naturalistic emotion recognition: a relatively simple, reproducible, compute-efficient pipeline that reaches competitive leaderboard performance without large ensembles. The explicit release of code and the systematic ablation structure are strengths, as is the use of soft labels to leverage the full training set rather than discarding no-agreement samples. However, the significance is currently limited by two issues: the headline claim about outperforming 95% of submissions is internally inconsistent with the reported top-15 single-system rank, and the design-relevant ablation differences are presented without any variance estimate. The paper is therefore more a promising system description than a fully supported set of design recommendations.

major comments (4)
  1. [Abstract, Introduction, and §4.3] The abstract states that 'even a single system (without ensembling) can outperform more than 95% of the submissions,' but Section 4.3 says the single system 'achieves a top-15 ranking on the leaderboard,' and the Introduction attributes the 'outperforming 95% submissions' result to the three-system ensemble. With 166 submissions reported in Section 4.1, a top-15 rank outperforms 151/166 ≈ 91.0% of submissions, not more than 95%. The central quantitative claim is therefore internally inconsistent and must be corrected, together with exact official leaderboard ranks and submission counts.
  2. [§3.2, Tables 3 and 4] All configurations are trained once with a fixed seed. The reported macro-F1 differences are as small as 0.401 vs. 0.406 in Table 3 and 0.406 vs. 0.411 in Table 4, and the minority-class mAP differences are similarly small. Without multiple seeds, confidence intervals, or significance tests, the paper's design recommendations (that augmentation and multi-task learning improve SER performance) are not statistically supported. Please report variance estimates or soften the causal claims.
  3. [§2.3] The target distribution for mixed audio is set to d_mix = (d_maj + d_min)/2, and annotation dropout drops 20% of majority-class annotations based on the assumption that this moves the distribution closer to the oracle. Both are untested heuristics. Because the paper's main contribution is the set of design choices, these assumptions need at least a sensitivity analysis or some empirical validation (e.g., comparison against human judgments for mixed audio) before being claimed as effective.
  4. [§3.1 and §4.2] The manuscript is inconsistent about whether development-set samples were used for training. Section 3.1 says the authors 'choose not to do so' to ensure consistent findings, but Section 4.2 states 'in this experiment, we included speech samples with other and no agreement labels from the development set as training data.' Please clarify which configurations used development data, and whether the Table 5 ensemble results include them, for reproducibility.
minor comments (5)
  1. [§6] The conclusion contains a typo: 'categorized motion prediction' should read 'categorized emotion prediction.'
  2. [Figure 3] The text and caption reference 'Figure 3: Our proposed data augmentation technique,' but no Figure 3 appears in the manuscript; the figure is missing.
  3. [§3.1] Please define the term 'test-3' and clarify how it relates to the leaderboard results; the text later refers simply to 'the test set.'
  4. [Table 5] The caption says the 3-system ensemble achieves macro-F1 above 0.41 on the test set, but the table also lists a 'Best Single System' row. Please clarify whether the single-system value is a test-set score and how it relates to the top-15 leaderboard rank.
  5. [§4.1 and §4.3] Ranking expressions are inconsistent: 'top 30 ranks,' 'top-15 ranking,' and 'top-3 performing team' should be formatted and defined consistently, with the number of submissions explicitly stated in each context.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the system is evaluated against an external, held-out challenge test set, and the authors' self-citations are ordinary method reuse rather than load-bearing evidence.

full rationale

This paper is an empirical systems paper for a shared task. The central results are macro-F1 scores and leaderboard ranks obtained on the INTERSPEECH 2025 Emotion Recognition Challenge test set (test-3), whose ground-truth labels are not public. The design choices described in Section 2—soft-label KL-divergence training, annotation dropout, audio mixing, inverse-frequency distribution reweighting, and multitask auxiliary labels—are training inputs and modeling decisions, not quantities defined in terms of the reported test outcomes. There is no equation in which a predicted quantity is defined as its input, no fitted parameter renamed as a prediction, and no uniqueness theorem imported from the authors' prior work to force a particular choice. The authors reuse their own previously published downstream architecture and hyperparameters [15], and they cite their own prior work [14] when motivating speech and text foundation model choices, but these are normal, transparent borrowings: the leaderboard performance is independently measured against an external benchmark and does not depend on whether the reused architecture was derived from first principles. The abstract's 'more than 95% of submissions' phrasing versus the paper's later 'top-15 ranking' is a reporting inconsistency, not a circularity, because it does not involve a derivation reducing to its own inputs. Overall, the paper's claimed contribution is self-contained with respect to external evaluation, so the circularity score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim depends on pretrained foundation models, on the task-specific assumption that mixed audio has the averaged emotion distribution, and on a 20% annotation dropout rate. These are domain assumptions and hand-chosen hyperparameters, not derived facts.

free parameters (6)
  • audio_mixing_probability p_a = not reported
    Controls how often majority-class samples are mixed with minority samples; no value is given in Section 2.3, which blocks exact reproduction.
  • annotation_dropout_rate = 0.2
    Fraction of majority-class annotations dropped per sample during training; chosen by hand as a robustness heuristic.
  • learning_rate = 0.0005
    Fixed across all experiments and taken from prior work [15] without tuning for this task.
  • training_epochs = 15
    Fixed across all experiments and taken from prior work [15].
  • downstream_conv_filter_size = 256
    One fixed filter size used for the downstream convolutional module across all models.
  • silence_overlap_duration_t = sampled uniformly from [0,2] seconds
    Random duration for silence or overlap in audio mixing; the exact sampling distribution is not fully specified.
assumptions (4)
  • domain assumption Pretrained Whisper-Large V3, WavLM Large, and RoBERTa-Large representations carry emotion-relevant information for this podcast corpus.
    The whole system relies on these external models; the paper only validates this through downstream task performance.
  • ad hoc to paper The perceived emotion distribution of mixed audio is the arithmetic mean of the two source distributions, d_mix = (d_maj + d_min)/2.
    Section 2.3 defines this target for mixed samples; no perceptual validation is offered.
  • ad hoc to paper Dropping 20% of majority-class annotations yields a target distribution closer to the true oracle distribution.
    Section 2.3 states that this improves robustness, but the exact rate is chosen by hand and not systematically tuned.
  • domain assumption A single fixed seed and one run per configuration is representative of model performance.
    Section 3.2 fixes the seed for all experiments; no confidence intervals or repeated runs are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Developing a Top-tier Framework in Naturalistic Conditions Challenge for Categorized Emotion Prediction: From Speech Foundation Models and Learning Objective to Data Augmentation and Engineering Choices." pith.science (2026). https://pith.science/paper/N2O36RNB

@misc{pith2026250522133,
  author       = {Pith},
  title        = {Pith review of: Developing a Top-tier Framework in Naturalistic Conditions Challenge for Categorized Emotion Prediction: From Speech Foundation Models and Learning Objective to Data Augmentation and Engineering Choices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N2O36RNB}},
  note         = {Machine review of arXiv:2505.22133}
}
read the original abstract

Speech emotion recognition (SER), particularly for naturally expressed emotions, remains a challenging computational task. Key challenges include the inherent subjectivity in emotion annotation and the imbalanced distribution of emotion labels in datasets. This paper introduces the \texttt{SAILER} system developed for participation in the INTERSPEECH 2025 Emotion Recognition Challenge (Task 1). The challenge dataset, which contains natural emotional speech from podcasts, serves as a valuable resource for studying imbalanced and subjective emotion annotations. Our system is designed to be simple, reproducible, and effective, highlighting critical choices in modeling, learning objectives, data augmentation, and engineering choices. Results show that even a single system (without ensembling) can outperform more than 95\% of the submissions, with a Macro-F1 score exceeding 0.4. Moreover, an ensemble of three systems further improves performance, achieving a competitively ranked score (top-3 performing team). Our model is at: https://github.com/tiantiaf0627/vox-profile-release.

Figures

Figures reproduced from arXiv: 2505.22133 by the authors.

Figure 2
Figure 2. Our proposed soft-labeling approach. This leads to the use of the distribution learning loss (e.g., KL-Divergence). augmentation method could potentially increase the SER per￾formance by a large margin. Here, we define majority classes as neutral, happy, sad, and angry, while the remaining emotions are minority classes. We introduce two novel data augmenta￾tions to effectively address the data imbalance issue: annot… view at source ↗
Figure 3
Figure 3. Our proposed data augmentation technique to address data imbalance in SER modeling [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 16 canonical work pages

  1. [18]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inInternational conference on machine learning. PMLR, 2023, pp. 28 492–28 518

    2023

  2. [22]

    Peft-ser: On the use of parameter efficient trans- fer learning approaches for speech emotion recognition using pre- trained speech models,

    T. Fenget al., “Peft-ser: On the use of parameter efficient trans- fer learning approaches for speech emotion recognition using pre- trained speech models,” in2023 11th International Conference on Affective Computing and Intelligent Interaction (ACII). IEEE, 2023, pp. 1–8

    work page 2023

  3. [1]

    Introduction Despite notable advances in speech emotion recognition (SER) driven by speech foundation models [1, 2], accurately recogniz- ing emotions from speech remains a challenging computational problem in machine learning [3]. Key challenges include in- herent ambiguity in emotion expressions [4] and subjectivity in annotating emotions [5, 6], imbala...

    work page

  4. [2]

    Developing a Top-tier Framework in Naturalistic Conditions Challenge for Categorized Emotion Prediction: From Speech Foundation Models and Learning Objective to Data Augmentation and Engineering Choices

    Method 2.1. Speech Foundation Model The use of speech foundation models, such as Whisper [11] and WavLM [12], have demonstrated effectiveness in SER. Many recent studies [13, 14] show that simply leveraging pre-trained speech representations is adequate to achieve competitive per- formance compared to traditional hand-crafted features. Fur- arXiv:2505.221...

    work page Pith review arXiv 2025

  5. [3]

    Dataset The IS2025 Emotion Recognition Challenge used the MSP- Podcast dataset v1.12 [19, 10]

    Dataset and Experiment 3.1. Dataset The IS2025 Emotion Recognition Challenge used the MSP- Podcast dataset v1.12 [19, 10]. The dataset consists of podcast data from the Internet, including spontaneous speech with nat- ural human emotion expressions. The dataset is annotated with different emotion attributes. The dataset consists of five sub- sets: the tra...

    work page

  6. [4]

    Do Speech Foundation Models Impact SER? As suggested by [14], we first investigate whether the choice of speech foundation models impacts the SER performance

    Results 4.1. Do Speech Foundation Models Impact SER? As suggested by [14], we first investigate whether the choice of speech foundation models impacts the SER performance. Specifically, we compare the performance of WavLM Large and Table 3:Comparisons of models trained without augmentation to those with audio mixing, annotation dropout, and both com- bine...

    work page

  7. [5]

    One is to study pre-trained speech models with emotional speech data like Emotion2Vec [21]

    On Further Improvements While our current system shows competitive performance, we highlight several promising and easy-to-prototype directions that researchers can explore for developing next-generation, state-of-the-art SER systems. One is to study pre-trained speech models with emotional speech data like Emotion2Vec [21]. Learning Objective.While we us...

    work page

  8. [6]

    Experimental results show thatSAILERis highly competitive in the IS25-SER challenge, achieving top-tier per- formance with minimum system complexity

    Conclusion In this work, we describe theSAILERframework, a simple and reproducible-friendly SER model for categorized motion pre- diction in task 1 of the IS25-SER challenge.SAILERconsiders designs from speech foundation models and learning objectives to data augmentation and engineering choices to tackle imbal- anced data. Experimental results show thatS...

    work page

Show all 29 references
  1. [7]

    Acknowledgment We gratefully acknowledge support from IARPA ARTS (award number 140D0424C0067, JHU subcontract) from the Office of the Director of National Intelligence and NSF Grant (SCH with award number 2204942)

  2. [8]

    On the opportunities and risks of foundation models,

    R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskill et al., “On the opportunities and risks of foundation models,” arXiv preprint arXiv:2108.07258, 2021

    2021 arXiv

  3. [9]

    Dawn of the trans- former era in speech emotion recognition: closing the valence gap,

    J. Wagner, A. Triantafyllopoulos, H. Wierstorf, M. Schmitt, F. Burkhardt, F. Eyben, and B. W. Schuller, “Dawn of the trans- former era in speech emotion recognition: closing the valence gap,”IEEE Transactions on Pattern Analysis and Machine Intel- ligence, vol. 45, no. 9, pp. ...

    2023

  4. [10]

    An engineering view on emotions and speech: From analysis and pre- dictive models to responsible human-centered applications,

    C.-C. Lee, T. Chaspari, E. M. Provost, and S. S. Narayanan, “An engineering view on emotions and speech: From analysis and pre- dictive models to responsible human-centered applications,”Pro- ceedings of the IEEE, vol. 111, no. 10, pp. 1142–1158, 2023

    2023

  5. [11]

    Interpreting ambiguous emotional expressions,

    E. Mower, A. Metallinou, C.-C. Lee, A. Kazemzadeh, C. Busso, S. Lee, and S. S. Narayanan, “Interpreting ambiguous emotional expressions,” inProceedings of the International Conference on Affective Computing and Intelligent Interaction (ACII), Amster- dam, The Netherlands, sep 2009

    2009

  6. [12]

    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). IEEE, 2024, pp. 11 861–11 865

    2024

  7. [13]

    People make mistakes: Ob- taining accurate ground truth from continuous annotations of sub- jective constructs,

    B. M. Booth and S. S. Narayanan, “People make mistakes: Ob- taining accurate ground truth from continuous annotations of sub- jective constructs,”Behavior Research Methods, vol. 56, no. 8, pp. 8784–8800, 2024

    2024

  8. [14]

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

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

    2024

  9. [15]

    1st place solution to odyssey emotion recognition chal- lenge task1: Tackling class imbalance problem,

    M. Chen, H. Zhang, Y . Li, J. Luo, W. Wu, Z. Ma, P. Bell, C. Lai, J. D. Reiss, L. Wang, P. C. Woodland, X. Chen, H. Phan, and T. Hain, “1st place solution to odyssey emotion recognition chal- lenge task1: Tackling class imbalance problem,” inThe Speaker and Language Recognitio...

    2024

  10. [16]

    Double multi-head attention multimodal system for odyssey 2024 speech emotion recognition challenge,

    F. Costa, M. India, and J. Hernando, “Double multi-head attention multimodal system for odyssey 2024 speech emotion recognition challenge,” inThe Speaker and Language Recognition Workshop (Odyssey 2024), 2024, pp. 266–273

    2024

  11. [17]

    The interspeech 2025 challenge on speech emotion recognition in naturalistic conditions,

    A. R. Naini, L. Goncalves, A. N. Salman, P. Mote, I. R. ¨Ulgen, T. Thebaud, L. Velazquez, L. P. Garcia, N. Dehak, B. Sisman, and C. Busso, “The interspeech 2025 challenge on speech emotion recognition in naturalistic conditions,” inInterspeech 2025, vol. To appear, Rotterdam, ...

    2025

  12. [19]

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

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

    2022

  13. [20]

    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,” inInterspeech 2021, 2021, pp. 3400–3404

    2021

  14. [21]

    Foundation model assisted automatic speech emotion recognition: Transcribing, annotating, and aug- menting,

    T. Feng and S. Narayanan, “Foundation model assisted automatic speech emotion recognition: Transcribing, annotating, and aug- menting,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 12 116–12 120

    2024

  15. [23]

    Fusing asr outputs in joint training for speech emotion recognition,

    Y . Li, P. Bell, and C. Lai, “Fusing asr outputs in joint training for speech emotion recognition,” inICASSP 2022-2022 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 7362–7366

    2022

  16. [24]

    A framework for automatic human emotion classification using emotional profiles,

    E. Mower, M. J. Mataric, and S. S. Narayanan, “A framework for automatic human emotion classification using emotional profiles,” IEEE Transactions on Audio, Speech and Language Processing, vol. 19, no. 5, pp. 1057–1070, 2011

    2011

  17. [25]

    Minority views matter: Evaluating speech emotion classifiers with human subjective annotations by an all- inclusive aggregation rule,

    H.-C. Chou, L. Goncalves, S.-G. Leem, A. N. Salman, C.-C. Lee, and C. Busso, “Minority views matter: Evaluating speech emotion classifiers with human subjective annotations by an all- inclusive aggregation rule,”IEEE Transactions on Affective Com- puting, 2024

    2024

  18. [26]

    Building naturalistic emotionally bal- anced speech corpus by retrieving emotional speech from existing podcast recordings,

    R. Lotfian and C. Busso, “Building naturalistic emotionally bal- anced speech corpus by retrieving emotional speech from existing podcast recordings,”IEEE Transactions on Affective Computing, vol. 10, no. 4, pp. 471–483, 2017

    2017

  19. [27]

    Roberta: A robustly optimized bert pretraining approach,

    Y . Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V . Stoyanov, “Roberta: A robustly optimized bert pretraining approach,”arXiv preprint arXiv:1907.11692, 2019

    1907 arXiv

  20. [28]

    emotion2vec: Self-supervised pre-training for speech emotion representation,

    Z. Ma, Z. Zheng, J. Ye, J. Li, Z. Gao, S. Zhang, and X. Chen, “emotion2vec: Self-supervised pre-training for speech emotion representation,” inFindings of the Association for Computational Linguistics ACL 2024, 2024, pp. 15 747–15 760

    2024

  21. [29]

    Emix: a data augmentation method for speech emotion recognition,

    A. Dang, T. H. Vu, J.-C. Wanget al., “Emix: a data augmentation method for speech emotion recognition,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

Pith tools

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