Pith. sign in

REVIEW 5 major objections 8 minor 1 cited by

Token-Level Logits Matter: A Closer Look at Speech Foundation Models for Ambiguous Emotion Recognition

T0 review · 5 major / 8 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 foundation models encode ambiguous emotions in hidden token-level logits, even when their text responses get the emotion wrong.

desk verdict A genuinely new token-level probing idea for ambiguous emotion in speech foundation models, but Eq. 2's normalization is not a probability model and the quantitative claims rest on it. read the letter →

arxiv 2505.18484 v1 pith:INGTXCYZ submitted 2025-05-24 cs.SD eess.AS

Jule Valendo Halim , Siyi Wang , Hong Jia , Ting Dang This is my paper · ORCID
classification cs.SDeess.AS
keywords ambiguousemotionrecognitionspeechfoundationmodelstoken-levellogitsdistributionzero-shotIEMOCAPmultimodallargelanguageaffectivecomputing
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

Speech foundation models are typically evaluated on single-label emotion recognition, but human emotion is often ambiguous and better described by a distribution over emotion categories. This paper asks whether pretrained speech foundation models can recover such ambiguous emotion distributions without any fine-tuning. It finds that the models' generated text often misreads the emotion, yet their token-level logits—the raw scores assigned to emotion-related subword tokens as the response is generated—contain a distribution closer to human soft labels. The paper proposes extracting these distributions by averaging the emotion-token logits across the generated sequence and sum-normalizing over the four emotion classes, and reports that this approach beats the text-based reading and is robust to prompt changes.

What carries the argument

The central object is the token-level logit vector restricted to emotion-related subword tokens. At each autoregressive step, the model's output layer produces logits over its whole vocabulary; the paper keeps only logits for subwords of the four emotion words (e.g., "Ang" and "er" for anger), averages the subword logits for each emotion word, averages these four vectors over the entire generated token sequence, and converts them into a distribution by sum-normalization. This extraction turns the model's internal next-token predictions into an emotion distribution before text commitment, which is what carries the argument: it makes visible an emotion posterior that the sampling or decoding to text does not preserve.

What would settle it

Recompute the token-level distributions on the same 4,373 IEMOCAP utterances after replacing the sum-normalization in Eq. (2) with a softmax over the four emotion-token scores, and check whether the reported KL, BD, and R2 advantages over text-level analysis survive; if they do not, the central result depends on the uncalibrated normalization rather than on emotional information in the logits.

Watch

Extended reading notes

Core claim

The central claim is that a pretrained speech foundation model holds an implicit posterior over emotion categories that is more faithful than the text it outputs. Using the IEMOCAP corpus and four emotion classes (anger, happiness, sadness, neutral), the paper compares a text-level extraction (parsing percentages from generated responses) with a token-level extraction (averaging the logits of emotion-related subword tokens across all generated tokens, then sum-normalizing via $\phi_{e_n} = z_{e_n} / \sum_{n=1}^N z_{e_n}$). Token-level extraction improves KL divergence from 2.05 to 0.99 and Bhattacharyya distance from 0.51 to 0.47 relative to text-level extraction, and it remains stable when the prompt is changed from ambiguous-emotion to single-emotion wording. The paper concludes that this internal knowledge comes from pretraining and is not fully translated into the models' articulations, making token-level readout an effective way to infer ambiguous emotions.

Load-bearing premise

The argument assumes that averaging raw next-token scores of hand-picked emotion-related subwords across the generated output, then dividing by their sum over four emotions, produces a probability distribution that can be fairly compared with the human soft-label distribution; if those raw scores are negative or sum to zero, the result is not a valid distribution, and the 2.1% of outputs that produced invalid distributions were omitted.

Editorial extensions

If this is right

  • Token-level analysis outperforms text-level analysis on ambiguous emotion distributions, with relative improvements of 51.71% in KL divergence, 7.84% in Bhattacharyya distance, and 11.76% in R2.
  • Token-level emotion distributions are stable across prompting strategies: switching from an ambiguous-emotion prompt to a single-emotion prompt improves all metrics (KL 0.99 to 0.75, BD 0.47 to 0.31, R2 0.38 to 0.53).
  • Including logits from numerical percentage tokens, not just emotion-word tokens, improves the extracted distribution, so the full output sequence carries emotional information.
  • The same token-level distributions can be collapsed to a single-emotion prediction that reaches accuracy comparable to zero-shot text baselines without any fine-tuning.

Reading between the lines

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

  • If the finding generalizes to other speech foundation models, logit-level readouts could be used as a calibration signal during decoding, steering the model's text toward the emotional distribution it already computed internally.
  • A direct next experiment would test the same extraction on a second speech foundation model and a second corpus with more than three annotators, to see whether the 51.71% KL improvement is model-specific or a general property of pretrained speech models.
  • The paper's omission of the 2.1% of outputs that produced invalid distributions means a robustness check should re-run the comparison on the full set, treating invalid outputs as a separate failure mode rather than dropping them.
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 / 8 minor

Summary. This paper investigates whether speech foundation models (SFMs) can produce distributional emotion predictions on ambiguous speech. Using LTU-AS as the SFM and 4,373 IEMOCAP utterances with four emotion classes, the authors compare two extraction approaches. The text-level approach parses the percentages that the model writes in its response into a distribution; the token-level approach averages the raw logits of hand-selected emotion-related subword tokens over all generated output positions and then sum-normalizes the averaged logits (Eq. 2) to form a four-class 'posterior' distribution. Both are evaluated against human soft-label distributions using KL divergence, Bhattacharyya distance, and R2. The paper reports that the token-level approach improves KL by 51.71% relative to the text-level approach, that single-emotion prompting improves token-level results further (Table 2), and concludes that SFMs encode ambiguous emotion priors at the token level even when their text responses are inaccurate.

Significance. If the core claims held, the paper would make a useful contribution: internal token-level scores of a pretrained speech LLM would carry soft emotion information partially lost in the text output, extractable without fine-tuning. The design has genuine strengths: external human soft labels as targets (no circularity), no fitted parameters, and a concrete falsifiable comparison. The evidence, however, does not yet establish the claims: Eq. (2) is not a valid logit-to-distribution conversion; the 2.1% exclusion and the Section 5.3 test-set-based choice introduce selection bias; and the absence of error bars or significance tests leaves small differences unassessable. The central hypothesis is worth testing with a softmax-based recalibration and significance testing; the present evaluation is suggestive but not reliable.

major comments (5)
  1. [§3.4.2, Eq. (2)] The conversion φ_en = z_en / Σ_n z_n does not generally produce a probability distribution. Raw logits are unbounded, can be negative, and their sum over the four emotion classes can be zero or negative, in which case φ has negative components or is undefined; the text itself acknowledges the goal of a 'valid emotion distribution' without achieving it. No softmax, temperature, or calibration is applied, so referring to these vectors as 'posterior probability distributions' (Abstract, §1, §3.4.2) is inaccurate. Because every headline result in Tables 1–2 and Figure 3 evaluates KL, BD, and R2 against φ, the central claim that token-level logits encode ambiguous emotion distributions is currently supported only by an ad hoc rescaling. The authors should replace Eq. (2) with a proper transformation (e.g., softmax over the emotion-token logits, with temperature reported) and re-run all distributional comparisons.
  2. [§5.1] The paper states that 2.1% of outputs produced invalid distributions and were omitted, but it does not specify whether this applies to the text-level or token-level pipeline, nor what 'invalid' means operationally (sum of z_n equal to zero? negative components for some emotions? percentages not summing to 100?). If the excluded cases are systematically the most ambiguous utterances or those with strongly negative logits, all reported KL/BD/R2 values are biased by the exclusion. The authors should report the exclusion rate per condition, characterize the excluded utterances, and note that a softmax conversion (which is always valid) would make the exclusion unnecessary.
  3. [§5.3, Fig. 3] The extraction scheme used in the main results (logits from emotion-related text plus numerical percentage tokens) is selected by comparing it with the emotion-text-only variant on the evaluation set and keeping the better performer. This is test-set-based method selection, and the reported advantages—including the 51.71% relative KL improvement claimed in §5.1—inherit an optimistic bias. The scheme should be pre-specified, chosen on a held-out validation set, or reported with both variants shown without presenting the better one as the method's expected performance.
  4. [§5.1, Tables 1–3] All results are point estimates from a single run of a single model (LTU-AS), without error bars, bootstrap confidence intervals, or significance tests, yet §5.1 uses 'significantly outperforms.' With 4,373 utterances, a paired per-utterance analysis (e.g., Wilcoxon signed-rank or paired bootstrap) is feasible and should be reported. The KL gap (0.99 vs 2.05) is large, but the BD gap (0.47 vs 0.51) and the R2 gap (0.34 vs 0.38) are small enough that they could be within noise, and the same applies to the cross-prompt comparison in Table 2 (KL 0.99 to 0.75, BD 0.47 to 0.31, R2 0.38 to 0.53).
  5. [§5.2, Conclusion] Two claims outrun the data. First, 'robust ... across different prompts' (Abstract) is difficult to reconcile with the magnitude of the Table 2 differences, where the single-emotion prompt changes all three metrics by roughly 20–35%; the authors should analyze whether these differences are significant and temper the robustness claim accordingly. Second, the Conclusion's statement that the approach 'outperform[s] state-of-the-art zero-shot baselines' is not supported under the ambiguous prompt, where Table 1 shows the baseline R2 (0.51) exceeds the token-level R2 (0.38); the claim holds only for the single-prompt variant in Table 2. In addition, the assertion in §1 that text responses are 'highly dependent on prompts' is never tested, since no text-level single-prompt condition is reported.
minor comments (8)
  1. [Abstract/§1] There are grammar errors such as 'an token-level' (should be 'a token-level'); the paper needs a careful copy-edit.
  2. [§3.4.2] Equation (1) is garbled in the typeset text ('zen j = 1 K PK k=1 zen j,k'); define K explicitly as the number of subword tokens per emotion word and note that the token split is model-specific (e.g., 'Anger' becomes 'Ang' + 'er').
  3. [§3.4.2/Fig. 2] Figure 2 shows non-target tokens such as 'car', 'mo', 'tion' alongside the target tokens; clarify the rule for selecting emotion-related tokens from the vocabulary (exact string matching? prefix matching?) and confirm that the same token set is used for both prompts.
  4. [§4.1] Specify the utterance-selection criterion for 'utterances that only include four emotion classes'—for instance, whether all annotator labels must fall within the four classes—and report how many utterances were excluded by this filter.
  5. [§5.1] The sentence 'it is possibly due to their use of the advanced Gemini-1.5 model' is speculative; either remove it or support it with an ablation that varies the model family or scale.
  6. [Table 3] The Text (Ambiguous) F1 of 0.06 deserves a brief explanation, since a reader could otherwise mistake it for a bug rather than the model's failure to produce majority-vote-matching single labels under the ambiguous prompt.
  7. [§4.2] The paper provides no decoding details (temperature, sampling seed, maximum number of generated tokens, number of runs) and no code release; these are needed to reproduce the token-level averaging exactly.
  8. [§4.1] IEMOCAP ground-truth distributions are based on roughly three annotators per utterance, so the targets are very coarse (masses that are multiples of 1/3, with many unanimous utterances); a sentence noting this and its effect on absolute KL values would help interpretation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the token-level emotion distributions are a fixed, parameter-free function of model logits and are evaluated against external human soft labels.

full rationale

The paper's derivation chain is not circular. The predicted token-level distribution is computed by a fixed transformation of the SFM's output logits: Eq. (1) averages logits over subword tokens and the text describes averaging over all generated output tokens, then Eq. (2) normalizes by the sum over the four emotion classes. No parameter is fitted to the evaluation data, and no result is imported from the authors' prior work as a load-bearing premise. The ground truth is the external IEMOCAP human soft-label distribution, so the KL, BD, and R2 comparisons measure agreement with an independently defined target rather than with the method's own inputs. The baseline from [15] and related-work citations involving the same research group are used for comparison or background, not to justify the central token-level extraction, so they do not constitute self-citation circularity. The paper's admitted omission of 2.1% of outputs with invalid distributions (Section 5.1) and the concern that raw logits normalized by Eq. (2) may not be calibrated probabilities are validity or correctness caveats, not circularity: even if Eq. (2) is not a proper probability model, the predicted quantity is still derived from the model's logits and compared with external labels rather than being defined in terms of the target. The central empirical claim is self-contained as a study of what a fixed logit aggregation yields.

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

The central claim rests on two unverified premises: that Eq. 2 turns logits into valid probabilities, and that the unspecified conversion of three annotator labels into a soft label is appropriate. There are no fitted numeric constants and no invented entities; the method's design choices are hand-selected rather than learned.

assumptions (3)
  • ad hoc to paper The average of raw logits of emotion-related tokens across the generated output sequence, normalized by their sum, represents a categorical emotion distribution.
    Eq. 2; raw logits can be negative or sum to zero, so this is not a standard probability transform and no justification is given.
  • domain assumption The ground truth emotion distribution can be obtained from the IEMOCAP annotator labels, although the conversion rule is not specified.
    Section 4.1 only says annotations by three annotators and selected files; the soft-label construction is required for KL, BD, and R2 comparisons.
  • domain assumption The hand-picked subword token lists, such as "ang", "er", "ne", "ut", "ral", are sufficient to represent each emotion in the vocabulary.
    Section 3.4.2; performance depends on this mapping, and no validation of alternate token lists is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Token-Level Logits Matter: A Closer Look at Speech Foundation Models for Ambiguous Emotion Recognition." pith.science (2026). https://pith.science/paper/INGTXCYZ

@misc{pith2026250518484,
  author       = {Pith},
  title        = {Pith review of: Token-Level Logits Matter: A Closer Look at Speech Foundation Models for Ambiguous Emotion Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/INGTXCYZ}},
  note         = {Machine review of arXiv:2505.18484}
}
read the original abstract

Emotional intelligence in conversational AI is crucial across domains like human-computer interaction. While numerous models have been developed, they often overlook the complexity and ambiguity inherent in human emotions. In the era of large speech foundation models (SFMs), understanding their capability in recognizing ambiguous emotions is essential for the development of next-generation emotion-aware models. This study examines the effectiveness of SFMs in ambiguous emotion recognition. We designed prompts for ambiguous emotion prediction and introduced two novel approaches to infer ambiguous emotion distributions: one analysing generated text responses and the other examining the internal processing of SFMs through token-level logits. Our findings suggest that while SFMs may not consistently generate accurate text responses for ambiguous emotions, they can interpret such emotions at the token level based on prior knowledge, demonstrating robustness across different prompts.

Figures

Figures reproduced from arXiv: 2505.18484 by the authors.

Figure 1
Figure 1. System overview. Speech utterances are processed by SFMs to generate emotion distributions, which are then com￾pared with the ground truth inferred from M human annotators. based on traditional modelling methods. In the era of large-scale SFMs, employing a universal SFM trained on extensive speech data has opened up new possibili￾ties for emotion understanding, moving away from reliance on task-specific small models… view at source ↗
Figure 2
Figure 2. Framework for ambiguous emotion recognition using SFMs. By providing a prompt alongside speech, it enables the extraction of both the generated text and posterior probabilities at i) text-level and ii) token-level, respectively. where K is the number of tokens for each emotion word en, and z en j is the averaged logits for the emotion en. This leads to a vector zj = [z e1 j , z e2 j , . . . , z eN j ] for each outpu… view at source ↗
Figure 3
Figure 3. Performance comparison utilizing logits of i) both emotion-related text and numerical percentage and ii) only emotion-related text [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Beyond Scale: Small Language Models are Comparable to GPT-4 in Mental Health Understanding

    cs.CL 2025-07 reject novelty 4.0 of 10

    On six public mental health text datasets, small language models reach macro F1 within 0.02 of large LLM baselines on binary tasks, but the comparison relies on previously published scores.

Reference graph

Works this paper leans on

36 extracted references · 31 canonical work pages · cited by 1 Pith paper

  1. [1]

    articulate

    Introduction Speech emotion recognition (SER) has experienced signifi- cant growth in areas such as mental health monitoring [1] and human-computer interaction (HCI) [2]. Recent advancements in large language models (LLMs) have led to the development of advanced speech foundation models (SFMs), which integrate speech inputs with LLMs. These models incorpo...

    work page

  2. [2]

    delivery skills

    Related Work While the development of SER systems for single emotion classification has advanced for decades [16, 17, 18, 19], the progress in recognizing ambiguous emotions is still lagging behind. Emotion complexity and ambiguity were first recog- nized in [20], which recommended classifying emotions based on soft-labels rather than relying solely on ca...

    arXiv 2025

  3. [3]

    ang”, followed by “er

    Methodology 3.1. Problem definition As shown in Figure 1, given a speech utterance xt, the objec- tive is to infer the ambiguous emotion distribution p(ˆy|xt, θ) using a SFM parameterized by θ. The predicted distribution encompasses N emotion classes, with each probability indicat- ing the likelihood of a specific emotion, while the overall distri- bution...

    work page

  4. [4]

    Text-Level Analysis 6 er

    work page

  5. [5]

    iness 6.46 2.44 0.43 0.08 0.7

    Token-Level Analysis Ang Emotion-related Tokens Logit Ang, er Ne, ut, ral Sad, ness H, app, iness 6.46 , 2.44 3.00, 0.63, 1.090.39, 1.04 -0.53, 2.64, 3.32 Ang er car mo tion .......... iness 6.46 2.44 0.43 0.08 0.7 .......... -3.32 Vocabulary Logit Target Tokens Averaged logits Anger Neutral Sadness Happiness 4.45 1.57 0.71 1.81 Predicted Distribution (To...

    work page

  6. [6]

    Experimental setup 4.1. Dataset The Interactive Emotional Dyadic Motion Capture (IEMOCAP) database [25] is used, consisting of approximately 12 hours of audiovisual recordings of dyadic conversations between ac- tors. Each recording is annotated by three annotators each. Specifically, we select utterances that only include four emo- tion classes: Happines...

    work page

  7. [7]

    Performance on ambiguous SER Table 1 presents the performance of ambiguous SER using ambiguity-based metrics, evaluating the entire distribution pre- dictions

    Results 5.1. Performance on ambiguous SER Table 1 presents the performance of ambiguous SER using ambiguity-based metrics, evaluating the entire distribution pre- dictions. 2.1% of the outputs produced invalid distributions and were omitted from the evaluation. Our approach at the token level achieves the best KL divergence and BD, outperforming the basel...

    work page

  8. [8]

    Conclusion Our study investigated the extent to which pretrained SFMs can interpret ambiguity in SER based on their prior knowledge and introduced two approaches to infer ambiguous emotion distri- butions at both the text and token levels. Our findings sug- gest that SFMs can recognize the nuanced ambiguity present in emotional speech due to its prior kno...

    work page

Show all 36 references
  1. [9]

    Speech emotion recognition using su- pervised deep recurrent system for mental health monitoring,

    N. Elsayed, Z. ElSayed, N. Asadizanjani, M. Ozer, A. Abdel- gawad, and M. Bayoumi, “Speech emotion recognition using su- pervised deep recurrent system for mental health monitoring,” 2024

    2024

  2. [10]

    Speech emotion recogni- tion approaches in human computer interaction,

    S. Ramakrishnan and I. M. El Emary, “Speech emotion recogni- tion approaches in human computer interaction,”Telecommunica- tion Systems, vol. 52, pp. 1467–1478, 2013

    2013

  3. [11]

    Beyond silent letters: Amplifying llms in emotion recognition with vocal nuances,

    Z. Wu, Z. Gong, L. Ai, P. Shi, K. Donbekci, and J. Hirschberg, “Beyond silent letters: Amplifying llms in emotion recognition with vocal nuances,” 2024

    2024

  4. [12]

    From text to emotion: Unveiling the emotion annotation capabilities of llms,

    M. Niu, M. Jaiswal, and E. M. Provost, “From text to emotion: Unveiling the emotion annotation capabilities of llms,” INTER- SPEECH, 2024

    2024

  5. [13]

    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,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 12 116–12 120

    2024

  6. [14]

    Interpreting ambiguous emotional ex- pressions,

    E. Mower, A. Metallinou, C. M. Lee, A. Kazemzadeh, C. Busso, S. Lee, and S. Narayanan, “Interpreting ambiguous emotional ex- pressions,” in Proceedings of the International Conference on Af- fective Computing and Intelligent Interaction , 2009, pp. 1–8

    2009

  7. [15]

    Emotion recog- nition systems must embrace ambiguity,

    J. Wu, T. Dang, V . Sethu, and E. Ambikairajah, “Emotion recog- nition systems must embrace ambiguity,” 2024

    2024

  8. [17]

    An investi- gation of emotion prediction uncertainty using gaussian mixture regression

    T. Dang, V . Sethu, J. Epps, and E. Ambikairajah, “An investi- gation of emotion prediction uncertainty using gaussian mixture regression.” in INTERSPEECH, 2017, pp. 1248–1252

    2017

  9. [18]

    Dynamic multi-rater gaussian mixture regression incorporating temporal dependencies of emotion uncertainty using kalman filters,

    T. Dang, V . Sethu, and E. Ambikairajah, “Dynamic multi-rater gaussian mixture regression incorporating temporal dependencies of emotion uncertainty using kalman filters,” in 2018 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018, ...

    2018

  10. [19]

    Child emotion recognition using probabilistic neural network with effective features,

    M. N. Mohanty and H. K. Palo, “Child emotion recognition using probabilistic neural network with effective features,” Measure- ment, vol. 152, p. 107369, 2020

    2020

  11. [20]

    Text emotion distribution learning via multi-task convolutional neural network

    Y . Zhang, J. Fu, D. She, Y . Zhang, S. Wang, and J. Yang, “Text emotion distribution learning via multi-task convolutional neural network.” in IJCAI, 2018, pp. 4595–4601

    2018

  12. [21]

    A novel se- quential monte carlo framework for predicting ambiguous emo- tion states,

    J. Wu, T. Dang, V . Sethu, and E. Ambikairajah, “A novel se- quential monte carlo framework for predicting ambiguous emo- tion states,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2022, pp. 8567–8571

    2022

  13. [22]

    Dual-constrained dynamical neural odes for ambiguity-aware continuous emotion prediction,

    J. Wu, T. Dang, V . Sethu, and E. Ambikairajah, “Dual-constrained dynamical neural odes for ambiguity-aware continuous emotion prediction,” in Proc. Interspeech 2024, 2024, pp. 3185–3189

    2024

  14. [23]

    Aer-llm: Ambiguity- aware emotion recognition leveraging large language models,

    X. Hong, Y . Gong, V . Sethu, and T. Dang, “Aer-llm: Ambiguity- aware emotion recognition leveraging large language models,” in ICASSP, 2024

    2024

  15. [24]

    Robust speech emotion recognition using cnn+ lstm based on stochastic fractal search optimization algorithm,

    A. A. Abdelhamid, E.-S. M. El-Kenawy, B. Alotaibi, G. M. Amer, M. Y . Abdelkader, A. Ibrahim, and M. M. Eid, “Robust speech emotion recognition using cnn+ lstm based on stochastic fractal search optimization algorithm,” Ieee Access, vol. 10, pp. 49 265– 49 284, 2022

    2022

  16. [25]

    Speech emo- tion recognition based on multi-feature speed rate and lstm,

    Z. Yang, Z. Li, S. Zhou, L. Zhang, and S. Serikawa, “Speech emo- tion recognition based on multi-feature speed rate and lstm,”Neu- rocomputing, vol. 601, p. 128177, 2024

    2024

  17. [26]

    Multimodal speech emotion recognition based on large language model,

    C. Fang, Y . Jin, G. Chen, Y . Zhang, S. Li, Y . Ma, and Y . Xie, “Multimodal speech emotion recognition based on large language model,” IEICE TRANSACTIONS on Information and Systems , vol. 107, no. 11, pp. 1463–1467, 2024

    2024

  18. [27]

    Mul- timodal emotion recognition using feature fusion: An llm-based approach,

    C. O. Kumar, N. Gowtham, M. Zakariah, and A. Almazyad, “Mul- timodal emotion recognition using feature fusion: An llm-based approach,” IEEE Access, 2024

    2024

  19. [28]

    Interpreting ambiguous emotional expressions,

    E. Mower, A. Metallinou, C.-C. Lee, A. Kazemzadeh, C. Busso, S. Lee, and S. Narayanan, “Interpreting ambiguous emotional expressions,” in 2009 3rd International Conference on Affective Computing and Intelligent Interaction and Workshops . IEEE, 2009, pp. 1–8

    2009

  20. [29]

    Multi-classifier interactive learning for ambiguous speech emotion recognition,

    Y . Zhou, X. Liang, Y . Gu, Y . Yin, and L. Yao, “Multi-classifier interactive learning for ambiguous speech emotion recognition,” IEEE/ACM transactions on audio, speech, and language process- ing, vol. 30, pp. 695–705, 2022

    2022

  21. [30]

    Handling ambiguity in emotion: From out- of-domain detection to distribution estimation,

    W. Wu, B. Li, C. Zhang, C.-C. Chiu, Q. Li, J. Bai, T. N. Sainath, and P. C. Woodland, “Handling ambiguity in emotion: From out- of-domain detection to distribution estimation,” arXiv preprint arXiv:2402.12862, 2024

    2024 arXiv

  22. [31]

    Using gaussian processes with lstm neural networks to predict continuous-time, dimen- sional emotion in ambiguous speech,

    M. Atcheson, V . Sethu, and J. Epps, “Using gaussian processes with lstm neural networks to predict continuous-time, dimen- sional emotion in ambiguous speech,” in 2019 8th International Conference on Affective Computing and Intelligent Interaction (ACII). IEEE, 2019, pp. 718–724

    2019

  23. [32]

    Emotionally numb or empathetic? evaluating how llms feel using emotionbench,

    J.-t. Huang, M. H. Lam, E. J. Li, S. Ren, W. Wang, W. Jiao, Z. Tu, and M. R. Lyu, “Emotionally numb or empathetic? evaluating how llms feel using emotionbench,” 2023

    2023

  24. [33]

    Iemocap: Interactive emotional dyadic motion capture database,

    C. Busso, M. Bulut, C.-C. Lee, A. Kazemzadeh, E. Mower, S. Kim, J. N. Chang, S. Lee, and S. S. Narayanan, “Iemocap: Interactive emotional dyadic motion capture database,”Language resources and evaluation, vol. 42, pp. 335–359, 2008

    2008

  25. [34]

    Joint audio and speech understanding,

    Y . Gong, A. H. Liu, H. Luo, L. Karlinsky, and J. Glass, “Joint audio and speech understanding,” in2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2023, pp. 1–8

    2023

  26. [35]

    Robust speech recognition via large-scale weak su- pervision,

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

    2023

  27. [36]

    Listen, think, and understand,

    Y . Gong, H. Luo, A. H. Liu, L. Karlinsky, and J. Glass, “Listen, think, and understand,” 2024

    2024

  28. [37]

    Af- fect recognition in conversations using large language models,

    S. Feng, G. Sun, N. Lubis, W. Wu, C. Zhang, and M. Ga ˇsi´c, “Af- fect recognition in conversations using large language models,” arXiv preprint arXiv:2309.12881, 2023

    2023 arXiv

Pith tools

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