Pith. sign in

REVIEW 3 major objections 31 references

Can Conversational Temporal Dynamics Improve Depression Detection in Dyads? A Preliminary Investigation in Multi-Modality Perspectives

T0 review · 3 major / 0 minor · reviewed 2026-07-12 · grok-4.5

Pith's one-line read A 24-feature map of how people take turns can match or beat large speech and text models for depression screening in clinical interviews.

desk verdict Solid small-scale transfer of fixed dyadic timing features to DAIC-WOZ depression screening; CTD beats frozen SSL unimodals on dev and fuses cleanly with text, but wizard-controlled ask-side timing and tiny n leave the transfer claim provisional. read the letter →

arxiv 2607.03744 v1 pith:YZ7QSJXU submitted 2026-07-04 cs.AI

classification cs.AI
keywords depressiondetectionconversationaltemporaldynamicsturn-takingmultimodallatefusionDAIC-WOZWavLMRoBERTa
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

Most automatic depression screening from clinical interviews focuses on what the participant says or how their voice sounds. This paper asks whether the timing of the exchange itself — how long each person holds the floor, how long they hesitate before answering, how much silence sits inside a turn — carries a depression signal of its own. The authors build a fixed, 24-number summary of dyadic turn-pair timing, treat it as a full modality next to frozen large acoustic and language models, and show that on the standard DAIC-WOZ interview corpus this tiny timing module is the strongest single stream on the development set. When its session scores are fused with the text model by a simple convex weight search, performance rises further, while the acoustic stream is driven to zero weight. The practical claim is that conversational timing is a lightweight, readable complement that should be treated as a first-class input for dyadic depression screening.

What carries the argument

Conversational temporal dynamics (CTD): a fixed, task-agnostic 24-feature bank of dyadic Ask/Res turn-pair timing (turn durations and ratios, voiced-versus-silence ratios, response hesitation, and silence/backchannel counts) averaged per session and classified by regularized logistic regression, then fused at the probability level with other modality scores.

What would settle it

Recompute the same CTD features using only participant-side response timing (or on interviews with a fully scripted, non-adaptive interviewer) and check whether the unimodal CTD advantage and the fusion gains over text alone disappear.

Watch

Extended reading notes

Core claim

On DAIC-WOZ, a compact 24-dimensional conversational temporal dynamics module built from Ask/Res turn-pair durations, silences, response latency, and backchannel counts is the strongest single modality on the development set (macro-F1 0.746), and convex-weighted late fusion with a frozen RoBERTa-large semantic detector reaches 0.804 development and 0.669 test macro-F1 while assigning zero weight to the acoustic stream.

Load-bearing premise

The timing features, including interviewer-side durations and silences, mainly reflect the participant's depression rather than the virtual interviewer's wizard-controlled prompt choices or other session 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

3 major / 0 minor

Summary. The paper asks whether dyadic conversational temporal dynamics (CTD)—a fixed 24-dimensional bank of Ask/Res turn-pair timing descriptors (durations, voiced/silence ratios, response latency, backchannel/silence counts)—can serve as a first-class modality for depression screening on DAIC-WOZ. Under a subject-independent, leakage-safe protocol with documented session exclusions, a compact L2-logistic CTD detector is compared to frozen WavLM-large (utterance-level acoustic probe) and frozen RoBERTa-large (turn-level semantic probe). On the development set CTD is the strongest unimodal system (macro-F1 0.746 vs. 0.690 RoBERTa and 0.673 WavLM). Convex-weighted late fusion of session probabilities reaches 0.804 / 0.669 macro-F1 on dev/test and drives the acoustic weight to exactly zero, so the deployed system reduces to RoBERTa+CTD. The authors frame the study as preliminary and discuss statistical and transfer caveats in Sec. VIII.

Significance. If the result holds beyond this cohort, treating transcript-level turn timing as an equal-footing modality is a useful, lightweight, and interpretable complement to large self-supervised encoders for dyadic clinical screening. Strengths that support taking the work seriously include: (i) reuse of a previously published, task-agnostic CTD feature bank with no DAIC-specific feature engineering; (ii) a carefully documented, subject-independent protocol with train-only statistics, published integrity exclusions, and both parameter-free mean-prob and dev-tuned convex fusion; (iii) explicit reporting of bootstrap CIs, paired prediction changes, and seed variability; and (iv) public code with seed/hyperparameter/deployed-model details. These practices raise the bar relative to much of the DAIC-WOZ literature even if the absolute gains remain small-sample and indicative.

major comments (3)
  1. Sec. IV-C / Table I and Sec. VIII: Roughly half of the 24 CTD features are Ask-side (ask_d, the six ask voiced/silence ratios, ask_bt/st, and all duration diffs/ratios involving Ask), and res_h is the Ask-end→Res-start gap. Because Ellie is wizard-controlled, these quantities can encode prompt selection, scripted pacing, or operator reaction rather than participant psychomotor retardation. The manuscript correctly flags this in Sec. VIII but never reports a response-side-only or Ask-ablated detector. Without that ablation (or an equivalent prompt-controlled check), the claim that CTD is a transferable ‘first-class modality for dyadic depression screening’ rests on a DAIC-specific artifact risk. This is load-bearing for the abstract and conclusion; a response-side-only row in Table II (or a clear statement that the result is DAIC-Ellie-specific) is needed before the interpretation can sta
  2. Sec. VII-D and Table II: Dev n=33 and test n=45 yield wide, overlapping bootstrap CIs (e.g., proposed system test macro-F1 [0.509, 0.806] vs. CTD alone [0.472, 0.771]), and the paired intervals for RoBERTa+CTD vs. each unimodal arm include or touch zero. The paper already notes this, yet the abstract and bolded claims still present 0.804 / 0.669 as an unambiguous improvement and CTD as ‘the highest single-modality performance.’ The central ordering should be stated as indicative on this split, with the more robust observations (consistent dev lift, parameter-free mean_prob[T+CTD] test 0.650, learned zero acoustic weight) foregrounded over point-estimate ranking.
  3. Sec. IV-A and Table II/III: The acoustic stream is weak on test (0.545) and receives weight 0.0 in the three-way convex solution. The authors attribute this partly to utterance- vs. turn-level pooling and to the frozen SUPERB-style probe, but they do not show that a stronger acoustic baseline (fine-tuned WavLM, different pooling, or a published competitive acoustic system under the same split) would still be down-weighted. The zero-weight conclusion is therefore specific to this probe/cohort; either strengthen the acoustic baseline or qualify the claim that ‘acoustics contributes little’ as probe-dependent rather than modality-general.

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circularity: CTD is a fixed a-priori feature bank from prior work, evaluated empirically on a new task with held-out test; fusion weights/thresholds are standard dev-tuning, not predictions forced by construction.

  1. self citation load bearing [Sec. II-C / Sec. IV-C (CTD feature adoption)]
    "we adopt a fixed, previously published dyadic timing descriptor set originally designed for a different task and corpus (Sec. II). This set is a priori and task-agnostic, as it is neither tuned nor added to or removed from based on DAIC-WOZ performance... The formulation originates in the deception-detection work of Chou et al. [13], [14]"

    The CTD feature bank is taken from prior papers co-authored by one of the present authors. This is ordinary self-citation of a reusable descriptor set; it is not load-bearing for the depression claim because the features are held completely fixed (no DAIC-WOZ tuning) and the performance numbers are new empirical measurements on a different task and corpus. Included only as the single minor self-citation that raises the score from 0 to 1.

full rationale

The paper's central claims are empirical performance numbers (CTD single-modality macro-F1, convex late-fusion gains, zero acoustic weight) obtained under a leakage-safe subject-independent protocol on DAIC-WOZ. The 24-d CTD descriptors are imported unchanged from the authors' prior deception-detection papers and are explicitly held fixed with no selection or engineering on DAIC-WOZ labels or performance; any predictive value is therefore an out-of-sample transfer result, not a tautology. Acoustic and semantic baselines are frozen SSL encoders with standard lightweight probes; fusion is ordinary score-level convex combination whose weights and decision threshold are grid-searched on the development split only and then frozen for a single train+dev refit evaluation on the official held-out test set. A parameter-free mean-probability fusion is also reported and already shows the CTD+semantics gain, confirming the result is not an artifact of weight tuning. No equation equates a fitted quantity to a claimed prediction, no uniqueness theorem is imported to forbid alternatives, and no known empirical pattern is merely renamed. The sole self-citation (Chou et al. on the CTD bank) supplies a reusable feature definition for a different task; it is not load-bearing for the depression-screening claim, which rests on the new measurements. Minor residual optimism from small-n dev tuning is acknowledged by the authors themselves and does not constitute circularity under the stated criteria. Score 1 reflects only that non-load-bearing self-citation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claim rests on standard ML and clinical-interview assumptions plus a small set of free parameters (fusion weights, thresholds, regularization) tuned on the development split. No new physical entities are invented. The load-bearing modeling choices are that the fixed Chou et al. Ask/Res descriptor bank transfers to depression and that DAIC-WOZ PHQ-8 binarization plus the cleaned official split are adequate for the comparison.

free parameters (4)
  • convex fusion weights (A/T/CTD) = [0.0, 0.3, 0.7]
    Grid-searched on development with step 0.1; the selected three-way weights (0.0, 0.3, 0.7) define the proposed system and drive the acoustic stream to zero.
  • decision thresholds per detector and fusion = WavLM 0.65; RoBERTa 0.524; CTD 0.50; fusion joint
    Chosen on development alone (grid 0.05–0.95) to maximize macro-F1; WavLM uses 0.65, RoBERTa 0.524, CTD 0.50, fusion re-tuned jointly with weights.
  • CTD logistic-regression C = 0.3
    L2 regularization strength for the 24-d session-mean classifier; set to 0.3 with class-balanced weighting.
  • probe hyperparameters (lr, dropout, layer weighting, attention size) = WavLM lr 5e-5, dropout 0.3; RoBERTa lr 1e-3, dropout 0.1
    AdamW learning rates, dropouts, and attention hidden size for the frozen WavLM and RoBERTa probes are chosen by the authors rather than exhaustively optimized; they affect the unimodal baselines against which CTD is compared.
assumptions (5)
  • domain assumption PHQ-8 ≥ 10 is a valid binary depression-screening label for the task.
    Used throughout as the ground-truth target; the paper notes these are self-report scores not clinical diagnoses.
  • domain assumption The fixed 24-feature Ask/Res timing bank from Chou et al. is task-agnostic and leakage-safe when transferred unchanged to DAIC-WOZ.
    Sec. IV-C: features are adopted a priori with no addition/removal based on DAIC performance.
  • domain assumption Excluding the ten sessions catalogued by Patapati for integrity problems does not bias the depression signal.
    Sec. III: sessions removed for desynchronization, missing utterances, label mismatch, etc.
  • ad hoc to paper Score-level convex combination of independent modality probabilities is a fair way to measure complementarity of a 24-d module against 1024-d encoders.
    Sec. V: chosen to avoid dimensionality imbalance that would swamp CTD under early fusion.
  • domain assumption Frozen WavLM-large and RoBERTa-large with lightweight probes are representative strong unimodal baselines.
    Sec. IV: baselines are intentionally standard SUPERB-style probes, not exhaustively optimized.
invented entities (1)
  • Conversational Temporal Dynamics (CTD) as a first-class modality
    purpose: To treat dyadic turn-pair timing as a standalone input stream fused on equal footing with acoustics and semantics.
    The underlying features pre-exist in Chou et al.; the paper reifies them as a named modality for depression screening rather than inventing new physics or latent variables. Independent evidence is partial (prior deception results) but not depression-specific outside this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can Conversational Temporal Dynamics Improve Depression Detection in Dyads? A Preliminary Investigation in Multi-Modality Perspectives." pith.science (2026). https://pith.science/paper/YZ7QSJXU

@misc{pith2026260703744,
  author       = {Pith},
  title        = {Pith review of: Can Conversational Temporal Dynamics Improve Depression Detection in Dyads? A Preliminary Investigation in Multi-Modality Perspectives},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YZ7QSJXU}},
  note         = {Machine review of arXiv:2607.03744}
}
read the original abstract

Automatic depression detection from clinical interviews typically models the semantic content and acoustic characteristics of participant speech. However, the interactional timing between the clinician and participant remains comparatively under-modeled. We investigate conversational temporal dynamics, specifically dyadic turn-pair timing, as a primary modality fused with self-supervised encoders. Evaluated on the DAIC-WOZ dataset, we compare a compact 24-dimensional timing module against frozen WavLM-large and RoBERTa-large baseline detectors. This temporal module achieves the highest single-modality performance on the development set. Furthermore, a convex-weighted late fusion strategy improves overall performance to 0.804 and 0.669 macro-F1 on the development and test sets, respectively. The learned fusion effectively assigns zero weight to acoustics, demonstrating that conversational timing serves as a lightweight, interpretable complement for dyadic depression screening.

Figures

Figures reproduced from arXiv: 2607.03744 by the authors.

Figure 1
Figure 1. Tri-modal pipeline. All three detectors share one Ask/Res turn segmentation of the same interview, but differ in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. One Ask/Res turn pair. For each turn we measure duration [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 4 linked inside Pith

  1. [1]

    Depression and other common mental disorders: Global health estimates,

    World Health Organization, “Depression and other common mental disorders: Global health estimates,” World Health Organization, Geneva, Switzerland, Tech. Rep. WHO/MSD/MER/2017.2, 2017. [Online]. Available: https://apps.who.int/iris/handle/10665/254610

  2. [2]

    A VEC 2013: The continuous audio/visual emotion and depression recognition challenge,

    M. Valstar, B. Schuller, K. Smith, F. Eyben, B. Jiang, S. Bilakhia, S. Schnieder, R. Cowie, and M. Pantic, “A VEC 2013: The continuous audio/visual emotion and depression recognition challenge,” inProceed- ings of the 3rd ACM International Workshop on Audio/Visual Emotion Challenge. Barcelona, Spain: ACM, 2013, pp. 3–10

  3. [3]

    A VEC 2016: Depression, mood, and emotion recognition workshop and chal- lenge,

    M. Valstar, J. Gratch, B. Schuller, F. Ringeval, D. Lalanne, M. Tor- res Torres, S. Scherer, G. Stratou, R. Cowie, and M. Pantic, “A VEC 2016: Depression, mood, and emotion recognition workshop and chal- lenge,” inProceedings of the 6th International Workshop on Au- dio/Visual Emotion Challenge. Amsterdam, The Netherlands: ACM, 2016, pp. 3–10

  4. [4]

    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. Xiao, J. Wu, L. Zhou, S. Ren, Y . Qian, Y . Qian, J. Wu, M. Zeng, X. Yu, and F. Wei, “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

  5. [5]

    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

  6. [6]

    Using speech recognition technology to investigate the association between timing- related speech features and depression severity,

    M. Yamamoto, A. Takamiya, K. Sawada, M. Yoshimura, M. Kitazawa, K.-c. Liang, T. Fujita, M. Mimura, and T. Kishimoto, “Using speech recognition technology to investigate the association between timing- related speech features and depression severity,”PLOS ONE, vol. 15, no. 9, p. e0238726, 2020

  7. [7]

    V ocal acoustic biomarkers of depression severity and treatment response,

    J. C. Mundt, A. P. V ogel, D. E. Feltner, and W. R. Lenderking, “V ocal acoustic biomarkers of depression severity and treatment response,” Biological Psychiatry, vol. 72, no. 7, pp. 580–587, 2012

  8. [8]

    A review of depression and suicide risk assessment using speech analysis,

    N. Cummins, S. Scherer, J. Krajewski, S. Schnieder, J. Epps, and T. F. Quatieri, “A review of depression and suicide risk assessment using speech analysis,”Speech Communication, vol. 71, pp. 10–49, 2015

Show all 31 references
  1. [9]

    Identifying mood episodes using dialogue features from clinical interviews,

    Z. Aldeneh, M. Jaiswal, M. Picheny, M. G. McInnis, and E. Mower Provost, “Identifying mood episodes using dialogue features from clinical interviews,” inProceedings of Interspeech 2019, Graz, Austria, 2019, pp. 1926–1930

  2. [10]

    Beyond short-frame acoustic features: Capturing long-term speech patterns for depression detection,

    S. Fushimi, M. A. Azani, M. Chiba, and Y . Okada, “Beyond short-frame acoustic features: Capturing long-term speech patterns for depression detection,”Technologies, vol. 14, no. 4, p. 198, 2026

  3. [11]

    Natural language processing methods for acoustic and landmark event-based features in speech-based depression detection,

    Z. Huang, J. Epps, D. Joachim, and V . Sethu, “Natural language processing methods for acoustic and landmark event-based features in speech-based depression detection,”IEEE Journal of Selected Topics in Signal Processing, vol. 14, no. 2, pp. 435–448, 2020

  4. [12]

    When LLMs meets acoustic landmarks: An efficient approach to integrate speech into large language models for depression detection,

    X. Zhang, H. Liu, K. Xu, Q. Zhang, D. Liu, B. Ahmed, and J. Epps, “When LLMs meets acoustic landmarks: An efficient approach to integrate speech into large language models for depression detection,” inProceedings of the 2024 Conference on Empirical Methods in Natural Language ...

  5. [13]

    Joint learning of conversational temporal dynamics and acoustic features for speech deception detection in dialog games,

    H.-C. Chou, Y .-W. Liu, and C.-C. Lee, “Joint learning of conversational temporal dynamics and acoustic features for speech deception detection in dialog games,” inProceedings of the Asia-Pacific Signal and Informa- tion Processing Association Annual Summit and Conference, Lan...

  6. [14]

    Automatic deception detec- tion using multiple speech and language communicative descriptors in dialogs,

    H.-C. Chou, Y .-W. Liu, and C.-C. Lee, “Automatic deception detec- tion using multiple speech and language communicative descriptors in dialogs,”APSIPA Transactions on Signal and Information Processing, vol. 10, p. e5, 2021

  7. [15]

    SpeechT-RAG: Reliable depression detection in LLMs with retrieval-augmented genera- tion using speech timing information,

    X. Zhang, H. Liu, Q. Zhang, B. Ahmed, and J. Epps, “SpeechT-RAG: Reliable depression detection in LLMs with retrieval-augmented genera- tion using speech timing information,”arXiv preprint arXiv:2502.10950, 2025

  8. [16]

    The distress analysis interview corpus of human and computer interviews,

    J. Gratch, R. Artstein, G. Lucas, G. Stratou, S. Scherer, A. Nazarian, R. Wood, J. Boberg, D. DeVault, S. Marsella, D. Traum, S. Rizzo, and L.-P. Morency, “The distress analysis interview corpus of human and computer interviews,” inProceedings of the Ninth International Confer...

  9. [17]

    Detecting depression with audio/text sequence modeling of interviews,

    T. Al Hanai, M. M. Ghassemi, and J. R. Glass, “Detecting depression with audio/text sequence modeling of interviews,” inProceedings of Interspeech 2018, Hyderabad, India, 2018, pp. 1716–1720

  10. [18]

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

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” in Advances in Neural Information Processing Systems, vol. 33, 2020, pp. 12 449–12 460

  11. [19]

    SUPERB: Speech processing universal PERformance benchmark,

    S.-w. Yang, P.-H. Chi, Y .-S. Chuang, C.-I. J. Lai, K. Lakhotia, Y . Y . Lin, A. T. Liu, J. Shi, X. Chang, G.-T. Lin, T.-H. Huang, W.-C. Tseng, K.-t. Lee, D.-R. Liu, Z. Huang, S. Dong, S.-W. Li, S. Watanabe, A. Mohamed, and H.-y. Lee, “SUPERB: Speech processing universal PERfo...

  12. [20]

    Topic modeling based multi-modal depression detection,

    Y . Gong and C. Poellabauer, “Topic modeling based multi-modal depression detection,” inProceedings of the 7th Annual Workshop on Audio/Visual Emotion Challenge. Mountain View, CA, USA: ACM, 2017, pp. 69–76

  13. [21]

    Automatic depression detection: An emotional audio-textual corpus and a GRU/BiLSTM-based model,

    Y . Shen, H. Yang, and L. Lin, “Automatic depression detection: An emotional audio-textual corpus and a GRU/BiLSTM-based model,” in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, Singapore, 2022, pp. 6247–6251

  14. [22]

    Self-supervised representations in speech-based depression detection,

    W. Wu, C. Zhang, and P. C. Woodland, “Self-supervised representations in speech-based depression detection,” inProceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, 2023

  15. [23]

    Hierarchical self-supervised repre- sentation learning for depression detection from speech,

    Y . Li, E. S. Chng, and C. Guan, “Hierarchical self-supervised repre- sentation learning for depression detection from speech,”arXiv preprint arXiv:2510.08593, 2025

  16. [24]

    Test-time training for speech-based depression detection,

    S. H. Dumpala, C. S. Sastry, R. Uher, and S. Oore, “Test-time training for speech-based depression detection,” inProceedings of Interspeech 2025, 2025, pp. 479–483

  17. [25]

    Domain adversarial training for mitigating gender bias in speech-based mental health detection,

    J.-W. Kim, H. Yoon, W. Oh, D. Jung, S.-H. Yoon, D.-J. Kim, D.-H. Lee, S.-Y . Lee, and C.-M. Yang, “Domain adversarial training for mitigating gender bias in speech-based mental health detection,”arXiv preprint arXiv:2505.03359, 2025

  18. [26]

    The PHQ-8 as a measure of current depression in the general population,

    K. Kroenke, T. W. Strine, R. L. Spitzer, J. B. W. Williams, J. T. Berry, and A. H. Mokdad, “The PHQ-8 as a measure of current depression in the general population,”Journal of Affective Disorders, vol. 114, no. 1–3, pp. 163–173, 2009

  19. [27]

    Common pitfalls and recommendations for use of machine learning in depression severity estimation: DAIC-WOZ study,

    I. Danylenko and O. Unold, “Common pitfalls and recommendations for use of machine learning in depression severity estimation: DAIC-WOZ study,”Applied Sciences, vol. 16, no. 1, p. 422, 2026

  20. [28]

    Integrating large language models into a tri-modal architecture for automated depression classification on the DAIC-WOZ,

    S. V . Patapati, “Integrating large language models into a tri-modal architecture for automated depression classification on the DAIC-WOZ,” arXiv preprint arXiv:2407.19340, 2024

  21. [29]

    DAIC-WOZ: On the validity of using the therapist’s prompts in automatic depression detection from clinical interviews,

    S. Burdisso, E. Reyes-Ram ´ırez, E. Villatoro-tello, F. S ´anchez-Vega, A. Lopez Monroy, and P. Motlicek, “DAIC-WOZ: On the validity of using the therapist’s prompts in automatic depression detection from clinical interviews,” inProceedings of the 6th Clinical Natural Language...

  22. [30]

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

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

  23. [31]

    Confidence intervals for evaluation in ma- chine learning,

    L. Ferrer and P. Riera, “Confidence intervals for evaluation in ma- chine learning,” https://github.com/luferrer/ConfidenceIntervals, 2024, computer software

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.