Pith. sign in

REVIEW 4 major objections 6 minor 29 references

Aggregation methods for speech-based depression detection cannot be ranked by average accuracy on one model: a third of setups collapse to always predicting the same class, and rankings flip with backbone and seed.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 06:13 UTC pith:2FKZXMUE

load-bearing objection Solid multi-backbone aggregation benchmark that makes collapse and seed fragility visible; test-set checkpointing and tiny speaker counts weaken the rankings but do not invent the core finding. the 4 major comments →

arxiv 2607.02904 v1 pith:2FKZXMUE submitted 2026-07-03 eess.AS

Speaker-Aware Temporal Aggregation Strategies on Segment Representations for Depression Detection in Dyadic Interaction: A Benchmark Study

Anisha Pattanayak , Huang-Cheng Chou , Shrikanth Narayanan , Sudarsana Reddy Kadiri This is my paper
classification eess.AS
keywords depression detectiontemporal aggregationpoolingclinical speechself-supervised learningspeaker-level evaluationbenchmark robustness
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper asks whether the ranking of methods that turn short speech clips into one speaker-level depression decision is real, or an artifact of the usual fixed-encoder, hand-picked-layer pipeline. It builds DEPOOL: six aggregation architectures crossed with six frozen self-supervised speech backbones on an English and a Mandarin depression corpus, with each setup learning which encoder layers matter instead of fixing one by hand. Across the 72 configurations, roughly one third collapse into predicting a single class for every speaker, a failure linked to the backbone at least as much as to the architecture. The head that looks most stable under a single seed fails when seeds are varied. The authors conclude that robustness across backbones and seeds, not average accuracy on one pipeline, should be the first-class criterion for benchmarking temporal aggregation in clinical speech.

Core claim

In a controlled 72-cell grid of six temporal aggregation architectures, six frozen self-supervised speech backbones, and two depression corpora, about one third of configurations collapse into single-class prediction for every test speaker. Collapse is concentrated in particular architecture-backbone pairs and is at least as backbone-driven as architecture-driven; the architecture that never collapses under a single seed becomes unreliable under seed replication. Aggregation architecture therefore cannot be evaluated in isolation from its backbone or random seed.

What carries the argument

DEPOOL: the cross-product of six aggregation heads (mean pooling, statistical pooling, self-attention, bidirectional GRU with attention, NetVLAD, Transformer encoder) with six frozen SSL backbones on E-DAIC and MODMA, using learned softmax layer weights so every head sees a controlled 256-dimensional clip embedding.

Load-bearing premise

The collapse rates and architecture rankings remain trustworthy even though the main grid uses one fixed seed and picks the best-F1 checkpoint on the tiny held-out test sets rather than a pure validation criterion.

What would settle it

A full multi-seed sweep of all 72 cells with strict validation-only checkpointing that yields near-zero collapse rates and stable architecture rankings across seeds and backbones would falsify the claim that backbone-and-seed robustness must be first-class.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Benchmarks reporting only mean accuracy for one head on one backbone can hide widespread collapse and produce misleading rankings.
  • Backbone choice can dominate architecture choice; some backbones collapse on most heads.
  • Single-seed stability does not guarantee multi-seed stability, so seed replication is required before recommending any head as default.
  • Comparative claims about aggregation for clinical speech should report collapse rates and seed variance alongside average metrics.
  • Future aggregation studies should adopt multi-backbone, multi-seed protocols with speaker-independent splits.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same collapse and seed-sensitivity pattern may appear in other low-data paralinguistic tasks that freeze SSL backbones on small speaker sets.
  • Residual confounds from ASR fine-tuning of some checkpoints could still shape the collapse map even with learned layer weights.
  • Utility scores that overweight sensitivity will favor heads that over-predict the depressed class, amplifying differences that are collapse artifacts.
  • Extending the grid to more languages or within-clip temporal encodings would test whether backbone-driven collapse generalizes beyond the two corpora.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper introduces DEPOOL, a controlled 72-cell benchmark of six temporal aggregation architectures (mean, statistical, self-attention, Bi-GRU+attention, NetVLAD, Transformer encoder) crossed with six frozen SSL speech backbones on E-DAIC (English) and MODMA (Mandarin). Layer choice is replaced by a learned softmax over all hidden layers (Eq. 1), with speaker-independent stratified splits and a clinical utility score U. The central empirical claim is that roughly one third of single-seed configurations collapse to single-class prediction (sens/spec = (1,0) or (0,1)), that collapse is at least as backbone-dependent as architecture-dependent (Tables I–II, Figs. 2–3), and that Bi-GRU+attention’s apparent single-seed stability fails multi-seed replication on a focal subset (Table III); therefore robustness to backbone and seed should be a first-class criterion for temporal aggregation in clinical speech.

Significance. If the robustness finding holds under cleaner evaluation, the paper supplies a useful methodological corrective for clinical speech ML: single-pipeline aggregation comparisons are under-powered and can mis-rank methods. Strengths include the full architecture×backbone×corpus grid, SUPERB-style learned layer featurization that removes hand-picked-layer confounds, strictly speaker-independent splits, dual-language corpora, an explicit collapse operationalization, and an open release of pipeline, splits, and results. These are concrete contributions even if absolute metrics are noisy. The work is more a carefully designed empirical warning than a new detection SOTA, but that warning is timely for the field.

major comments (4)
  1. [§III-G, §VI] §III-G and Limitations §VI: Checkpoint selection tracks best-F1 on the held-out test partition rather than validation. With class-weighted CE, few training sequences, and test sets of only 23 (E-DAIC) and 10 (MODMA) speakers, this protocol can preferentially retain degenerate constant-label solutions whenever they briefly maximize test F1. Collapse rates in Figs. 2–3 and architecture rankings in Tables I–II are therefore not cleanly separable from the evaluation rule itself. Comparative claims require re-running the grid (or at least the collapse-prone cells) with validation-only checkpointing and reporting how collapse counts and rankings change.
  2. [§IV-D, Table III] §IV-A–C vs §IV-D / Table III: The headline 33% collapse rate and the claim that Bi-GRU+attention “never collapses” rest on a single fixed seed for all 72 cells. Multi-seed evidence covers only four architecture/backbone pairs per corpus (Table III), where Bi-GRU F1 std reaches 0.42 and the single-seed stability result fails. The paper’s own conclusion that seed robustness is first-class is undercut by not extending multi-seed replication to the full grid (or a stratified sample of all architecture×backbone cells) before ranking methods by collapse.
  3. [§III-A, Tables I–II, Figs. 2–3] §III-A, §III-G, §V-D: Held-out sets are 23 and 10 speakers. One misclassified participant moves accuracy by ~4–10 points; collapse is defined at the speaker-prediction level on these tiny sets. Mean metrics averaged over six backbones (Tables I–II) and collapse fractions of 12 runs (Figs. 2–3) therefore have high participant-level variance. The manuscript should report binomial/bootstrap uncertainty on collapse rates and avoid treating point estimates of Acc/F1/U as stable architecture rankings without that uncertainty.
  4. [§III-C, Table II, Fig. 3] §III-C / Appendix Table IV: Two of six backbones are ASR-fine-tuned (HuBERT-Large, Data2Vec-Audio-Large) while others are purely SSL; Wav2Vec2-Robust’s domain-invariance objective is also distinct. The paper notes this confound but still attributes collapse primarily to “backbone” identity (Fig. 3, 83% for Wav2Vec2-Robust). A controlled re-analysis separating SSL-only vs ASR-FT checkpoints (or reporting layer-weight distributions by backbone type) is needed before the backbone-dominance claim is load-bearing for the robustness criterion.
minor comments (6)
  1. [§III-E] §III-E: NetVLAD uses K=2 clusters with a 64-dim bottleneck; Transformer uses a single layer, 4 heads, 64-dim bottleneck. These capacity choices are under-motivated relative to the low-data regime discussion in §V-B and may themselves drive collapse; a short ablation or justification would help.
  2. [Table I] Table I caption: “macro or positive-class F1” is ambiguous; state which F1 is reported in each corpus block.
  3. [§III-G, Eq. (3)] Eq. (3): Clinical utility U = (2·Sens + Spec)/3 is reasonable but ad hoc; cite prior clinical-speech use or note it as a paper-specific weighted score so readers do not treat it as a standard metric.
  4. [Figs. 2–3] Fig. 1 pipeline diagram is clear; Figs. 2–3 would benefit from absolute counts (n collapsed / 12) next to rates and error bars if multi-seed data exist for any cells.
  5. [§II-B] §II-B: Attentive statistics pooling and NetVLAD are cited for speaker embeddings; a one-sentence note on why attentive stats was not included as a seventh head would close a natural gap.
  6. [Introduction, §III-F] Minor wording: “DEPOOLand” missing space (Introduction); “F . Speaker-Independent” stray space in heading III-F.

Circularity Check

0 steps flagged

No significant circularity: empirical benchmark reports trained metrics and operational collapse flags without reducing claims to fitted inputs or self-definitional identities.

full rationale

DEPOOL is a controlled empirical grid (6 architectures × 6 frozen SSL backbones × 2 corpora) that trains aggregation heads and learned layer weights on speaker-independent train splits, selects checkpoints, and reports standard classification metrics plus an explicitly defined utility U = (2·Sensitivity + Specificity)/3 on held-out speakers. Collapse is an operational flag—(sensitivity, specificity) = (1,0) or (0,1)—not a tautology derived from the loss or architecture equations. No first-principles derivation, uniqueness theorem, or ansatz is invoked whose conclusion is forced by construction from the paper’s own inputs or from load-bearing self-citations; citations are to external datasets, SSL models, and prior benchmarks (SUPERB, EMO-SUPERB). Methodological concerns such as test-set checkpointing or small speaker counts affect reliability of the rankings but do not constitute circular reduction of a claimed prediction to its defining fit. The paper is therefore self-contained against external benchmarks with score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

The central robustness claim rests on standard ML training choices, clinical-speech domain assumptions, and a few paper-specific operational definitions (collapse flag, utility U, fixed windowing). No new physical entities are postulated. Free parameters are ordinary hyperparameters and design choices that can change absolute metrics and, in principle, collapse rates; the paper does not claim parameter-free universality.

free parameters (6)
  • AdamW learning rate and weight decay
    Fixed at 1e-3 and 1e-4 for all 72 configs; optimization landscape and collapse risk can depend on these choices.
  • Training epochs and best-F1 checkpoint rule
    40 epochs with best-F1 selection; Limitations admit selection used the test partition, which can inflate absolute metrics and affect which runs look collapsed.
  • Clip window 3 s with 1.5 s stride
    Hand-chosen segmentation that defines the sequence each head sees; different windows could change temporal structure available to aggregation.
  • Projection dimension 256 and head-specific sizes (NetVLAD K=2, Transformer 64-dim/4-head, GRU 128/dir)
    Architecture capacities chosen by authors; under-capacity routing is part of their collapse hypothesis for Transformer/NetVLAD.
  • PHQ-8 binarization threshold ≥10 (E-DAIC)
    Label definition for the English corpus; changes class balance and what 'depression' means for evaluation.
  • Clinical utility U = (2·Sensitivity + Specificity)/3
    Paper-defined weighted score used in tables; not a free fit to data but an ad hoc clinical preference weight.
axioms (5)
  • domain assumption Frozen SSL speech representations contain depression-relevant paralinguistic information that can be recovered by a small trained head.
    Assumed throughout §III–IV; without it the grid is uninformative rather than a test of aggregation.
  • domain assumption Speaker-independent stratified group splits prevent identity leakage from dominating depression labels.
    §III-F; standard clinical-speech practice but load-bearing for interpreting speaker-level metrics as depression rather than speaker ID.
  • ad hoc to paper A run with (sensitivity, specificity) = (1,0) or (0,1) is a meaningful 'collapse' failure of temporal aggregation.
    §III-G operational definition driving the one-third headline; alternative thresholds or soft-confidence criteria could reclassify borderline runs.
  • domain assumption Learned softmax layer weights remove hand-picked-layer selection as a confound across backbones.
    §III-D and contribution list; mitigates but does not remove ASR-fine-tuning confounds the authors themselves note.
  • standard math Standard optimization and evaluation math (AdamW, cross-entropy, ROC-AUC, StratifiedGroupKFold) behave as in the scikit-learn / deep-learning literature.
    Background tools cited via Pedregosa et al. and common practice; not re-proved.
invented entities (2)
  • DEPOOL benchmark (72-cell architecture × backbone × corpus grid with semi-fine-tuned layer aggregation) no independent evidence
    purpose: Provide a controlled comparison of temporal aggregation methods for depression detection without fixing one backbone or layer.
    Named protocol introduced by the paper; independent evidence is the released splits/results rather than external prior existence.
  • Clinical utility score U no independent evidence
    purpose: Summarize sensitivity/specificity with double weight on sensitivity for clinical cost asymmetry.
    Defined in Eq. (3); a paper-specific scalar, not an independently validated clinical instrument.

pith-pipeline@v1.1.0-grok45 · 17918 in / 3823 out tokens · 41350 ms · 2026-07-12T06:13:40.401921+00:00 · methodology

0 comments
read the original abstract

Speech-based depression detection compresses features from short audio segments into one speaker-level decision, a step called temporal aggregation rarely studied on its own. Most benchmarks fix a single self-supervised encoder and a single hand-picked layer, so a reported gain may reflect the pipeline rather than the aggregation method itself. We introduce DEPOOL, a controlled benchmark that compares six aggregation architectures with six frozen speech backbones on an English and a Mandarin depression corpus, where each configuration learns which backbone layers matter rather than fixing one by hand. Across the resulting 72-configuration grid, a third of configurations collapse into predicting a single class for every speaker, a failure tied to the backbone as much as to the method, and the architecture that is most stable in a single-seed run becomes unreliable when training repeats across seeds. Robustness to backbone and seed, rather than average accuracy across a single pipeline, should be a first-class benchmarking criterion for temporal aggregation in clinical speech.

Figures

Figures reproduced from arXiv: 2607.02904 by Anisha Pattanayak, Huang-Cheng Chou, Shrikanth Narayanan, Sudarsana Reddy Kadiri.

Figure 1
Figure 1. Figure 1: DEPOOL pipeline: Components are color-coded by training status. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Share of runs (of 12: 6 backbones × 2 corpora) in which an architecture collapses to a single class, under single-seed training. Bi-GRU with Attention never collapses here, but Section IV-D shows this does not survive seed replication. its 6 architecture pairings per corpus: the clearest single signal that backbone choice can dominate architecture choice. B. Best Single Configurations On the E-DAIC, the be… view at source ↗
Figure 4
Figure 4. Figure 4: Seed replication on the focal subset of Table III (ED = E-DAIC, [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 6 linked inside Pith

  1. [1]

    Depressive disorder (depression),

    World Health Organization, “Depressive disorder (depression),” WHO Fact Sheet, Sep. 2023

    2023

  2. [2]

    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

    2015

  3. [3]

    Automated assessment of psychiatric disorders using speech: A systematic review,

    D. M. Low, D. Bentley, and S. Ghosh, “Automated assessment of psychiatric disorders using speech: A systematic review,”Laryngoscope Investigative Otolaryngology, vol. 5, no. 1, pp. 96–116, 2020

    2020

  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. Li, and F. Wei, “WavLM: Large-scale self-supervised pre-training for full stack speech processing,”IEEE J. Sel. Topics Signal Process., vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  5. [5]

    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,” inProc. LREC 2014, 2014

    2014

  6. [6]

    MODMA dataset: A multi-modal open dataset for mental-disorder analysis,

    H. Cai, Z. Yuan, Y . Gao, S. Sun, N. Li, F. Tian, H. Xiao, J. Li, Z. Yang, X. Li, Q. Zhu, Z. Xiang, and N. Sun, “MODMA dataset: A multi-modal open dataset for mental-disorder analysis,”arXiv preprint arXiv:2002.09283, 2020

    Pith/arXiv arXiv 2002

  7. [7]

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

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,”arXiv preprint arXiv:2106.07447, 2021

    Pith/arXiv arXiv 2021

  8. [8]

    Robust wav2vec 2.0: Analyzing domain shift in self-supervised pre-training,

    W.-N. Hsu, A. Sriram, A. Baevski, T. Likhomanenko, Q. Xu, V . Pratap, J. Kahn, A. Lee, R. Collobert, G. Synnaeve, and M. Auli, “Robust wav2vec 2.0: Analyzing domain shift in self-supervised pre-training,” inProc. Interspeech 2021, pp. 721–725, 2021

    2021

  9. [9]

    data2vec: A general framework for self-supervised learning in speech, vision and language,

    A. Baevski, W.-N. Hsu, Q. Xu, A. Babu, J. Gu, and M. Auli, “data2vec: A general framework for self-supervised learning in speech, vision and language,” inProc. ICML 2022, pp. 1298–1312, 2022

    2022

  10. [10]

    XLS-R: Self-supervised cross-lingual speech representation learning at scale,

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. von Platen, Y . Saraf, J. Pino, A. Baevski, A. Conneau, and M. Auli, “XLS-R: Self-supervised cross-lingual speech representation learning at scale,”arXiv preprint arXiv:2111.09296, 2021

    Pith/arXiv arXiv 2021

  11. [11]

    Inves- tigation of layer-wise speech representations in self-supervised learning models: A cross-lingual study in detecting depression,

    B. Maji, R. Guha, A. Routray, S. Nasreen, and D. Majumdar, “Inves- tigation of layer-wise speech representations in self-supervised learning models: A cross-lingual study in detecting depression,” inProc. Inter- speech 2024, pp. 3020–3024, 2024

    2024

  12. [12]

    SUPERB: Speech processing universal performance benchmark,

    S.-W. Yang, P.-H. Chi, Y .-S. Chuang, C.-I. J. Lai, K. Qian, P. Hira, and J. Glass, “SUPERB: Speech processing universal performance benchmark,” inProc. Interspeech 2021, pp. 1194–1198, 2021

    2021

  13. [13]

    EMO-SUPERB: An in-depth look at speech emotion recognition,

    H. Wu, H.-C. Chou, K.-W. Chang, L. Goncalves, J. Du, J.-S. R. Jang, C.-C. Lee, and H.-y. Lee, “EMO-SUPERB: An in-depth look at speech emotion recognition,”arXiv preprint arXiv:2402.13018, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Cross- lingual speech emotion recognition: Humans vs. self-supervised mod- els,

    Z. Han, T. Geng, H. Feng, J. Yuan, K. Richmond, and Y . Li, “Cross- lingual speech emotion recognition: Humans vs. self-supervised mod- els,”arXiv preprint arXiv:2409.16920, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    Association between acoustic speech features and non-severe levels of anxiety and depression symptoms across lifespan,

    L. Albuquerque, A. R. S. Valente, A. Teixeira, D. Figueiredo, P. Sa- Couto, and C. Oliveira, “Association between acoustic speech features and non-severe levels of anxiety and depression symptoms across lifespan,”PLOS ONE, vol. 16, no. 4, p. e0248842, 2021

    2021

  16. [16]

    Detecting depression with audio/text sequence modeling of interviews,

    T. Alhanai, M. Ghassemi, and J. Glass, “Detecting depression with audio/text sequence modeling of interviews,” inProc. Interspeech 2018, pp. 1716–1720, 2018

    2018

  17. [17]

    DEPA: Self-supervised audio embedding for depression detection,

    P. Zhang, M. Wu, H. Dinkel, and K. Yu, “DEPA: Self-supervised audio embedding for depression detection,” inProc. ACM MM 2021, pp. 135– 143, 2021

    2021

  18. [18]

    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

    Pith/arXiv arXiv 2021

  19. [19]

    Tracking depression-related mental state using multimodal analysis,

    J. R. Williamson, T. F. Quatieri, B. S. Helfer, J. Perricone, S. S. Ghosh, and G. Ciccarelli, “Tracking depression-related mental state using multimodal analysis,” inProc. ACM ICMI 2013, pp. 279–286, 2013

    2013

  20. [20]

    A VEC 2019 Workshop and Challenge: State-of-Mind, Detecting Depression with AI, and Cross-Cultural Affect Recognition,

    F. Ringeval, B. Schuller, M. Valstar, N. Cummins, R. Cowie, L. Tavabi, M. Schmitt, S. Alisamir, S. Amiriparian, E.-M. Messner, S. Song, S. Liu, Z. Zhao, A. Mallol-Ragolta, Z. Ren, M. Soleymani, and M. Pantic, “A VEC 2019 Workshop and Challenge: State-of-Mind, Detecting Depression with AI, and Cross-Cultural Affect Recognition,” inProc. ACM MM AVEC 2019, p...

    2019

  21. [21]

    Speaker normalization for speech-based depression detection,

    Y . F. A. Gaus, S. Ballard, N. Cummins, and B. Schuller, “Speaker normalization for speech-based depression detection,” inProc. ICASSP 2021, pp. 7278–7282, 2021

    2021

  22. [22]

    DepAudioNet: An efficient deep model for audio based depression classification,

    X. Ma, H. Yang, Q. Chen, D. Huang, and Y . Wang, “DepAudioNet: An efficient deep model for audio based depression classification,” inProc. ACM AVEC 2016, pp. 35–42, 2016

    2016

  23. [23]

    Automatic speech emotion recognition using recurrent neural networks with local attention,

    S. Mirsamadi, E. Barsoum, and C. Zhang, “Automatic speech emotion recognition using recurrent neural networks with local attention,” in Proc. ICASSP 2017, pp. 2227–2231, 2017

    2017

  24. [24]

    NetVLAD: CNN architecture for weakly supervised place recognition,

    R. Arandjelovic, P. Gronat, A. Torii, T. Pajdla, and J. Sivic, “NetVLAD: CNN architecture for weakly supervised place recognition,” inProc. CVPR 2016, pp. 5297–5307, 2016

    2016

  25. [25]

    Attentive statistics pooling for deep speaker embedding,

    K. Okabe, T. Koshinaka, and K. Shinoda, “Attentive statistics pooling for deep speaker embedding,” inProc. Interspeech 2018, pp. 2252–2256, 2018

    2018

  26. [26]

    BERT: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidirectional transformers for language understanding,” inProc. NAACL 2019, pp. 4171–4186, 2019

    2019

  27. [27]

    The INTERSPEECH 2021 computational paralinguistics challenge,

    B. W. Schuller, A. Batliner, C. Bergler, C. Mascolo, J. Han, I. Lefter, H. Kaya, S. Amiriparian, A. Baird, L. Stappen, S. Ottl, M. Gerczuk, P. Tzirakis, C. Brown, J. Chauhan, A. Grammenos, A. Hasthanasombat, D. Spathis, T. Xia, P. Cicuta, et al., “The INTERSPEECH 2021 computational paralinguistics challenge,” inProc. Interspeech 2021, pp. 431–435, 2021

    2021

  28. [28]

    Prevalence of neural collapse during the terminal phase of deep learning training,

    V . Papyan, X. Y . Han, and D. L. Donoho, “Prevalence of neural collapse during the terminal phase of deep learning training,”Proc. Natl. Acad. Sci., vol. 117, no. 40, pp. 24652–24663, 2020

    2020

  29. [29]

    Scikit-learn: Machine learning in Python,

    F. Pedregosa, G. Varoquaux, A. Gramfort, V . Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V . Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay, “Scikit-learn: Machine learning in Python,”J. Mach. Learn. Res., vol. 12, pp. 2825–2830, 2011. APPENDIX Table IV summarizes the six self-supervised...

    2011