Pith. sign in

REVIEW 5 major objections 5 minor 40 references

Recognizing Dementia from Neuropsychological Tests with State Space Models

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A state-space audio model outscores prior dementia screening by 21% relative AUC, the paper reports, by reading hour-long neuropsychological interviews in a single pass.

desk verdict The SSM-for-dementia application is plausible and the paper has useful analyses, but the headline 21% claim is not statistically supported because the test set is tiny and the authors select their best checkpoints and fusion weight on it. read the letter →

arxiv 2507.10311 v1 pith:FWZK5ATT submitted 2025-07-14 cs.LG cs.AI

Liming Wang , Saurabhchand Bhati , Cody Karjadi , Rhoda Au , James Glass This is my paper · ORCID
classification cs.LGcs.AI
keywords dementiaclassificationstatespacemodelsMambaneuropsychologicaltestsspeechbiomarkersaudio-textfusionFraminghamHeartStudylong-sequencemodeling
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 tries to establish that a state-space-model audio classifier, Demenba, can classify dementia from hour-long neuropsychological test recordings more accurately than the previous best approach, while using comparable or fewer parameters. Trained on over 1,000 hours of Framingham Heart Study recordings labeled normal, mild cognitive impairment, or dementia, Demenba reports 3-class AUC 0.83 with speaker-diarization boundaries versus 0.69 for an EfficientNet b6 baseline, a 21% relative gain, and reaches 0.95 AUC when its audio scores are fused with a fine-tuned BERT transcript classifier. A sympathetic reader would care because the system reads long interviews in one pass, which prior models could not do without fragmenting context, and because improved fine-grained separation of MCI from dementia is clinically the harder and more useful distinction.

What carries the argument

The central object is the selective state-space model (Mamba/VMamba) encoder applied to 128-bin mel-filterbank frames of up to 360-second segments, with four Mamba layer groups each followed by downsampling by two and mean-pooling to class logits. Around it sit three supporting mechanisms: speaker-diarization boundaries that preserve turn-taking and prosodic continuity, selective majority voting that keeps the top-k segments by peak probability rather than averaging all segments, and late fusion $p_{\text{audio+text}} = (1-\lambda)p_{\text{audio}} + \lambda p_{\text{text}}$ with a tuned weight. The argument that long context is what matters rests on segment-duration curves showing AUC gains from 30s to 180s and diminishing returns beyond.

What would settle it

Run Demenba-medium and EfficientNet b6 on a held-out test set of at least 100 recordings per class that is never used for checkpoint selection, segment-count tuning, or fusion-weight selection, and compare 3-class speaker-diarization AUC; if the gap shrinks to within confidence-interval overlap, the 21% relative gain is an artifact of test-set selection.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that linear-complexity selective state-space models overcome the context fragmentation that limited prior dementia classifiers: using Mamba/VMamba backbones on mel-filterbank segments up to 360 seconds, with speaker-diarization boundaries and selective majority voting over the most informative segments, Demenba-medium achieves 0.83 AUC in 3-class normal/MCI/dementia classification versus 0.69 for the EfficientNet b6 baseline, and 0.87 versus 0.82 in 2-class impaired detection. The paper also reports that fusing audio probabilities with BERT text probabilities yields 0.95 AUC, that including silence improves 2-class discrimination by 10-15% absolute AUC, and that interviewer speech alone carries diagnostic signal. The scaling experiments show AUC still rising with training hours for the medium model and saturating beyond roughly 180 seconds of segment duration.

Load-bearing premise

The reported advantage rests on test AUCs computed from just ten recordings per class, with the best checkpoints and fusion weights chosen using that same test set.

Editorial extensions

If this is right

  • An hour-long neuropsychological interview can be classified in one forward pass, so automatic dementia screening no longer depends on manual sentence-level segmentation or forced alignment.
  • Fine-grained 3-class classification (normal versus MCI versus dementia) improves more than coarse 2-class detection, which matters for catching early decline.
  • Audio and text carry complementary evidence; fusing prosodic cues with lexical cues gives the single best reported result of 0.95 AUC.
  • Scaling behavior suggests that collecting more labeled hours would continue to improve the medium model, giving a concrete data-collection target.
  • Interviewer speech contributing to the decision implies interviewers react measurably to participants, so standardizing interviews could matter for fair screening.

Reading between the lines

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

  • If the 21% gain is caused by selective majority voting and long segments rather than by the SSM architecture itself, then applying the same voting and 360-second segments to the EfficientNet baseline would isolate the source; the paper does not report that control.
  • The reported test AUCs come from best checkpoints selected on a test set of only ten recordings per class, with the fusion weight swept on the same test set; a pre-registered held-out evaluation with more recordings would be needed before trusting the 0.83 and 0.95 numbers as deployment-level.
  • The finding that silence boosts 2-class but not 3-class accuracy suggests silence features track dementia severity extremes; a testable extension is to use pause and silence duration statistics as explicit input features to see whether they mediate the gain.
  • Because the Framingham recordings come from one longitudinal cohort, the method's portability to other dialects, recording devices, and test protocols is unknown; cross-dataset evaluation would settle how general the SSM advantage is.
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 / 5 minor

Summary. The paper proposes Demenba, an automatic dementia classification system based on state-space models (SSMs) that processes long segments of audio from hour-long neuropsychological interviews, optionally fusing with text-based classifiers. The authors report that Demenba outperforms an EfficientNet baseline by up to 21% relative AUC in 3-class dementia classification on the Framingham Heart Study dataset, while also presenting scaling experiments, ablation studies, and an evaluation on a separate 92-recording subset. The main claimed contributions are the SSM-based long-context audio model, the improved fine-grained classification results, and the analysis of segment types and modality fusion.

Significance. If the empirical claims were supported by a sound evaluation, the paper would be a meaningful step in automatic dementia classification: it trains on over 1,000 hours of data, uses a controlled feature and optimization setup, includes scaling analyses, checks for keyword leakage, and provides an anonymous code release. However, the central quantitative claim is not currently supported because the test set is used for model and hyperparameter selection, the test set is very small, and no uncertainty quantification is provided. The architectural idea is plausible and interesting, but the evidence as presented does not establish the claimed advantage over prior work.

major comments (5)
  1. [Section IV-A, Tables II-IV] The manuscript explicitly reports results selected on the test set: Section IV-A states 'Results from the best models on the test set are reported', and Table IV's caption says 'best checkpoints from the test set'. Because the test set is used for checkpoint selection, the AUCs in Tables II and III are not unbiased estimates of generalization, and the abstract's 21% relative improvement in 3-class classification rests on this biased protocol. The authors must select checkpoints and hyperparameters on a held-out validation set and report test-set results only once.
  2. [Fig. 5, Eq. (1), Table III] The fusion weight lambda in Eq. (1) is swept over the test set in Fig. 5, whose y-axis is labeled 'Test AUC', and the 0.95 audio+text AUC in Table III appears to be the peak of this sweep. This constitutes hyperparameter tuning on the test set and adds another layer of optimistic bias to the headline fusion result. The fusion weight should be selected on a validation set or reported as a curve with a validation-based choice.
  3. [Table I, Section IV-A] The test sets are extremely small: the 3-class test set has 10 recordings per class (30 total) and the 2-class test set has 20 total, yet no confidence intervals, bootstrap estimates, or significance tests are reported. With 30 test recordings, the difference between 0.83 and 0.69 AUC (the basis of the 21% relative claim) is well within sampling variability; the claim is not statistically supported. The authors should enlarge the test set, or at minimum report bootstrap/DeLong confidence intervals and paired significance tests.
  4. [Table I] Table I shows test sets with more recordings than participants (3-class: 18 participants but 30 recordings; 2-class: 10 participants but 20 recordings). If any participant contributes multiple recordings, the recordings are not independent, which further reduces the effective sample size and can bias AUC estimates. The authors should report the number of unique participants per class and use subject-level clustering or analysis.
  5. [Abstract, Table II] The abstract's claim of 'outperforming prior approaches by 21% while using fewer parameters' is inconsistent with Table II: the 21% relative improvement is achieved by Demenba-medium (48M parameters) against EfficientNet b6 (40M parameters), so Demenba-medium does not use fewer parameters. Demenba-small (29M) does use fewer parameters, but its 3-class SD AUC is 0.81, which is a 17% relative improvement, not 21%. The claim should be corrected and the specific model and comparison should be stated.
minor comments (5)
  1. [Section II, Section IV-A] There are typos: 'Fragmingham' should be 'Framingham', and 'occurences' should be 'occurrences'.
  2. [Fig. 2b] The 'Best AUC' plotted as the highest AUC for all segment durations up to the current value is an optimistic cumulative statistic; plotting the actual AUC at each segment duration would be more informative and interpretable.
  3. [Table II caption] The caption 'only AUCs for the better of BCE and CE losses are shown' indicates selection without specifying the selection criterion; both losses should be reported or the selection should be explicitly based on a validation set.
  4. [Tables II and IV] The eval-set results in Table IV use gold speaker boundaries and different segment lengths, while Table II uses predicted SD or VAD boundaries; the two tables are not directly comparable, so the text should not imply a uniform comparison across them.
  5. [References] References [10] and [14] are the same paper (arXiv and ICLR versions) and should be consolidated.

Circularity Check

3 steps flagged · score 6.0 of 10

Reported 21% relative-AUC improvement is obtained by selecting best checkpoints and fusion weights on the test set, so the headline numbers are fitted rather than predicted.

  1. fitted input called prediction [Section IV-A ('Dataset and training settings') and Table II]
    "Results from the best models on the test set are reported, including those on the eval set."

    The test-set AUCs that support the main claim (Demenba-medium 0.83 vs EfficientNet b6 0.69 in Table II, described as up to 21% relative improvement) are maxima over checkpoints selected using the test labels. The test set therefore participates in model selection, so the reported 'prediction' on the test set is an optimistically biased fit, not an unbiased estimate. With only 10 recordings per class, selection noise is large, making the 0.83 vs 0.69 gap a selected maximum rather than a reliable comparison.

  2. fitted input called prediction [Eq. (1), Section IV-E, Fig. 5, Table III]
    "Fig. 5 examines how AUC varies as we sweep the fusion weight between the audio-only Demenba and the text-only classifier, including LLaMA and Qwen2-7B-Instruct."

    The fusion weight lambda in paudio+text = (1-lambda)paudio + lambda*ptext is swept on the test set (Fig. 5 y-axis is 'Test AUC'). The reported 0.95 audio+BERT AUC in Table III is the peak of this test-set sweep. Thus the best multimodal result is obtained by fitting lambda to the test labels and then reported as a performance number, reducing the fusion 'prediction' to an in-sample maximum.

1 more flagged steps
  1. fitted input called prediction [Table II caption]
    "only AUCs for the better of BCE and CE losses are shown."

    For each model, the reported AUC is the higher of two runs trained with different losses (BCE vs CE), selected by test-set AUC. This makes every comparison in Table II a comparison of test-selected maxima rather than a pre-specified model, further inflating the reported gap between Demenba and EfficientNet.

full rationale

The central architectural claim is not circular: the SSM backbone is taken from prior work (DASS), but the comparison against EfficientNet is an external empirical benchmark, and no uniqueness theorem or definitional identity forces the result. The circularity is in the evaluation protocol: the test set is used to choose checkpoints, loss variants, and fusion weights, and the same test-set numbers are then presented as the paper's headline results. This is the 'fitted input called prediction' pattern: the reported 21% improvement and the 0.95 fusion AUC reduce to maxima over test-set selections. The eval-set results in Table IV are less circular because the checkpoint selection uses the test set and evaluation uses a disjoint eval set, though that protocol still makes the 'test' set a validation set. Overall, the paper's main empirical contribution is partially circular, not fully definitional. Score 6.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central empirical claim rests on a handful of tuned hyperparameters (segment count k, fusion weight lambda, loss weights) and on domain assumptions about label accuracy, diarization correctness, ASR informativeness, and split representativeness. No new physical or mathematical entities are introduced.

free parameters (3)
  • selective voting parameter k = tuned on held-out validation, value not reported
    Number of top segments chosen for soft majority vote; influences recording-level scores and is tuned on a held-out validation set (Section III).
  • fusion weight lambda = model-dependent, e.g., about 0.5 for Qwen2, about 0.1 for Llama (Fig. 5)
    In Eq. (1), the audio-text fusion weight is swept and the peak on the test set is shown in Fig. 5; the final lambda is selected per model.
  • weighted CE loss weights = (1, 3, 3) for normal/MCI/dementia
    Weights chosen to maximize validation performance (Section IV-A).
assumptions (5)
  • domain assumption Adjudicated FHS clinical labels are correct ground truth for normal/MCI/dementia.
    Section IV-A: labels come from adjudicated review; any label noise transfers to the reported AUCs.
  • domain assumption The 10-per-class random test split is representative and participants do not overlap with training.
    Section IV-A: participants do not overlap, but no statement is made about interviewers or recording conditions; with 10 recordings per class, AUC estimates have large variance.
  • ad hoc to paper Speaker diarization labels the first speaker with more than 20ms as the interviewer and all others as the participant.
    Section IV-A; an incorrect role assignment would shift speaker-role attribution in the analysis of Fig. 4.
  • domain assumption Whisper-Large v2 ASR transcripts with about 60% CER remain informative enough for the text branch.
    Section IV-A; the text classifiers are trained on noisy ASR output and evaluated on the same type of transcripts.
  • domain assumption Interviewer identity and test administration procedures do not leak diagnostic information to the model.
    The paper checks for dementia keywords but does not control for interviewer overlap between train and test, despite citing evidence that interviewer speech is informative (refs [6]-[8]).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Recognizing Dementia from Neuropsychological Tests with State Space Models." pith.science (2026). https://pith.science/paper/FWZK5ATT

@misc{pith2026250710311,
  author       = {Pith},
  title        = {Pith review of: Recognizing Dementia from Neuropsychological Tests with State Space Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FWZK5ATT}},
  note         = {Machine review of arXiv:2507.10311}
}
read the original abstract

Early detection of dementia is critical for timely medical intervention and improved patient outcomes. Neuropsychological tests are widely used for cognitive assessment but have traditionally relied on manual scoring. Automatic dementia classification (ADC) systems aim to infer cognitive decline directly from speech recordings of such tests. We propose Demenba, a novel ADC framework based on state space models, which scale linearly in memory and computation with sequence length. Trained on over 1,000 hours of cognitive assessments administered to Framingham Heart Study participants, some of whom were diagnosed with dementia through adjudicated review, our method outperforms prior approaches in fine-grained dementia classification by 21\%, while using fewer parameters. We further analyze its scaling behavior and demonstrate that our model gains additional improvement when fused with large language models, paving the way for more transparent and scalable dementia assessment tools. Code: https://anonymous.4open.science/r/Demenba-0861

Figures

Figures reproduced from arXiv: 2507.10311 by the authors.

Figure 1
Figure 1. Overall architecture of the proposed ADC model. The model consists of a frozen speech segmenter that divides the hour-long recording into shorter segments, a trainable SSM-based audio classifier and a trainable text-based text classifier. The predictions from the two classifiers are then combined via late fusion. allows noisy or uninformative segments to dilute the overall decision. To mitigate this, we adopt a sele… view at source ↗
Figure 2
Figure 2. illustrates how Demenba performance evolves as we vary (a) the amount of training data and (b) the maximum segment duration. In Fig. 2a, we plot AUC as a function of total training hours per class, from 50 to 400 hours, for both 2-class and 3-class ADC. Both Demenba-small and Demenba￾medium show steadily increasing AUC up to 400 hours per class. Even at the largest data point we tested, the curves have not plateaued… view at source ↗
Figure 3
Figure 3. Effect of including silences in audio on dementia classification performance for different model sizes and number of classes. All models are trained with 100-hour/class subset and 360s segments. M (k) and S (k) stand for Demenba￾medium (k-class) and Demenba-small (k-class) respectively. Demenba-medium Demenba-small Model Name 0.60 0.65 0.70 0.75 0.80 Eval AUC Segment Type Both Participant only Interviewer only [PIT… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Effect of including different speaker role segments for 2-class classification on the eval set. All models are trained with 360-second segments. class), whereas Demenba-small improves by 8%. This suggests that Demenba-medium’s deeper architecture better captures long-r…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 32 canonical work pages

  1. [1]

    Nonsteroidal antiinflammatory drugs for the prevention of alzheimer’s disease: a systematic review,

    C. A. Szekely, J. E. Thorne, P. P. Zandi, M. Ek, E. Messias, J. C. Breitner, and S. N. Goodman, “Nonsteroidal antiinflammatory drugs for the prevention of alzheimer’s disease: a systematic review,” Neuroepi- demiology, vol. 23, no. 4, pp. 159–169, 2004

    work page 2004

  2. [2]

    Midlife adiposity predicts earlier onset of alzheimer’s dementia, neuropathology and presymptomatic cerebral amyloid accumulation,

    Y . F. Chuang, Y . An, M. Bilgel, D. F. Wong, J. C. Troncoso, R. J. O’Brien, J. C. Breitner, L. Ferrucci, S. M. Resnick, and M. Tham- bisetty, “Midlife adiposity predicts earlier onset of alzheimer’s dementia, neuropathology and presymptomatic cerebral amyloid accumulation,” Molecular Psychiatry, vol. 21, pp. 910–915, 2016

    work page 2016

  3. [3]

    The mini-mental state examination (mmse),

    L. Kurlowicz and M. Wallace, “The mini-mental state examination (mmse),” Journal of Gerontological Nursing , vol. 25, no. 5, pp. 8–9, 1999

    work page 1999

  4. [4]

    De- tecting cognitive decline using speech only: The ADReSSo challenge,

    S. Luz, F. Haider, S. de la Fuente, D. Fromm, and B. MacWhinney, “De- tecting cognitive decline using speech only: The ADReSSo challenge,” in Interspeech, pp. 3780–3784, 2021

    work page 2021

  5. [5]

    Spoken language biomarkers for detect- ing cognitive impairment,

    T. Alhanai, R. Au, and J. Glass, “Spoken language biomarkers for detect- ing cognitive impairment,” in 2017 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) , pp. 409–416, 2017

    work page 2017

  6. [6]

    Detecting dementia from long neuropsychological interviews,

    N. Dawalatabad, Y . Gong, S. Khurana, R. Au, and J. Glass, “Detecting dementia from long neuropsychological interviews,” in Findings of the Association for Computational Linguistics: EMNLP 2022 , (Abu Dhabi, United Arab Emirates), pp. 5270–5283, Association for Computational Linguistics, 2022

    work page 2022

  7. [7]

    Role-specific language models for processing recorded neuropsychological exams,

    T. Al Hanai, R. Au, and J. Glass, “Role-specific language models for processing recorded neuropsychological exams,” in Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers) (M. Walker, H. Ji, and A. Stent, eds.), (New Orleans, Louisiana), pp. ...

    work page 2018

  8. [8]

    Influence of the interviewer on the automatic assessment of alzheimer’s disease in the context of the adresso challenge,

    P. P ´erez-Toro, S. Bayerl, T. Arias-Vergara, J. V ´asquez-Correa, P. Klumpp, M. Schuster, E. N ¨oth, J. Orozco-Arroyave, and K. Ried- hammer, “Influence of the interviewer on the automatic assessment of alzheimer’s disease in the context of the adresso challenge,” in Proceedings of Interspeech , pp. 3785–3789, 2021

    work page 2021

Show all 40 references
  1. [9]

    Attention is all you need,

    Vaswani et al., “Attention is all you need,” in NeurIPS, p. 6000–6010, 2017

    2017

  2. [10]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, and C. R ´e, “Efficiently modeling long sequences with structured state spaces,” arXiv preprint arXiv:2111.00396 , 2021

    2021 arXiv

  3. [11]

    Comparing acoustic-based approaches for alzheimer’s disease detection,

    A. Balagopalan and J. Novikova, “Comparing acoustic-based approaches for alzheimer’s disease detection,” in Interspeech, pp. 3800–3804, 2021

    2021

  4. [12]

    Leveraging pretrained representations with task-related keywords for alzheimer’s disease detection,

    J. Li, K. Song, J. Li, B. Zheng, D. Li, X. Wu, X. Liu, and H. Meng, “Leveraging pretrained representations with task-related keywords for alzheimer’s disease detection,” in ArXiv, 2023

    2023

  5. [13]

    Dass: Distilled audio state space models are stronger and more duration- scalable learners,

    S. Bhati, Y . Gong, L. Karlinsky, H. Kuehne, R. Feris, and J. Glass, “Dass: Distilled audio state space models are stronger and more duration- scalable learners,” in SLT, 2024

    2024

  6. [14]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, and C. R ´e, “Efficiently modeling long sequences with structured state spaces,” ICLR, 2022

    2022

  7. [15]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023

    2023 arXiv

  8. [16]

    Alzheimer’s dementia recognition through spontaneous speech: The ADReSS Challenge,

    S. Luz, F. Haider, S. de la Fuente, D. Fromm, and B. MacWhinney, “Alzheimer’s dementia recognition through spontaneous speech: The ADReSS Challenge,” in Interspeech, (Shanghai, China), 2020

    2020

  9. [17]

    Alzheimer’s dementia recogni- tion using acoustic, lexical, disfluency and speech pause features robust to noisy inputs,

    M. Rohanian, J. Hough, and M. Purver, “Alzheimer’s dementia recogni- tion using acoustic, lexical, disfluency and speech pause features robust to noisy inputs,” in Interspeech, pp. 3820–3824, 2021

    2021

  10. [18]

    Detection of dementia on voice recordings using deep learning: a framingham heart study,

    C. Xue, C. Karjadi, I. C. Paschalidis, R. Au, and V . B. Kolachalama, “Detection of dementia on voice recordings using deep learning: a framingham heart study,” Alzheimer’s Research & Therapy , vol. 13, p. 146, Aug. 2021

    2021

  11. [19]

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

    A. Baevski, H. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” in NeurIPS, 2020

    2020

  12. [20]

    Robust speech recognition via large-scale weak super- vision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak super- vision,” in ICML, 2023

    2023

  13. [21]

    Classifying alzheimer’s disease using audio and text-based representations of speech,

    R. Haulcy and J. Glass, “Classifying alzheimer’s disease using audio and text-based representations of speech,” Frontiers in Psychology, vol. 11, p. 624137, 2021

    2021

  14. [22]

    Vmamba: Visual state space model,

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, and Y . Liu, “Vmamba: Visual state space model,” arXiv preprint arXiv:2401.10166, 2024

    2024 arXiv

  15. [23]

    Vision mamba: Efficient visual representation learning with bidirectional state space model,

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang, “Vision mamba: Efficient visual representation learning with bidirectional state space model,” arXiv preprint arXiv:2401.09417 , 2024

    2024 arXiv

  16. [24]

    Audio mamba: Bidirectional state space model for audio representation learning,

    M. H. Erol, A. Senocak, J. Feng, and J. S. Chung, “Audio mamba: Bidirectional state space model for audio representation learning,” IEEE Signal Processing Letters , vol. 31, pp. 2975–2979, 2024

    2024

  17. [25]

    Audio mamba: Pretrained audio state space model for audio tagging,

    J. Lin and H. Hu, “Audio mamba: Pretrained audio state space model for audio tagging,” arXiv preprint arXiv:2405.13636 , 2024

    2024 arXiv

  18. [26]

    Ssamba: Self- supervised audio representation learning with mamba state space model,

    S. Shams, S. S. Dindar, X. Jiang, and N. Mesgarani, “Ssamba: Self- supervised audio representation learning with mamba state space model,” arXiv preprint arXiv:2405.11831 , 2024

    2024 arXiv

  19. [27]

    Mamba in speech: Towards an alternative to self-attention,

    X. Zhang, Q. Zhang, H. Liu, T. Xiao, X. Qian, B. Ahmed, E. Am- bikairajah, H. Li, and J. Epps, “Mamba in speech: Towards an alternative to self-attention,” IEEE Transactions on Audio, Speech and Language Processing, 2025

    2025

  20. [28]

    Speech slytherin: Examining the performance and efficiency of mamba for speech separation, recognition, and synthesis,

    X. Jiang, Y . A. Li, A. N. Florea, C. Han, and N. Mesgarani, “Speech slytherin: Examining the performance and efficiency of mamba for speech separation, recognition, and synthesis,” in ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing ...

    2025

  21. [29]

    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,” in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nol...

    2019

  22. [30]

    Powerset multi-class cross entropy loss for neural speaker diarization,

    A. Plaquet and H. Bredin, “Powerset multi-class cross entropy loss for neural speaker diarization,” in Proc. INTERSPEECH 2023 , 2023

    2023

  23. [31]

    pyannote.audio 2.1 speaker diarization pipeline: principle, benchmark, and recipe,

    H. Bredin, “pyannote.audio 2.1 speaker diarization pipeline: principle, benchmark, and recipe,” in Proc. INTERSPEECH 2023 , 2023

    2023

  24. [32]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 248–255, IEEE, 2009

    2009

  25. [33]

    Audio set: An ontology and human-labeled dataset for audio events,

    J. F. Gemmeke, D. P. W. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter, “Audio set: An ontology and human-labeled dataset for audio events,” in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), (New Orle...

    2017

  26. [34]

    Efficientnet: Rethinking model scaling for convolutional neural networks,

    M. Tan and Q. V . Le, “Efficientnet: Rethinking model scaling for convolutional neural networks,” in ICML, vol. 97, pp. 6105–6114, PMLR, 2019

    2019

  27. [35]

    LLaMA: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lanchaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample, “LLaMA: Open and efficient foundation language models,” arXiv preprint arXiv:2302.13971 , 2023

    2023 arXiv

  28. [36]

    Qwen2 technical report,

    A. Yang et al. , “Qwen2 technical report,” tech. rep., Qwen Team, Alibaba Group, 2024. Technical Report

    2024

  29. [37]

    Phi-4-mini technical report: Compact yet powerful multi- modal language models via mixture-of-loras,

    Microsoft, “Phi-4-mini technical report: Compact yet powerful multi- modal language models via mixture-of-loras,” Technical Report, 2024

    2024

  30. [38]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,” 2021

    2021

  31. [39]

    Adam: A method for stochastic optimiza- tion,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimiza- tion,” in Proceedings of the 3rd International Conference on Learning Representations (ICLR), (San Diego, CA), 2015

    2015

  32. [40]

    Scaling laws for neural language models,

    J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei, “Scaling laws for neural language models,” arXiv preprint arXiv:2001.08361 , 2020

    2001 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.