Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Zero-Shot Cognitive Impairment Detection from Speech Using AudioLLM

T0 review · 4 major / 6 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 AudioLLM spots cognitive decline zero-shot, rivaling supervised models.

desk verdict Genuinely new zero-shot AudioLLM application, but best-of-five prompt selection on the evaluation labels inflates the headline numbers; the held-out test comparison is more credible. read the letter →

arxiv 2506.17351 v1 pith:PT6SK4VB submitted 2025-06-20 cs.SD cs.AIcs.CLcs.MMeess.AS

Mostafa Shahin , Beena Ahmed , Julien Epps This is my paper
classification cs.SDcs.AIcs.CLcs.MMeess.AS
keywords zero-shotclassificationcognitiveimpairmentdetectionAudioLLMQwen2-Audiospeech-basedscreeningpromptengineeringChain-of-Thoughtpromptingmultilingualspeechanalysis
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

The paper asks whether a general-purpose audio language model, never fine-tuned for dementia screening, can classify speech as normal cognition or cognitive impairment by following plain-English prompts. It answers yes: on the TAUKADIAL and PROCESS benchmarks, the zero-shot Qwen2-Audio model reaches accuracy comparable to supervised classifiers and transfers across English, Mandarin, picture description, and verbal fluency tasks. The central claim is that prompt design alone, not parameter updates, can extract clinically relevant acoustic and linguistic cues from a pretrained audio LLM. This would matter because cognitive screening today depends on scarce, privacy-sensitive labelled speech data, and a zero-shot prompt-based route could lower that barrier.

What carries the argument

The mechanism is the pretrained AudioLLM itself, Qwen2-Audio, which concatenates Whisper audio encodings with tokenized text prompts and generates a single-word verdict. The authors steer it only through natural-language instructions; the five prompt types vary the amount of context and the demand for internal step-by-step reasoning. A chat-style template wraps the audio between special tokens and the text prompt, and for each prompt type five reworded variants are generated to test sensitivity to phrasing. The load-bearing operation is the model's ability to ground a diagnostic decision in acoustic and linguistic properties of the audio without any update to its weights.

What would settle it

Recompute unweighted average recall using prompts and majority-vote rules chosen only on the official training splits, then apply them to the untouched official test splits; if the scores fall materially below the reported 59% and 57.5%, the zero-shot claim has been overstated.

Watch

Extended reading notes

Core claim

The central discovery is that Qwen2-Audio, an audio-and-text-to-text model built on the Qwen LLM with a Whisper-large-v3 encoder, can perform zero-shot binary classification of speech samples into normal cognition or cognitive impairment across two challenge datasets. The authors design five prompt archetypes (Direct, Contextual, Informative, Chain-of-Thought, and CoT with demographic context) and find that contextual prompts improve accuracy over minimal ones, CoT prompting does not beat direct prompting, and fluency tasks yield stronger cues than picture description. Reported unweighted average recall reaches 57.5% on TAUKADIAL overall and 59% on the TAUKADIAL test set when majority-voting the best prompt per type, described as comparable to or higher than in-domain supervised baselines; on PROCESS, picture description reaches 56.3% UAR while fluency tasks score higher. The authors conclude that the zero-shot AudioLLM approach generalizes across languages, tasks, and datasets.

Load-bearing premise

The evaluation protocol assumes that reporting only the best of five reworded prompts per prompt type, chosen on the same merged train and test data, does not materially inflate the reported accuracy.

Editorial extensions

If this is right

  • A single general audio LLM can act as a cross-lingual, cross-task CI screener without retraining, provided the prompt supplies task context.
  • Prompt engineering becomes a primary performance lever: adding task, age, gender, and language context shifts UAR by several points, so prompt choice matters as much as model choice.
  • Fluency-based tasks (semantic and phonemic) carry stronger CI cues for this model than picture description, suggesting that task selection can improve zero-shot screening.
  • Prompt rewording changes results notably, yet the same wording tends to perform consistently across datasets, so a small set of well-validated prompts could be reused.
  • If the results hold, deployment of CI screening no longer requires annotated speech from the target population; only the pretrained model and a set of prompt templates are needed.

Reading between the lines

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

  • In our reading, the same prompt that works across datasets implies that a portable 'screening instruction' could be distilled and tested on other audio LLMs beyond Qwen2-Audio; the paper does not test this transfer.
  • The higher Mandarin scores may stem from the model's pretraining data distribution, a possibility the paper raises only in passing; if so, language-aware prompt tuning could close the gap.
  • A natural extension the authors do not pursue is attribution analysis: checking whether verdicts correlate with pause duration, speech rate, or word-finding behavior would show whether the model reasons about genuine CI cues rather than dataset artifacts.
  • Under a strictly conservative protocol, the best-prompt-per-type selection on combined train/test data would be replaced by prompt selection on a held-out development set; we expect the reported gaps to narrow, though the rank order across prompt types may persist.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. This manuscript proposes a zero-shot method for binary cognitive-impairment (CI) detection from speech using the pretrained Qwen2-Audio AudioLLM. The authors design five prompt types (Direct, Contextual, Informative, CoT, CoT-Informative), generate five ChatGPT-reworded variants per type, and evaluate on the TAUKADIAL dataset (English/Mandarin picture description) and the PROCESS dataset (three English tasks). The paper reports UAR/macro-F1 for each prompt type, compares Qwen2-Audio on the TAUKADIAL test split with official challenge baselines, and reports PROCESS results per task. The central claims are that the zero-shot AudioLLM achieves performance comparable to supervised methods and that it generalizes across languages, tasks, and datasets.

Significance. If the reported performance held under a proper evaluation protocol, the result would be of practical interest: a general audio-language model could screen cognitive status without task-specific fine-tuning, and prompt design alone could yield competitive accuracy on picture-description and fluency tasks. The paper is clearly written and uses two publicly available challenge datasets, and it provides a direct comparison with the official TAUKADIAL baseline on the test split. The main limitation is that the headline numbers in Table IV and Fig. 2 are obtained by selecting the best of five reworded prompts using the same merged data on which the metrics are computed, which is an optimistic selection procedure rather than an unbiased zero-shot estimate. The PROCESS results are also not compared with any supervised baseline. These issues affect the abstract's 'comparable to supervised' claim directly, so the current evidence is suggestive rather than conclusive.

major comments (4)
  1. [V, Table IV and Fig. 2; Section III-A] The evaluation protocol for Table IV and Fig. 2 reports only the best-performing prompt variant within each prompt type, selected using the same merged train/test (or train/dev) samples on which the UAR is computed. Because the labels are visible to the selection step, the 57.5% Contextual UAR is the maximum over five reworded variants, and the row-wise comparison is a maximum over 25 prompt configurations. This is an optimistic, selection-on-labels estimate and does not support the 'comparable to supervised methods' claim as a zero-shot result. Please either select prompts on a held-out split, report all five variants with their mean and spread, or apply a multiple-comparison correction.
  2. [V, Fig. 2; Section III-A] The PROCESS results are reported only as the best-prompt UAR per task, again selected on the merged evaluation data, and no supervised baseline for PROCESS is included. Consequently, the abstract's claims of 'comparable to supervised methods' and 'generalizability across datasets' are supported only by the TAUKADIAL test-set comparison in Table V. Please add official PROCESS challenge baselines (or clearly state that none are available) and use a prompt-selection protocol that does not touch the test labels.
  3. [IV; Section III-B] The zero-shot interpretation assumes Qwen2-Audio was not trained on TAUKADIAL or PROCESS samples. Qwen2-Audio was pretrained on a large public audio corpus, and both evaluation datasets were publicly available before the model's release; the paper provides no check for overlap between the model's pretraining data and these evaluation sets. Without such a check, the 'zero-shot' claim is vulnerable to contamination. Please provide evidence of no leakage, for example by querying the model with held-out metadata, using temporal separation, or performing a canary analysis.
  4. [V, Table V] The comparison with supervised baselines on the TAUKADIAL test set lacks any measure of uncertainty. The 59.0% MV UAR is within 0.2 points of the strongest baseline (w2v+ling+eGeMAPs hard fusion, 59.2%), and without confidence intervals or a significance test the statements 'outperformed nearly all baseline models' and 'comparable to supervised methods' are not quantitatively established. Also, the text says 'Using the training split, we selected the best prompt within each type' — this is a sounder procedure than Table IV, but the preceding phrase 'five best-performing prompts (one per type)' needs rewording to make it unambiguous that the test-set prompts were fixed before seeing test labels.
minor comments (6)
  1. [III-A] The description of which PROCESS subsets were merged is inconsistent: the text says 'training and testing subsets' while adjacent statistics describe only a training and development set; please clarify whether the test set was used.
  2. [V] The sentence 'This improved the test UAR from 55.2% (best individual Contextual prompt) to 59' is missing a percent sign and is grammatically incomplete.
  3. [Table I] The Healthy/CI column for TAUKADIAL lists 165/222, but the text reports 507 samples with 285 MCI and 222 NC; please distinguish subjects from samples consistently.
  4. [Table II and text] The label 'NC' is expanded as 'Normal Cognitive Decline' in several places; the standard term is 'normal cognition' or 'cognitively normal,' and the current wording is confusing.
  5. [Fig. 3] The caption says 'across four datasets/tasks,' but the bars are averages over four experimental conditions (TAUKADIAL, PROCESS-CTD, PROCESS-SFT, PROCESS-PFT); please rephrase to 'across four experiments.'
  6. [References and Section III-C] Reference [18] contains a garbled author list, and the text uses both 'ChatGPT-4.0' (Section III-C) and 'ChatGPT-4o' (Section V); please harmonize these.

Circularity Check

1 steps flagged · score 6.0 of 10

Best-of-five prompt selection on the evaluation labels carries the headline UARs in Table IV and Fig. 2, making those zero-shot numbers partly a selected maximum rather than a fixed configuration; Table V provides a partial independent check.

  1. fitted input called prediction [Section III-A (Datasets) and Section V (Results), Table IV and Fig. 2]
    "However, since our approach relies on zero-shot classification using an AudioLLM, we merge the train and test sets for evaluation. ... Table IV presents the Unweighted Average Recall (UAR) and macro-F1 scores for the TAUKADIAL dataset across all prompt types, reporting only the best-performing prompt (based on UAR) within each type. ... Here, we also report the UAR of the best prompt for each prompt type in each task, highlighting the most effective prompt variations for different cognitive assessment tasks."

    The zero-shot claim is that no task-specific tuning occurs, yet the prompt variant is selected by evaluating all five reworded variants on the same merged train-plus-test data and retaining the one with the highest UAR. The selection criterion is the UAR computed against the very labels being predicted, so each reported number in Table IV and Fig. 2 is a maximum over five prompt variants rather than the accuracy of a pre-specified zero-shot prompt. This is a discrete fit of the prompt configuration to the evaluation labels, renamed as zero-shot performance: the reported UAR is, by construction, the best of several guesses on the evaluation set.

full rationale

The derivation chain is: Qwen2-Audio is a frozen pretrained model; a text prompt is chosen; the model's outputs are scored as UAR; that UAR is presented as zero-shot CI detection performance. The circular step sits in the prompt choice: Section III-A explicitly merges train and test sets for evaluation, and Section V reports only the best-performing prompt within each type based on UAR. Because the same labels used to compute the reported metric are also used to select which prompt to report, the headline numbers in Table IV and Fig. 2 are selected maxima, not the performance of a fixed zero-shot configuration. This is a fitted input called a prediction. The paper is not fully circular because the model weights are never updated and because Table V evaluates a majority-vote combination on the official test split with prompt selection performed on the training split; that comparison (59.0 UAR vs the best supervised baseline's 59.2) gives the 'comparable to supervised' claim some independent support, though no significance test is reported. The absence of a pretraining-data leakage check is a validity risk, not a circularity. Overall, one important set of predictions reduces by construction to label-based prompt selection, so the circularity score is 6.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on no fitted model parameters, but it does rest on the dataset labels being trustworthy, on no data leakage from Qwen2-Audio's pretraining, and on an evaluation protocol that selects the best prompt variant against the evaluation labels. The prompt-variant selection is the main free choice and is the most consequential.

free parameters (1)
  • Best prompt variant index per prompt type (0-4) = Not reported; selected as the variant with maximum UAR on the evaluation data
    Section V states Table IV reports 'only the best-performing prompt (based on UAR) within each type' on the merged TAUKADIAL set, and the MV result selects the best prompt within each type on the training split. This choice is optimized against the labels and can inflate reported accuracy.
assumptions (4)
  • domain assumption Qwen2-Audio's pretraining data does not include TAUKADIAL or PROCESS recordings or closely similar cognitive-assessment speech, so the zero-shot evaluation is uncontaminated.
    The zero-shot claim in Sections I and III.B depends on the model never having seen these datasets. No overlap or leakage check is provided in Section IV.
  • domain assumption The official dataset labels (NC, MCI, dementia) are accurate ground truth for cognitive status.
    All metrics in Tables IV-V and Fig. 2 use the challenge-provided labels without independent clinical verification (Section III.A).
  • domain assumption Merging MCI and dementia into a single CI class preserves the binary detection task despite the prompts asking for NC vs MCI.
    Section III.A merges dementia into CI for PROCESS to unify labels, but the prompts in Table II instruct the model to choose between NC and MCI, so a dementia sample is not exactly captured by the MCI label.
  • ad hoc to paper The five ChatGPT-generated rewordings of each prompt preserve the intended prompt type and introduce no systematic bias.
    Section III.C and V rely on these variants to measure rephrasing sensitivity, but the variants are not included in the paper, so this assumption cannot be checked.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Zero-Shot Cognitive Impairment Detection from Speech Using AudioLLM." pith.science (2026). https://pith.science/paper/PT6SK4VB

@misc{pith2026250617351,
  author       = {Pith},
  title        = {Pith review of: Zero-Shot Cognitive Impairment Detection from Speech Using AudioLLM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PT6SK4VB}},
  note         = {Machine review of arXiv:2506.17351}
}
read the original abstract

Cognitive impairment (CI) is of growing public health concern, and early detection is vital for effective intervention. Speech has gained attention as a non-invasive and easily collectible biomarker for assessing cognitive decline. Traditional CI detection methods typically rely on supervised models trained on acoustic and linguistic features extracted from speech, which often require manual annotation and may not generalise well across datasets and languages. In this work, we propose the first zero-shot speech-based CI detection method using the Qwen2- Audio AudioLLM, a model capable of processing both audio and text inputs. By designing prompt-based instructions, we guide the model in classifying speech samples as indicative of normal cognition or cognitive impairment. We evaluate our approach on two datasets: one in English and another multilingual, spanning different cognitive assessment tasks. Our results show that the zero-shot AudioLLM approach achieves performance comparable to supervised methods and exhibits promising generalizability and consistency across languages, tasks, and datasets.

Figures

Figures reproduced from arXiv: 2506.17351 by the authors.

Figure 1
Figure 1. The architecture of Qwen2-Audio. The model is built on the [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The Unweighted Average Recall (UAR) and macro F1 (mF1) scores for the PROCESS dataset across three tasks: Cookie Theft Description (CTD), [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. The impact of rewording prompts on Qwen2-Audio performance. Each prompt type has five variants (indexed 0–4). The bars represent the average [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in 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. Toward Generalizable Cognitive Impairment Detection with Speech-Based Multimodal Large Language Models

    eess.SP 2026-07 reject novelty 4.0 of 10

    A multimodal LLM pipeline (Qwen audio + Qwen text embeddings, concatenated and classified) reaches 92.4% accuracy on a combined ADReSS20/ADReSSo21 test set, but the evaluation does not justify state-of-the-art or cros...

Reference graph

Works this paper leans on

51 extracted references · 30 canonical work pages · cited by 1 Pith paper

  1. [1]

    2023 Alzheimer’s disease facts and figures,

    Association Alzheimer’s, “2023 Alzheimer’s disease facts and figures,” Tech. Rep. 4, 4 2023. [Online]. Available: https://onlinelibrary.wiley.com/doi/full/10.1002/alz.13016 https://onlinelibrary.wiley.com/doi/abs/10.1002/alz.13016 https://alz- journals.onlinelibrary.wiley.com/doi/10.1002/alz.13016

    doi:10.1002/alz.13016 2023

  2. [2]

    Subtle changes in daily functioning predict conversion from normal to mild cognitive impairment or dementia: an analysis of the NACC database,

    M. A. Nowrangi, P. B. Rosenberg, and J. M. S. Leoutsakos, “Subtle changes in daily functioning predict conversion from normal to mild cognitive impairment or dementia: an analysis of the NACC database,” International psychogeriatrics, vol. 28, no. 12, p. 2009, 12 2016. [Online]. Available: https://pmc.ncbi.nlm.nih.gov/articles/PMC5628501/

    work page 2009

  3. [3]

    Characteristics of people living with undiagnosed dementia: findings from the CFAS Wales study,

    L. D. Gamble, F. E. Matthews, I. R. Jones, A. E. Hillman, B. Woods, C. A. Macleod, A. Martyr, R. Collins, C. Pentecost, J. M. Rusted, and L. Clare, “Characteristics of people living with undiagnosed dementia: findings from the CFAS Wales study,” BMC Geriatrics, vol. 22, no. 1, pp. 1–12, 12 2022. [Online]. Avail- able: https://bmcgeriatr.biomedcentral.com/...

    work page doi:10.1186/s12877-022- 2022

  4. [4]

    Speech based detection of Alzheimer’s disease: a survey of AI techniques, datasets and challenges,

    K. Ding, M. Chetty, Azadeh, N. Hoshyar, Tanusri Bhattacharya, and B. Klein, “Speech based detection of Alzheimer’s disease: a survey of AI techniques, datasets and challenges,”Artificial Intelligence Review 2024 57:12, vol. 57, no. 12, pp. 1–43, 10 2024. [Online]. Available: https://link.springer.com/article/10.1007/s10462-024-10961-6

    doi:10.1007/s10462-024-10961-6 2024

  5. [5]

    Artificial Intel- ligence, Speech, and Language Processing Approaches to Monitoring Alzheimer’s Disease: A Systematic Review,

    S. De La Fuente Garcia, C. W. Ritchie, and S. Luz, “Artificial Intel- ligence, Speech, and Language Processing Approaches to Monitoring Alzheimer’s Disease: A Systematic Review,”Journal of Alzheimer’s Disease, vol. 78, no. 4, pp. 1547–1574, 2020

    work page 2020

  6. [6]

    Tackling the ADReSS challenge: a multimodal approach to the automated recognition of Alzheimer’s dementia,

    M. Martinc and S. Pollak, “Tackling the ADReSS challenge: a multimodal approach to the automated recognition of Alzheimer’s dementia,” inInterspeech, 2020. [Online]. Available: http://dx.doi.org/10.21437/Interspeech.2020-2202

    doi:10.21437/interspeech.2020-2202 2020

  7. [7]

    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,” inInterspeech, 2020. [Online]. Available: http://dx.doi.org/10.21437/Interspeech.2020-2571

    doi:10.21437/interspeech.2020-2571 2020

  8. [8]

    Multiscale System for Alzheimer’s Dementia Recognition through Spontaneous Speech,

    E. Edwards, C. Dognin, B. Bollepalli, and M. Singh, “Multiscale System for Alzheimer’s Dementia Recognition through Spontaneous Speech,” inInterspeech, 2020. [Online]. Available: http://dx.doi.org/10.21437/Interspeech.2020-2781

    doi:10.21437/interspeech.2020-2781 2020

Show all 51 references
  1. [9]

    Connected Speech-Based Cognitive Assessment in Chinese and English,

    S. Luz, S. De, L. Fuente Garcia, F. Haider, D. Fromm, B. Macwhinney, A. Lanzi, Y .-N. Chang, C.-J. Chou, and Y .-C. Liu, “Connected Speech-Based Cognitive Assessment in Chinese and English,” in Interspeech, 2024. [Online]. Available: http://luzs.gitlab.io/taukadial/

    2024

  2. [10]

    Cognitive Insights Across Languages: Enhancing Multimodal Interview Analysis,

    D. Ortiz-Perez, J. Garcia-Rodriguez, and D. Tom ´as, “Cognitive Insights Across Languages: Enhancing Multimodal Interview Analysis,” inIn- terspeech. International Speech Communication Association, 9 2024, pp. 952–956

    2024

  3. [11]

    Multilingual Speech and Language Analysis for the Assessment of Mild Cognitive Impairment: Outcomes from the Taukadial Challenge,

    P. Andrea Perez-Toro, T. Arias-Vergara, P. Klumpp, T. Weise, M. Schuster, E. Noth, J. R. Orozco-Arroyave, and A. Maier, “Multilingual Speech and Language Analysis for the Assessment of Mild Cognitive Impairment: Outcomes from the Taukadial Challenge,” in Interspeech, 2024. [On...

    2024

  4. [12]

    The natural history of Alzheimer’s disease. Description of study cohort and accuracy of diagnosis,

    J. T. Becker, F. Boiler, O. L. Lopez, J. Saxton, and K. L. Mcgonigle, “The natural history of Alzheimer’s disease. Description of study cohort and accuracy of diagnosis,”Archives of neurology, vol. 51, no. 6, pp. 585–594, 1994. [Online]. Available: https://pubmed.ncbi.nlm.nih....

    1994

  5. [13]

    Early Dementia Detection Using Multiple Spontaneous Speech Prompts: The PROCESS Challenge,

    F. Tao, B. Mirheidari, M. Pahar, S. Young, Y . Xiao, H. Elghazaly, F. Peters, C. Illingworth, D. Braun, R. O’Malley, S. Bell, D. Blackburn, F. Haider, S. Luz, and H. Christensen, “Early Dementia Detection Using Multiple Spontaneous Speech Prompts: The PROCESS Challenge,” in IC...

    2025

  6. [14]

    A SURVEY ON SPEECH LARGE LANGUAGE MODELS,

    J. Peng, Y . Wang, Y . Xi, X. Li, X. Zhang, and K. Yu, “A SURVEY ON SPEECH LARGE LANGUAGE MODELS,”arXiv, 2024

    2024

  7. [15]

    WavLLM: Towards Robust and Adaptive Speech Large Language Model,

    S. Hu, L. Zhou, S. Liu, S. Chen, L. Meng, H. Hao, J. Pan, X. Liu, J. Li, S. Sivasankaran, L. Liu, and F. Wei, “WavLLM: Towards Robust and Adaptive Speech Large Language Model,”arXiv, 3 2024. [Online]. Available: http://arxiv.org/abs/2404.00656

    2024 arXiv

  8. [16]

    SALMONN: TOW ARDS GENERIC HEARING ABILI-TIES FOR LARGE LANGUAGE MODELS,

    C. Tang, W. Yu, G. Sun, X. Chen, T. Tan, W. Li, L. Lu, Z. Ma, and C. Zhang, “SALMONN: TOW ARDS GENERIC HEARING ABILI-TIES FOR LARGE LANGUAGE MODELS,” inICLR, 2024. [Online]. Available: https://github.com/bytedance/SALMONN

    2024

  9. [17]

    Qwen-Audio: Advancing Universal Audio Understanding via Unified Large-Scale Audio-Language Models,

    Y . Chu, J. Xu, X. Zhou, Q. Yang, S. Zhang, Z. Yan, C. Zhou, and J. Zhou, “Qwen-Audio: Advancing Universal Audio Understanding via Unified Large-Scale Audio-Language Models,”arXiv, 2023. [Online]. Available: https://qwen-audio.github.io/

    2023

  10. [18]

    Qwen2-Audio Technical Report,

    Y . Chu, J. Xu, Q. Yang, H. Wei Xipin Wei Zhifang Guo Yichong Leng Yuanjun Lv Jinzheng He Junyang Lin Chang Zhou, and J. Zhou, “Qwen2-Audio Technical Report,”arXiv, 2024. [Online]. Available: https://github.com/QwenLM/Qwen2-Audio

    2024

  11. [19]

    Qwen Technical Report,

    J. Bai, S. Bai, Y . Chu, Z. Cui, K. Dang, X. Deng, Y . Fan, W. Ge, Y . Han, F. Huang, B. Hui, L. Ji, M. Li, J. Lin, R. Lin, D. Liu, G. Liu, C. Lu, K. Lu, J. Ma, R. Men, X. Ren, X. Ren, C. Tan, S. Tan, J. Tu, P. Wang, S. Wang, W. Wang, S. Wu, B. Xu, J. Xu, A. Yang, H. Yang, J. ...

    2023 arXiv

  12. [20]

    A Comparison of Acoustic and Linguistics Methodologies for Alzheimer’s Dementia Recognition,

    N. Cummins, Y . Pan, Z. Ren, J. Fritsch, V . S. Nallanthighal, H. Christensen, D. Blackburn, B. W. Schuller, M. M. Doss, H. Strik, and A. H ¨arm¨a, “A Comparison of Acoustic and Linguistics Methodologies for Alzheimer’s Dementia Recognition,” inInterspeech, 2020. [Online]. Ava...

    2020 doi

  13. [21]

    Multi-Class Dementia Detection Using Acoustic Features - ICASSP-2025 PROCESS Challenge,

    M. A. Zafar, X. Zhang, M. Shahin, and B. Ahmed, “Multi-Class Dementia Detection Using Acoustic Features - ICASSP-2025 PROCESS Challenge,”ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–2, 4 2025. [Online]. Available: ...

    2025

  14. [22]

    Disfluencies and Fine-Tuning Pre-trained Language Models for Detection of Alzheimer’s Disease,

    J. Yuan, Y . Bian, X. Cai, J. Huang, Z. Ye, and K. Church, “Disfluencies and Fine-Tuning Pre-trained Language Models for Detection of Alzheimer’s Disease,” inInterspeech, 2020. [Online]. Available: http://dx.doi.org/10.21437/Interspeech.2020-2516

    2020 doi

  15. [23]

    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,” Advances in Neural Information Processing Systems, vol. 2020- December, 2020. [Online]. Available: https://arxiv.org/abs/2006.11477v3

    2020 arXiv

  16. [24]

    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,”IEEE/ACM Transactions on Audio Speech and Language Processing, vol. 29, pp. 3451–3460, 6 2021. [Onl...

    2021 arXiv

  17. [25]

    Robust Speech Recognition via Large-Scale Weak Supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust Speech Recognition via Large-Scale Weak Supervision,”Proceedings of Machine Learning Research, vol. 202, pp. 28 492–28 518, 12 2022. [Online]. Available: https://arxiv.org/abs/2212.04356v1

    2022 arXiv

  18. [26]

    Leveraging Universal Speech Representations for Detecting and Assessing the Severity of Mild Cognitive Impairment Across Languages,

    A. Favaro, T. Cao, N. Dehak, and L. Moro-Velazquez, “Leveraging Universal Speech Representations for Detecting and Assessing the Severity of Mild Cognitive Impairment Across Languages,” inInter- speech. International Speech Communication Association, 9 2024, pp. 972–976

    2024

  19. [27]

    Combining Acoustic Feature Sets for De- tecting Mild Cognitive Impairment in the Interspeech’24 TAUKADIAL Challenge,

    G. Gosztolya and L. T ´oth, “Combining Acoustic Feature Sets for De- tecting Mild Cognitive Impairment in the Interspeech’24 TAUKADIAL Challenge,” inInterspeech, 2024

    2024

  20. [28]

    Early Detection of Cognitive Decline Using V oice Assistant Commands,

    E. Kurtz, Y . Zhu, T. Driesse, B. Tran, J. A. Batsis, R. M. Roth, and X. Liang, “Early Detection of Cognitive Decline Using V oice Assistant Commands,”ICASSP , IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings, vol. 2023-June, 2023

    2023

  21. [29]

    Enriching Word Vectors with Subword Information,

    P. Bojanowski, E. Grave, A. Joulin, and T. Mikolov, “Enriching Word Vectors with Subword Information,”Transactions of the Association for Computational Linguistics, vol. 5, pp. 135–146, 12 2017. [Online]. Available: https://dx.doi.org/10.1162/tacl a 00051

    2017 doi

  22. [30]

    GloVe: Global Vectors for Word Representation,

    J. Pennington, R. Socher, and C. D. Manning, “GloVe: Global Vectors for Word Representation,” inEMNLP, 2014, pp. 1532–1543. [Online]. Available: http://nlp

    2014

  23. [31]

    Better Word Embeddings by Disentangling Contextual n-Gram Information,

    P. Gupta, M. Pagliardini, and M. Jaggi, “Better Word Embeddings by Disentangling Contextual n-Gram Information,” NAACL 2019 - Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, vol. 1, pp. 933–939, 2019. [Onl...

    2019

  24. [32]

    The Interspeech 2024 TAUKADIAL Challenge: Multilingual Mild Cogni- tive Impairment Detection with Multimodal Approach,

    B. Barrera-Altuna, D. Lee, Z. Zarnaz, J. Han, and S. Kim, “The Interspeech 2024 TAUKADIAL Challenge: Multilingual Mild Cogni- tive Impairment Detection with Multimodal Approach,” inInterspeech. International Speech Communication Association, 9 2024, pp. 967–971

    2024

  25. [33]

    Translingual Language Markers for Cognitive Assessment from Spontaneous Speech,

    B. Hoang, Y . Pang, H. Dodge, and J. Zhou, “Translingual Language Markers for Cognitive Assessment from Spontaneous Speech,” inInter- speech. International Speech Communication Association, 9 2024, pp. 977–981

    2024

  26. [34]

    To BERT or Not To BERT: Comparing Speech and Language-based Approaches for Alzheimer’s Disease Detection,

    A. Balagopalan, B. Eyre, F. Rudzicz, J. Novikova, and K. Li, “To BERT or Not To BERT: Comparing Speech and Language-based Approaches for Alzheimer’s Disease Detection,” inInterspeech, 2020. [Online]. Available: http://dx.doi.org/10.21437/Interspeech.2020-2557

    2020 doi

  27. [35]

    Multi-modal fusion with gating using audio, lexical and disfluency features for Alzheimer’s Dementia recognition from spontaneous speech,

    M. Rohanian, J. Hough, and M. Purver, “Multi-modal fusion with gating using audio, lexical and disfluency features for Alzheimer’s Dementia recognition from spontaneous speech,” inInterspeech, 2020. [Online]. Available: http://dx.doi.org/10.21437/Interspeech.2020-2721

    2020 doi

  28. [36]

    Pre-trained Feature Fusion and Matching for Mild Cognitive Impairment Detection,

    J. Duan, F. Wei, H. Li, and J. Liu, “Pre-trained Feature Fusion and Matching for Mild Cognitive Impairment Detection,” 2024. [Online]. Available: https://github.com/facebookresearch/

    2024

  29. [37]

    Goodglass and E

    H. Goodglass and E. Kaplan,Boston diagnostic aphasia examination booklet. Lea\& Febiger, 1983

    1983

  30. [38]

    Attention Is All You Need,

    A. Vaswani, G. Brain, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention Is All You Need,” inNIPS, 2017. [Online]. Available: https://dl.acm.org/doi/10.5555/3295222.3295349

    2017

  31. [39]

    Neural Machine Translation by Jointly Learning to Align and Translate,

    D. Bahdanau, K. H. Cho, and Y . Bengio, “Neural Machine Translation by Jointly Learning to Align and Translate,”3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings, 9 2014. [Online]. Available: https://arxiv.org/abs/1409.0473v7

    2015 arXiv

  32. [40]

    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,” NAACL HLT 2019 - 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies - ...

    2019 arXiv

  33. [41]

    Improving Language Understanding by Generative Pre-Training,

    A. R. Openai, K. N. Openai, T. S. Openai, and I. S. Openai, “Improving Language Understanding by Generative Pre-Training,” 2018. [Online]. Available: https://gluebenchmark.com/leaderboard

    2018

  34. [42]

    Large language models in medicine,

    A. J. Thirunavukarasu, D. S. J. Ting, K. Elangovan, L. Gutierrez, T. F. Tan, and D. S. W. Ting, “Large language models in medicine,”Nature Medicine 2023 29:8, vol. 29, no. 8, pp. 1930–1940, 7 2023. [Online]. Available: https://www.nature.com/articles/s41591-023-02448-8

    2023

  35. [43]

    Diagnostic reasoning prompts reveal the potential for large language model interpretability in medicine,

    T. Savage, A. Nayak, R. Gallo, E. Rangan, and J. H. Chen, “Diagnostic reasoning prompts reveal the potential for large language model interpretability in medicine,”npj Digital Medicine 2024 7:1, vol. 7, no. 1, pp. 1–7, 1 2024. [Online]. Available: https://www.nature.com/articl...

    2024

  36. [44]

    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 on Selected...

    2021

  37. [45]

    LORA: LOW-RANK ADAPTATION OF LARGE LAN-GUAGE MODELS,

    E. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LORA: LOW-RANK ADAPTATION OF LARGE LAN-GUAGE MODELS,”arXiv, 2021. [Online]. Available: https://github.com/microsoft/LoRA

    2021

  38. [46]

    BEATs: Audio Pre-Training with Acoustic Tokenizers,

    S. Chen, Y . Wu, C. Wang, S. Liu, D. Tompkins, Z. Chen, W. Che, X. Yu, and F. Wei, “BEATs: Audio Pre-Training with Acoustic Tokenizers,” Proceedings of Machine Learning Research, vol. 202, pp. 4672–4712, 12 2022. [Online]. Available: https://arxiv.org/abs/2212.09058v1

    2022 arXiv

  39. [47]

    Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena,

    L. Zheng, W.-L. Chiang, Y . Sheng, S. Zhuang, Z. Wu, Y . Zhuang, Z. Lin, Z. Li, D. Li, E. P. Xing, H. Zhang, J. E. Gonzalez, and I. Stoica, “Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena,” in Neural Information Processing Systems, 6 2023. [Online]. Available: https://...

    2023 arXiv

  40. [48]

    AudioBench: A Universal Benchmark for Audio Large Language Models,

    B. Wang, X. Zou, G. Lin, S. Sun, Z. Liu, W. Zhang, Z. Liu, A. Aw, and N. F. Chen, “AudioBench: A Universal Benchmark for Audio Large Language Models,”arXiv, 6 2024. [Online]. Available: http://arxiv.org/abs/2406.16020

    2024 arXiv

  41. [49]

    V oiceBench: Benchmarking LLM-Based V oice Assistants,

    Y . Chen, X. Yue, C. Zhang, X. Gao, R. T. Tan, and H. Li, “V oiceBench: Benchmarking LLM-Based V oice Assistants,”arXiv, 2024. [Online]. Available: https://github.com/

    2024

  42. [50]

    Qwen2.5 Technical Report,

    Qwen, :, A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, H. Lin, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Lin, K. Dang, K. Lu, K. Bao, K. Yang, L. Yu, M. Li, M. Xue, P. Zhang, Q. Zhu, R. Men, R. Lin, T. Li, T. Tang, T. Xia, ...

  43. [2024]

    Available: https://arxiv.org/abs/2412.15115v2

    [Online]. Available: https://arxiv.org/abs/2412.15115v2

    arXiv

Pith tools

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