Pith. sign in

REVIEW 3 major objections 7 minor 31 references

Adaptable Non-parametric Approach for Speech-based Symptom Assessment: Isolating Private Medical Data in a Retrieval Datastore

T0 review · 3 major / 7 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 retrieval datastore of speech embeddings can assess respiratory symptoms as accurately as fine-tuned models, without training on medical labels.

desk verdict Plausible new application of retrieval datastores to speech symptom assessment, but 'competitive' needs speaker-disjoint splits and uncertainty bounds. read the letter →

arxiv 2506.22972 v1 pith:5PLPHDIY submitted 2025-06-28 eess.AS

Yu-Wen Chen , Julia Hirschberg This is my paper
classification eess.AS
keywords speech-basedsymptomassessmentnon-parametricretrievalself-supervisedspeechrepresentationsdatastoreprivacy-preservinghealthHuBERTmetadata-awarerespiratoryscreening
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 proposes NoNPSA, a framework that replaces a trained symptom classifier with a retrieval datastore: a pre-trained self-supervised speech model extracts embeddings from a general dataset, and medical assessment is done by nearest-neighbor voting over stored labeled samples. The claim is that this reaches ROC AUC 0.704 on COVID-19 Sounds and 0.723 on Coswara, competitive with fine-tuned SSL baselines, while keeping private medical data out of model parameters. The attraction, if correct, is that medical data can be added, removed, or enriched with new metadata simply by editing the datastore, which matters for privacy and for settings where data changes over time.

What carries the argument

The central object is the retrieval datastore: each training utterance is stored as a key (temporal mean of a HuBERT-large layer 3, 4, or 5 embedding, including a version from the time-reversed waveform) and a value (symptom label plus age and sex metadata). Retrieval is $L_2$ nearest-neighbor search over keys; segment-level features come from $k$-means time clustering, and metadata-aware refinement discards retrieved samples whose age group or sex does not match the input. The final score is the fraction of retrieved samples labeled symptomatic, averaged across the three layers.

What would settle it

Shuffle the symptom labels among the training recordings before building the datastore and rerun the same pipeline: if the ROC AUC stays near the reported values instead of collapsing to chance, the retrieval score is driven by speaker identity or recording conditions rather than by symptom status.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that a non-parametric pipeline can perform speech-based symptom assessment at a level competitive with fine-tuned SSL models: ROC AUC 0.704 on COVID-19 Sounds and 0.723 on Coswara, without ever training a classifier on symptom labels. The assessment is produced by retrieving nearest neighbors from a datastore of SSL embeddings, filtering by metadata, and counting symptomatic neighbors; adding or deleting a sample is a datastore edit rather than a retraining event. The paper also reports that segment-level retrieval contributes the strongest signal, that reversed-signal retrieval adds complementary sensitivity, and that metadata refinement helps on the larger dataset but not the smaller one.

Load-bearing premise

The load-bearing premise is that two speech recordings close in the pretrained embedding space share a symptom label, so $L_2$ nearest-neighbor voting over the datastore is a valid medical measurement.

Editorial extensions

If this is right

  • Symptom screening can be deployed without training on labeled medical data, reducing the privacy surface to the datastore itself.
  • Adding new patient data is a datastore insertion, and removing a patient's data is a deletion, with no retraining or certified-removal machinery needed.
  • New metadata attributes can be used at inference time by re-filtering retrieved samples instead of redesigning and retraining a model.
  • Because private labels live only in the datastore, the pretrained speech model itself can be shared openly without exposing medical information.

Reading between the lines

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

  • One extension the paper leaves implicit: if $L_2$ geometry in SSL embeddings carries symptom information, the same retrieval scaffold should transfer to other acoustic health targets such as cough type, vocal fatigue, or depression by simply swapping the datastore labels; this is testable on existing open datasets.
  • The reversed-signal results hint that the symptom signal is partly prosodic or voice-quality information that survives time reversal; a direct test would compare retrieval using spectral-envelope features against lexical features.
  • Because the datastore stores raw labeled samples rather than a compressed model, it offers a built-in audit trail for explaining why a particular screening decision was made.
  • Averaging scores from layers 3, 4, and 5 is a simple ensemble; weighting layers by validation AUC or including more layers could push performance higher, a variant the paper does not explore.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. This paper proposes NoNPSA, a non-parametric framework for speech-based symptom assessment that avoids training on medical labels. The method builds retrieval datastores from HuBERT-large features (layers 3, 4, 5) of the original and temporally reversed waveforms. At inference, it performs segment-level retrieval via k-means clustering of time steps and utterance-level retrieval, both with L2 nearest-neighbor search, applies metadata-aware filtering, and computes a symptom score as the proportion of retrieved symptomatic samples. Experiments on COVID-19 Sounds and Coswara report ROC AUCs of 0.704 and 0.723, respectively, which the authors characterize as competitive with fine-tuned SSL baselines. The paper also includes ablations on segment-level versus utterance-level retrieval, metadata refinement, and SSL layer selection.

Significance. The core idea—isolating private medical data in a retrieval datastore so that model parameters need not encode symptom labels—is timely and potentially impactful for healthcare applications where data updates and privacy are concerns. The method is simple, uses public feature extractors and standard nearest-neighbor search, and the pipeline appears reproducible from the description. If the reported performance holds under rigorous evaluation, the paper would demonstrate that non-parametric retrieval can match fine-tuned models on these tasks, which is a meaningful result. However, as discussed below, the current evaluation has substantial gaps in speaker-disjointness, uncertainty quantification, and selection-bias control, so the central claim is not yet established.

major comments (3)
  1. [Section 3.1] The random 70/15/15 split for Coswara is not speaker-disjoint. Coswara contains multiple recordings per participant (e.g., cough, breathing, counting, phonation). If the same participant appears in both the datastore and the test set, the L2 nearest neighbors of a test recording are likely to be other recordings from the same speaker, making the retrieved labels correct by speaker identity rather than by symptom status. This would inflate the reported 0.723 AUC. Please re-run with a speaker-disjoint split (e.g., split by participant ID) or quantify the contribution of same-speaker retrieval by comparing results with and without same-speaker matches removed.
  2. [Sections 4.4 and 4.3, Table 1] The layer selection ('based on the ablation study in Section 4.4') and the metadata-refinement choice ('Based on empirical findings in Section 4.3') appear to be made using the same data partition as the reported test results, but the text does not state whether those ablation curves are computed on the validation or test sets. If the test set was inspected to choose layers 3–5 and the age-aware refinement for COVID-19, the headline AUCs are optimistically biased by selection. In addition, Table 1 reports only point estimates; on COVID-19, the NoNPSA AUC (0.704) and the HuBERT-base baseline AUC (0.708) differ by less than 0.01, which is far too small to support 'competitive' without confidence intervals or significance tests. Please report bootstrap confidence intervals or repeated-run standard errors, and clarify the partition used for all ablation-driven choices.
  3. [Sections 2.1.2 and 2.2] The central assumption that L2 distance between mean-pooled HuBERT embeddings is a valid similarity for symptom status is only justified empirically. The paper does not demonstrate that retrieval is driven by symptom-related acoustic cues rather than by speaker identity, age, sex, recording device, or background noise. Concretely, a permutation test (randomizing labels in the datastore) or a comparison of within-speaker versus cross-speaker retrieval accuracy would indicate how much of the signal is attributable to symptoms. Without such a control, even a speaker-disjoint split may still reflect confounds like age or sex, which the metadata-aware refinement partly addresses but does not eliminate. Please include a label-permutation baseline or a confounding analysis.
minor comments (7)
  1. [Section 3.2] The fine-tuned parametric baselines use mean squared error loss with a dense output layer for a binary classification task; cross-entropy loss would be more standard. Please justify or switch to cross-entropy, and report the number of epochs, batch size, and random seeds.
  2. [Section 2.1.1 and Figure 2] The phrase 'k-means-based segment- and utterance-level features' is misleading because utterance-level features use temporal averaging, not k-means. Please rephrase.
  3. [Section 4.2] The claim that 'setting a threshold of 0.6 yields high specificity, exceeding 0.9' is not supported by a table or figure; either add the operating-point analysis or remove the statement.
  4. [Figure 5] The x-vector and Whisper results are only shown for COVID-19 (utterance-level). It would be helpful to include Coswara results as well, or explicitly state that the layer analysis was limited to COVID-19.
  5. [Table 1] The table header is missing; the meaning of the checkmarks under 'Non parametric' is ambiguous (they mark NoNPSA rows, but the label 'Non parametric' applies only to NoNPSA). Please clarify the table structure.
  6. [References] Reference [20] contains a garbled author name ('J Ch'); please fix. Also, the footnote for the HuBERT-large model URL is incomplete.
  7. [Section 4.1] The sentence 'While our method performs better with HuBERT-large compared to HuBERT-base (Section 4.4)' is in tension with the baseline result that fine-tuned HuBERT-large performs worse than HuBERT-base; this point could be clarified, since the comparison in Table 1 uses different models for NoNPSA and baselines.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: NoNPSA is an externally evaluated retrieval pipeline; the only self-citation is a non-load-bearing baseline recipe.

full rationale

NoNPSA is an empirical retrieval pipeline, not a derived quantity, so there is no chain in which an output is defined in terms of a fitted parameter or a self-citation. The assessment score is defined in Section 2.2 as the proportion of symptomatic samples among retrieved top-k neighbors and is evaluated on held-out test splits of COVID-19 Sounds and Coswara using the datasets' standard splits (Section 3.1), so the reported AUCs are computed on samples not used to construct the datastore or to choose n and k; the paper states that 'Both n and k are selected based on validation set performance.' HuBERT layers are chosen by ablation in Section 4.4, which is a hyperparameter-selection procedure, not a fitted parameter renamed as a prediction; if that ablation were run on the test partition it would be a selection-bias or transparency concern, not definitional circularity. The only self-citation, reference [24], is used to justify the parametric baseline's fine-tuning recipe ('using a strategy proven effective for speech assessment tasks [23, 24]'); it does not supply the central nonparametric result, and the baseline strategy is independently instantiated in reference [23]. The conclusion's caveat that generalization to other tasks remains unknown is a limitation, not a circular step. No quantity in the paper is equivalent to its input by construction, and no uniqueness or ansatz claim is imported from the authors' prior work.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The paper introduces no new scientific entities and rests on standard empirical assumptions plus several domain assumptions about SSL embeddings, datastore privacy, label reliability, and metadata utility. The main free parameters are the retrieval size k, the number of k-means segments n, the selected HuBERT layers, the reversed-datastore choice, and the metadata-refinement mode; all were tuned on validation or chosen after empirical ablations. These choices are legitimate tuning but should be reported with a dedicated held-out set or a selection-bias analysis.

free parameters (5)
  • Number of segments n (per layer and dataset) = COVID-19: n=2,73,73 for layers 3,4,5; Coswara: n=2 for all layers
    Set from validation performance and Silhouette score, affects the number of retrieved samples and segment resolution. Section 3.2 states n is selected based on validation set performance.
  • Retrieval neighborhood size k = k=5 for segment-level retrieval; utterance-level retrieves n*k samples
    Selected on validation set; controls how many datastore labels are averaged into the score.
  • SSL layer set = HuBERT-large layers 3, 4, and 5
    Selected from ablation study in Section 4.4, which was run as utterance-level retrieval on COVID-19; these choices are then applied to both datasets.
  • Metadata refinement mode = Age-aware refinement for COVID-19, no refinement for Coswara
    Chosen based on empirical findings in Section 4.3 rather than a fixed rule; affects which retrieved samples are kept.
  • Reversed-waveform datastore usage = Used at utterance level, skipped at segment level
    Included because ablation in Section 4.1 showed improved performance; another data-dependent design choice.
assumptions (7)
  • domain assumption Self-reported symptom labels in COVID-19 Sounds and Coswara are reliable ground truth for symptomatic/asymptomatic status.
    The method treats dataset labels as correct when computing assessment scores and performance metrics (Section 3.1).
  • domain assumption L2 distance in HuBERT embedding space is a meaningful similarity measure for symptom status.
    Section 2.1.2 uses L2 distance for retrieval; if embeddings do not group by symptom state, the score is uninformative.
  • domain assumption A general-purpose pretrained HuBERT model contributes no private information and its features are safe to store for retrieval.
    The privacy argument in Section 1 assumes the SSL model is public and does not encode private information from the target datasets.
  • domain assumption Speakers uttering the same sentence means semantic content is controlled, so acoustic-only retrieval is sufficient.
    Section 2.1.1 states the study focuses solely on acoustic characteristics because all speakers utter the same sentence; this excludes semantic confounds by design.
  • standard math k-means with Silhouette-based n selection produces segments that retain symptom-relevant acoustic detail.
    Section 2.2 relies on k-means clustering and the Silhouette score to define segment-level features; the choice is algorithmic but not proven optimal for symptom discrimination.
  • domain assumption Reversing the waveform preserves symptom-relevant acoustic cues while destroying semantics.
    Section 2.1.1 adds reversed-signal datastores on the assumption that reversed audio offers a useful alternative acoustic view without adding semantic information.
  • domain assumption Age and sex metadata in the datasets are accurate and relevant enough to filter retrieved samples.
    Metadata-aware refinement in Sections 2.2 and 4.3 depends on these labels matching the input's metadata.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptable Non-parametric Approach for Speech-based Symptom Assessment: Isolating Private Medical Data in a Retrieval Datastore." pith.science (2026). https://pith.science/paper/5PLPHDIY

@misc{pith2026250622972,
  author       = {Pith},
  title        = {Pith review of: Adaptable Non-parametric Approach for Speech-based Symptom Assessment: Isolating Private Medical Data in a Retrieval Datastore},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5PLPHDIY}},
  note         = {Machine review of arXiv:2506.22972}
}
read the original abstract

The automatic assessment of health-related acoustic cues has the potential to improve healthcare accessibility and affordability. Although parametric models are promising, they face challenges in privacy and adaptability. To address these, we propose a NoN-Parametric framework for Speech-based symptom Assessment (NoNPSA). By isolating medical data in a retrieval datastore, NoNPSA avoids encoding private information in model parameters and enables efficient data updates. A self-supervised learning (SSL) model pre-trained on general-purpose datasets extracts features, which are used for similarity-based retrieval. Metadata-aware refinement filters the retrieved data, and associated labels are used to compute an assessment score. Experimental results show that NoNPSA achieves competitive performance compared to fine-tuning SSL-based methods, while enabling greater privacy, update efficiency, and adaptability--showcasing the potential of non-parametric approaches in healthcare.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 28 canonical work pages

  1. [1]

    Adaptable Non-parametric Approach for Speech-based Symptom Assessment: Isolating Private Medical Data in a Retrieval Datastore

    INTRODUCTION Automatic evaluation of health-related acoustic cues can im- prove healthcare accessibility, affordability, and scalability by enabling remote, equipment-free assessments [1, 2]. Re- searchers have been increasingly exploring the potential of this field [3]. For example, studies have developed auto- mated methods to diagnose pertussis from co...

    work page Pith review arXiv 2025

  2. [2]

    Figure 2 shows an overview of NoNPSA

    NONPSA Instead of training parametric models using symptomatic (pri- vate) data, NoNPSA leverages SSL models pretrained on a general-purpose open-source dataset to build a retrieval data- store and uses the retrieved labels to calculate assessment scores. Figure 2 shows an overview of NoNPSA. 2.1. Retrieval datastore 2.1.1. Datastore building Training dat...

    work page

  3. [3]

    I hope my data can help manage the virus pandemic

    EXPERIMENTAL SETUP 3.1. Data We used two open-source datasets, COVID-19 Sounds [20] (hereafter referred to as COVID-19) and Coswara [21]. The COVID-19 dataset contains audio recordings of participants instructed to say the sentence, “I hope my data can help manage the virus pandemic.” Each recording is labeled as ei- ther symptomatic or asymptomatic. Symp...

    work page

  4. [4]

    Model performance Table 1 presents a performance comparison between NoNPSA and baseline parametric approaches

    RESULTS 4.1. Model performance Table 1 presents a performance comparison between NoNPSA and baseline parametric approaches. For the parametric meth- ods, we fine-tuned the SSL model using a strategy proven effective for speech assessment tasks [23, 24]. Specifically, the average pooling was applied to the SSL models’ out- put embeddings, and a dense outpu...

    work page

  5. [5]

    CONCLUSION We propose a novel non-parametric speech-based symptom assessment (NoNPSA) framework that takes a step toward adaptable and privacy-preserving health assessments while maintaining competitive performance compared to paramet- ric fine-tuning SSL-based methods. Our ablation study high- lights specific SSL model layers that optimize performance fo...

    work page

  6. [6]

    Mak- ing cough count in tuberculosis care,

    Alexandra J Zimmer, C ´esar Ugarte-Gil, Rahul Pathri, Puneet Dewan, Devan Jaganath, Adithya Cattamanchi, Madhukar Pai, and Simon Grandjean Lapierre, “Mak- ing cough count in tuberculosis care,” Communications medicine, vol. 2, no. 1, pp. 83, 2022

    work page 2022

  7. [7]

    Human-centred artificial intelligence for mobile health sensing: challenges and opportunities,

    Ting Dang, Dimitris Spathis, Abhirup Ghosh, and Ce- cilia Mascolo, “Human-centred artificial intelligence for mobile health sensing: challenges and opportunities,” Royal Society Open Science, vol. 10, no. 11, pp. 230806, 2023

    work page 2023

  8. [8]

    Exploring machine learning for audio-based respiratory condition screening: A concise review of databases, methods, and open issues,

    Tong Xia, Jing Han, and Cecilia Mascolo, “Exploring machine learning for audio-based respiratory condition screening: A concise review of databases, methods, and open issues,” Experimental Biology and Medicine, vol. 247, no. 22, pp. 2053–2061, 2022

    work page 2022

Show all 31 references
  1. [9]

    A cough-based algorithm for automatic diagnosis of pertussis,

    Renard Xaviero Adhi Pramono, Syed Anas Imtiaz, and Esther Rodriguez-Villegas, “A cough-based algorithm for automatic diagnosis of pertussis,” PloS one, vol. 11, no. 9, pp. e0162128, 2016

    2016

  2. [10]

    Piece- wise position encoding in convolutional neural network for cough-based COVID-19 detection,

    Jiakun Shen, Xueshuai Zhang, Pengyuan Zhang, Yonghong Yan, Shaoxing Zhang, Zhihua Huang, Yanfen Tang, Yu Wang, Fujie Zhang, and Aijun Sun, “Piece- wise position encoding in convolutional neural network for cough-based COVID-19 detection,” inProc. ICASSP

  3. [11]

    Exploring self-supervised representation ensembles for COVID-19 cough classifi- cation,

    Hao Xue and Flora D Salim, “Exploring self-supervised representation ensembles for COVID-19 cough classifi- cation,” in Proc. 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining 2021 , pp. 1944– 1952

    2021

  4. [12]

    HeAR–health acoustic representations,

    Sebastien Baur, Zaid Nabulsi, Wei-Hung Weng, Jake Garrison, Louis Blankemeier, Sam Fishman, Christina Chen, Sujay Kakarmath, Minyoi Maimbolwa, Nsala Sanjase, et al., “HeAR–health acoustic representations,” arXiv preprint arXiv:2403.02522, 2024

    2024 arXiv

  5. [13]

    WavRx: a disease-agnostic, generalizable, and privacy-preserving speech dealth di- agnostic model,

    Yi Zhu and Tiago Falk, “WavRx: a disease-agnostic, generalizable, and privacy-preserving speech dealth di- agnostic model,” IEEE Journal of Biomedical and Health Informatics, 2024

    2024

  6. [14]

    Multi-objective non- intrusive hearing-aid speech assessment model,

    Hsin-Tien Chiang, Szu-Wei Fu, Hsin-Min Wang, Yu Tsao, and John HL Hansen, “Multi-objective non- intrusive hearing-aid speech assessment model,” The Journal of the Acoustical Society of America , vol. 156, no. 5, pp. 3574–3587, 2024

    2024

  7. [15]

    A re- view on subjective and objective evaluation of synthetic speech,

    Erica Cooper, Wen-Chin Huang, Yu Tsao, Hsin-Min Wang, Tomoki Toda, and Junichi Yamagishi, “A re- view on subjective and objective evaluation of synthetic speech,” Acoustical Science and Technology, pp. e24– 12, 2024

    2024

  8. [16]

    What does it mean for a language model to preserve privacy?,

    Hannah Brown, Katherine Lee, Fatemehsadat Mireshghallah, Reza Shokri, and Florian Tram `er, “What does it mean for a language model to preserve privacy?,” in Proc. ACM FAccT 2022, pp. 2280–2292

    2022

  9. [17]

    Certified data removal from machine learning models,

    Chuan Guo, Tom Goldstein, Awni Hannun, and Laurens Van Der Maaten, “Certified data removal from machine learning models,” in Proc. ICML 2020, pp. 3832–3842

    2020

  10. [18]

    Reliable, adaptable, and attributable language models with retrieval,

    Akari Asai, Zexuan Zhong, Danqi Chen, Pang Wei Koh, Luke Zettlemoyer, Hannaneh Hajishirzi, and Wen-tau Yih, “Reliable, adaptable, and attributable language models with retrieval,” arXiv preprint arXiv:2403.03187, 2024

    2024 arXiv

  11. [19]

    SILO language models: Isolating legal risk in a nonparametric datastore,

    Sewon Min, Suchin Gururangan, Eric Wallace, Wei- jia Shi, Hannaneh Hajishirzi, Noah A Smith, and Luke Zettlemoyer, “SILO language models: Isolating legal risk in a nonparametric datastore,” in Proc. ICLR 2024

    2024

  12. [20]

    Generalization through memorization: Nearest neighbor language models,

    Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis, “Generalization through memorization: Nearest neighbor language models,” in Proc. ICLR 2020

    2020

  13. [21]

    RAMP: Retrieval-augmented MOS prediction via confidence-based dynamic weighting,

    Hui Wang, Shiwan Zhao, Xiguang Zheng, and Yong Qin, “RAMP: Retrieval-augmented MOS prediction via confidence-based dynamic weighting,” in Proc. INTER- SPEECH 2023, pp. 1095–1099

    2023

  14. [22]

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

    Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrah- man Mohamed, “HuBERT: Self-supervised speech rep- resentation learning by masked prediction of hidden units,” IEEE/ACM transactions on audio, speech, and language processin...

    2021

  15. [23]

    Least squares quantization in pcm,

    Stuart Lloyd, “Least squares quantization in pcm,”IEEE transactions on information theory , vol. 28, no. 2, pp. 129–137, 1982

    1982

  16. [24]

    Silhouettes: a graphical aid to the interpretation and validation of cluster analysis,

    Peter J Rousseeuw, “Silhouettes: a graphical aid to the interpretation and validation of cluster analysis,” Jour- nal of computational and applied mathematics, vol. 20, pp. 53–65, 1987

    1987

  17. [25]

    COVID- 19 sounds: a large-scale audio dataset for digital respi- ratory screening,

    Tong Xia, Dimitris Spathis, J Ch, Andreas Grammenos, Jing Han, Apinan Hasthanasombat, Erika Bondareva, Ting Dang, Andres Floto, Pietro Cicuta, et al., “COVID- 19 sounds: a large-scale audio dataset for digital respi- ratory screening,” in Proc. NeurIPS 2021 datasets and benchm...

    2021

  18. [26]

    Coswara: A respiratory sounds and symptoms dataset for remote screening of SARS-CoV-2 infection,

    Debarpan Bhattacharya, Neeraj Kumar Sharma, Debot- tam Dutta, Srikanth Raj Chetupalli, Pravin Mote, Sriram Ganapathy, C Chandrakiran, Sahiti Nori, KK Suhail, Sadhana Gonuguntla, et al., “Coswara: A respiratory sounds and symptoms dataset for remote screening of SARS-CoV-2 infe...

    2023

  19. [27]

    The faiss library,

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazar ´e, Maria Lomeli, Lucas Hosseini, and Herv ´e J´egou, “The faiss library,” arXiv preprint arXiv:2401.08281, 2024

    2024 arXiv

  20. [28]

    Generalization ability of MOS pre- diction networks,

    Erica Cooper, Wen-Chin Huang, Tomoki Toda, and Ju- nichi Yamagishi, “Generalization ability of MOS pre- diction networks,” in Proc. ICASSP 2022 , pp. 8442– 8446

    2022

  21. [29]

    InQSS: a speech intelligi- bility and quality assessment model using a multi-task learning network,

    Yu-Wen Chen and Yu Tsao, “InQSS: a speech intelligi- bility and quality assessment model using a multi-task learning network,” in Proc. INTERSPEECH 2022

    2022

  22. [30]

    Spoken language recognition using x-vectors.,

    David Snyder, Daniel Garcia-Romero, Alan McCree, Gregory Sell, Daniel Povey, and Sanjeev Khudanpur, “Spoken language recognition using x-vectors.,” in Odyssey, vol. 2018, pp. 105–111

    2018

  23. [31]

    Self-supervised models of audio effectively explain hu- man cortical responses to speech,

    Aditya R Vaidya, Shailee Jain, and Alexander G Huth, “Self-supervised models of audio effectively explain hu- man cortical responses to speech,” in Proc. ICML 2022

    2022

Pith tools

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