Evaluating the Effectiveness of Pre-Trained Audio Embeddings for Classification of Parkinson's Disease Speech Data

Cristian Tejedor-Garcia; Emmy Postma

REVIEW 3 major objections 4 minor 35 references

Evaluating the Effectiveness of Pre-Trained Audio Embeddings for Classification of Parkinson's Disease Speech Data

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that on the NeuroVoz corpus, pre-trained OpenL3 audio embeddings outperform VGGish and Wav2Vec2.0 embeddings for classifying Parkinson's disease from speech, and that only Wav2Vec2.0 shows a statistically significant…

desk verdict Useful head-to-head benchmark of pre-trained audio embeddings for PD speech, but the inferential claims are oversold due to post-hoc classifier selection and uncorrected multiple comparisons. read the letter →

arxiv 2506.02078 v1 pith:NZGTTC5P submitted 2025-06-02 eess.AS cs.AI

Emmy Postma , Cristian Tejedor-Garcia This is my paper

classification eess.AScs.AI

keywords Parkinson'sdiseasespeechclassificationpre-trainedaudioembeddingsOpenL3VGGishWav2Vec2.0genderbiasdeepacousticfeatureextraction

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 asks whether three widely used pre-trained audio embeddings—OpenL3, VGGish, and Wav2Vec2.0—differ in how well they support automatic classification of Parkinson's disease from speech. On the NeuroVoz corpus, the answer is yes: OpenL3, a model trained on musical audio rather than speech, gives the highest accuracy and AUC for both the diadochokinesis task and the listen-and-repeat task. The paper also finds that only Wav2Vec2.0, a speech-trained model, produces a statistically significant gender gap, favoring male speakers on the DDK task. A sympathetic reader would care because the result suggests that representations tuned to sound quality rather than linguistic content may be more useful for detecting the acoustic signatures of Parkinson's disease, and that supposedly neutral speech representations can encode gender bias.

What carries the argument

The machinery is a comparison pipeline: raw recordings are resampled per model, split into windows, and passed through a pre-trained embedding model (OpenL3, VGGish, or Wav2Vec2.0); per-recording embeddings are averaged over windows, then classified by SVM, KNN, or ERT. Speaker-independent nested cross-validation keeps the same participant in one fold. The paper's significance tests compare only the best-performing classifier per embedding model, while agreement between pipelines is measured with the Matthews correlation coefficient.

What would settle it

Run the same nine pipelines with matched preprocessing—same resampling rate, window length, and hop size across OpenL3, VGGish, and Wav2Vec2.0—and if the OpenL3 accuracy advantage disappears or reverses, the paper's central comparison would not hold. For the gender-bias claim, a direct audit of the gender distribution in Wav2Vec2.0's pretraining data would test whether the proposed male-heavy training mechanism explains the male-favoring AUC gap.

Watch

Extended reading notes

Core claim

The central claim is that on the NeuroVoz dataset, pre-trained OpenL3 embeddings are more effective for binary Parkinson's disease classification than VGGish or Wav2Vec2.0 embeddings, in both the DDK and LR tasks. In terms of accuracy, the best OpenL3 pipelines reach 0.85 on DDK and 0.82 on LR, versus 0.81 and 0.78 for VGGish and 0.71 and 0.76 for Wav2Vec2.0; in terms of AUC, OpenL3 reaches 0.97 on DDK and 0.90 on LR. The paper further claims that among the nine model-classifier pipelines, only the Wav2Vec2.0-SVM pipeline shows a statistically significant gender difference on DDK, with an AUC of 0.82 when trained on male voices versus 0.45 on female voices (p=0.01), indicating a clear bias toward male speakers. It also claims that the hardest misclassified cases trace to atypical speech patterns such as hypophonia, stuttering in healthy controls, and smooth or fast pronunciation in PD speakers.

Load-bearing premise

The central comparison assumes the three embedding models are directly comparable even though they use different input resampling rates, window sizes, and hop sizes, and it compares only each model's best classifier; if those differences drive the accuracy gaps, the conclusion that OpenL3 extracts more effective embeddings would not follow.

Editorial extensions

If this is right

OpenL3 embeddings are a strong default choice for PD speech classification on tasks like DDK and LR, at least on NeuroVoz.
The Wav2Vec2.0-SVM pipeline should not be used without checking gender balance, because it can silently favor male speakers on DDK.
Low agreement between models means combining embeddings or ensembling classifiers may improve robustness on hard speakers.
The advantage of OpenL3 supports the idea that PD-relevant features are more acoustic than lexical, since all participants repeated the same words.

Reading between the lines

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

Because the three models were not matched on resampling rate, window size, and hop size, an ablation with identical preprocessing is needed before concluding that OpenL3 is inherently superior.
If Wav2Vec2.0's bias comes from male-heavy pretraining data, gender-balanced pretraining or fine-tuning could reduce it; the paper notes that the gender distribution of two of its pretraining corpora is unknown.
The finding rests on 108 speakers; replicating on other corpora would show whether the model ordering and the gender gap generalize.

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.

Desk Editor's Note

read the letter

Quick take: this is a genuinely useful benchmark paper, but the inferential claims are oversold. The direct comparison of OpenL3, VGGish, and Wav2Vec2.0 on the same public NeuroVoz dataset is new, and the speaker-independent nested CV design is a real strength. The raw numbers tell a fairly consistent story: OpenL3 and VGGish beat Wav2Vec2.0 on both tasks by a wide margin, and the pairwise agreement analysis (MCC) is a nice addition. The gender-specific evaluation is also a worthy question.

The soft spots are mostly about statistics. The pairwise significance tests compare the best-performing classifier per model, selected after looking at the same outer folds. That makes the reported p-values tests of the max of several pipelines, not of the embeddings themselves—the p-values are anti-conservative even before considering that multiple comparisons across tasks and metrics go uncorrected. The gender-bias claim rests on one significant Welch t-test out of nine pipeline-wise comparisons, with a female AUC of 0.45 ± 0.20 from small folds; a Bonferroni correction would knock it over the threshold. The raw gap (0.82 vs 0.45) looks real, but the support is weaker than the text suggests. Also, the three embedding models are used with their default window/hop sizes and resampling rates, so the comparison is really about off-the-shelf pipelines rather than the embeddings alone—a reasonable practical question, but the language should not imply a controlled comparison of the models.

That said, none of this kills the paper. The raw accuracy gaps are large enough that the ordering probably holds up, and a pre-specified comparison protocol would likely confirm much of it. The manual misclassification analysis is qualitative but clearly labeled as such. The code is linked in a footnote; hyperparameters and a classical feature baseline should be added for completeness.

Bottom line: worth a serious referee, and if published after a stats revision it would be a useful reference for anyone building PD speech classifiers with pre-trained embeddings.

Referee Report

3 major / 4 minor

Summary. This paper compares three pre-trained audio embeddings (OpenL3, VGGish, and Wav2Vec2.0) for binary Parkinson's disease classification using NeuroVoz speech data from two tasks, diadochokinesis (DDK) and listen-and-repeat (LR). For each embedding, three classifiers (SVM, KNN, ERT) are trained under a speaker-independent nested 5-fold cross-validation, and performance is reported via ACC, AUC, and MCC. The authors claim that OpenL3 outperforms the other embeddings on both tasks, and that only Wav2Vec2.0 shows a statistically significant gender bias on DDK data, favoring male speakers. The paper further analyzes misclassified cases and discusses speech characteristics such as hypophonia and fluency as potential sources of difficulty.

Significance. If the headline claims were properly supported, the paper would offer a useful direct comparison of three widely used pre-trained audio embeddings on the same PD dataset, with a speaker-independent evaluation protocol and a gender-focused analysis. The use of nested cross-validation, multiple complementary metrics (ACC, AUC, MCC), and a publicly available code repository are strengths that increase the reproducibility of the empirical measurements. The manual inspection of misclassified cases is also a valuable qualitative contribution. However, the statistical inferential framework currently does not support the two main claims: the reported pairwise p-values are computed after post-hoc selection of the best classifier per model, and the gender-bias finding is one uncorrected test among nine. The raw performance gaps may be real, but the paper's conclusions need to be backed by a pre-specified comparison protocol or corrected inference.

major comments (3)

[Section 3.1, Table 1] The pairwise significance tests compare the best-performing classifier for each embedding model, but this best classifier is selected post hoc from the outer-fold results. For example, for DDK AUC the selected pipelines are OL3-ERT, VGG-SVM, and W2V2-KNN, while for DDK ACC they are OL3-SVM/ERT, VGG-SVM, and W2V2-SVM; different selections are made for LR. Because the same data are used to select and to test, the Welch t-tests are effectively tests of the maximum of three correlated pipelines, not of the embedding models themselves, making the p-values anti-conservative. In addition, the three pairwise tests within each metric are uncorrected for multiplicity across ACC/AUC and across DDK/LR. To support the claim that OL3 outperforms VGG and W2V2, the authors should pre-specify one classifier per embedding (or report the full set of comparisons with an appropriate correction), or re-frame the results as descriptive without relying on these p-values.
[Section 3.2, Figure 3] The gender-bias conclusion for W2V2-SVM on DDK data rests on a single Welch t-test result (T=3.76, df=5.64, p=0.01) selected from nine pipeline-wise comparisons. No multiple-testing correction is reported; applying a Bonferroni correction for nine tests leaves p=0.09, which is not significant at the 0.05 level. The female-group AUC of 0.45±0.20 is based on a small and possibly imbalanced gender split, so the estimate is unstable. Therefore, the claim that "only Wav2Vec2.0 shows significant gender bias" is not supported by the reported statistical analysis. A more robust approach would be to test a gender-by-pipeline interaction in a model that accounts for the nested CV structure, use permutation tests with correction for multiplicity, and report confidence intervals or effect sizes for the gender difference.
[Section 2.3 and Section 3.1] The comparison of the three embedding models is confounded with differences in input preprocessing: OL3 resamples to 48 kHz and uses 1-second windows with a 0.1-second hop; VGGish resamples to 16 kHz and uses a 0.96-second hop; W2V2 resamples to 16 kHz and the window/hop settings are not specified. The conclusion that "OL3 extracts the most effective embeddings" is therefore not directly supported, because the observed performance gaps may be driven by window length, hop size, or pooling choices rather than by the intrinsic quality of the embedding spaces. The authors should either match these settings across models or analyze the sensitivity of the results to these choices, or at minimum explicitly acknowledge this confound as a limitation of the comparison.

minor comments (4)

[Section 3.1] The statement that Shapiro-Wilk and Kolmogorov-Smirnov tests "did not indicate significant deviations from normality" is uninformative when each distribution consists of only five outer-fold values; such tests have very low power, and the result is expected under the null even if the true distribution is non-normal.
[Section 2.3] For the Wav2Vec2.0 model, the window size and hop size used for embedding extraction are not specified; please provide this information so that the comparison with OL3 and VGGish is interpretable.
[Figure 6 caption] The caption for Figure 6 appears to contain stray axis values ("1.1 1 0.9") in the text; please clean up the caption.
[General formatting] The phrase "This paper is accepted (camera-ready) for Interspeech 2025" appears within the manuscript text; camera-ready papers should normally not include such a note in the body.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the benchmark compares externally pre-trained embeddings on nested-CV folds; only minor non-load-bearing self-citation and post-hoc classifier selection qualify the inferential claims.

full rationale

This paper is an empirical comparison of three externally pre-trained embedding models (OpenL3, VGGish, Wav2Vec2.0) on NeuroVoz DDK and LR tasks. The central results are measurements made with a speaker-independent 5-fold nested cross-validation, using embeddings produced by models whose weights were trained outside this study; no parameter is fitted to the test folds and then renamed as a prediction, and no equation in the paper defines the outcome in terms of the input. The only self-citation (reference [2], co-authored by Tejedor-Garcia) is used for background context on deep acoustic feature extraction and gender considerations, and it does not carry the load of any claimed result. The post-hoc choice of the best-performing classifier per embedding model for pairwise t-tests and the absence of multiple-comparison correction are legitimate statistical-validity concerns, but they concern the strength of the evidence rather than a circular derivation: the reported accuracy and AUC values are still computed on held-out data, not implied by construction. Accordingly, no circular step is identified.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper is an empirical benchmark with no invented entities. The main free parameters are unreported classifier hyperparameters and embedding window/hop choices, which are not the target of the claim but affect the comparison. The key domain assumptions are the correctness of the NeuroVoz labels, the validity of average pooling, and the reliability of significance tests on five fold-wise scores.

free parameters (4)

SVM hyperparameters (C, kernel) = not reported
Chosen by hand or left at defaults; not specified in the paper, yet SVM is the best classifier for each embedding model in the main comparisons.
KNN hyperparameters (k, distance metric) = not reported
Not specified; one of the three classifiers evaluated, and its performance varies by embedding model.
ERT hyperparameters (n_estimators, max_features) = not reported
Not specified; one of the three classifiers evaluated, and its performance varies by embedding model.
Embedding window and hop sizes = OL3: 1s window/0.1s hop; VGGish: 0.96s hop; W2V2: unspecified
Chosen by the authors following prior work; differences across models may confound the comparison of embedding effectiveness.

assumptions (3)

domain assumption NeuroVoz diagnostic labels (PD vs HC) are correct and consistent
The entire evaluation treats the dataset labels as ground truth, as stated in Section 2.1.
domain assumption Average pooling of window embeddings preserves task-relevant information
A single recording-level vector is formed by averaging all window embeddings (Section 2.2), following Wilkinghoff [20]; if this discards discriminative temporal detail, the comparison is affected.
domain assumption The five outer folds of nested CV provide exchangeable samples for Welch t-tests
Statistical significance is computed over the five outer-fold scores as if they are independent and normally distributed; with only five values per pipeline, normality and variance estimates are fragile (Section 3.1).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Evaluating the Effectiveness of Pre-Trained Audio Embeddings for Classification of Parkinson's Disease Speech Data." pith.science (2026). https://pith.science/paper/NZGTTC5P

@misc{pith2026250602078,
  author       = {Pith},
  title        = {Pith review of: Evaluating the Effectiveness of Pre-Trained Audio Embeddings for Classification of Parkinson's Disease Speech Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NZGTTC5P}},
  note         = {Machine review of arXiv:2506.02078}
}

read the original abstract

Speech impairments are prevalent biomarkers for Parkinson's Disease (PD), motivating the development of diagnostic techniques using speech data for clinical applications. Although deep acoustic features have shown promise for PD classification, their effectiveness often varies due to individual speaker differences, a factor that has not been thoroughly explored in the existing literature. This study investigates the effectiveness of three pre-trained audio embeddings (OpenL3, VGGish and Wav2Vec2.0 models) for PD classification. Using the NeuroVoz dataset, OpenL3 outperforms others in diadochokinesis (DDK) and listen and repeat (LR) tasks, capturing critical acoustic features for PD detection. Only Wav2Vec2.0 shows significant gender bias, achieving more favorable results for male speakers, in DDK tasks. The misclassified cases reveal challenges with atypical speech patterns, highlighting the need for improved feature extraction and model robustness in PD detection.

Figures

Figures reproduced from arXiv: 2506.02078 by the authors.

**Figure 4.** displays the AUC scores for all pipelines trained on LR data, including the full dataset, male-only data and femaleonly data. T-tests indicated no significant performance differences between models trained on embeddings from male and female voices. These findings suggest that LR speech task data results in comparable model performance across genders. 1.1 1 0.9 0.8 0.7 0.6 0.5 0.4 Model-Classifier [PITH_FULL_IMAGE… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages

[1]

Introduction Recent research has established speech as a critical and cost - effective biomarker of Parkinson’s disease (PD) [1, 2, 3]. Given that speech impairments are common symptoms of PD and of- ten emerge early in the disease progression, researchers have increasingly utilized speech data in machine learning (ML) ap- proaches to develop models capab...

work page 2025
[2]

pa -ta-ka

Methodology 2.1. Dataset We used the publicly available NeuroVoz1 dataset [19], which includes voice recordings and PD -related metadata from 108 Castilian Spanish speakers —53 with PD and 55 healthy con - trols (HC). Metadata includes gender, age, vocal and cephalic tremors, mandibular tremor, sialorrhea, dysphagia, hypophonic voice, time since diagnosis...

work page
[3]

Models Performance To answer RQ1, we compared the performances of all three embedding model (OL3, VGG, W2V2) and classifier (SVM, KNN, ERT) combinations on both DDK and LR data

Results and Discussion 3.1. Models Performance To answer RQ1, we compared the performances of all three embedding model (OL3, VGG, W2V2) and classifier (SVM, KNN, ERT) combinations on both DDK and LR data. Ta- ble 1 presents the mean and standard deviation (SD) of the ACC and AUC scores across the five outer folds of the nested -CV. To evaluate potential ...

work page arXiv 2025
[4]

Conclusions The findings of this study highlight the overall superior perfor- mance of OL3 pre-trained audio embeddings for PD classifica- tion using speech data from DDK and LR tasks from the Neu - roVoz dataset. Furthermore, despite OL3 and VGG achieving a high ACC, their low MCC agreement indicates that there is a clear speaker variance in PD data due ...

work page
[5]

The authors of this work would also like to thank Prof

Acknowledgements This publication is part of the project Responsible AI for Voice Diagnostics (RAIVD) with file number NGF.1607.22.013 of the research programme NGF AiNed Fellowship Grants which is financed by the Dutch Research Council (NWO). The authors of this work would also like to thank Prof. Dr. Martha Larson for her insightful feedback regarding t...

work page
[6]

Computerized analysis of speech and voice for parkinson’s disease: A systematic review,

Q. C. Ngo, M. A. Motin, N. D. Pah, P. Drota´r, P. Kempster, and D. Kumar, “Computerized analysis of speech and voice for parkinson’s disease: A systematic review,” Computer Methods and Programs in Biomedicine, vol. 226, p. 107133, 2022

work page 2022
[7]

Innovative Speech - Based Deep Learning Approaches for Parkinson’s Disease Classification: A Systematic Review,

L. van Gelderen and C. Tejedor -Garcia, “Innovative Speech - Based Deep Learning Approaches for Parkinson’s Disease Classification: A Systematic Review,” Applied Sciences, vol. 14, no. 17, p. 7873, 2024

work page 2024
[8]

Multilingual evalua- tion of interpretable biomarkers to represent language and speech patterns in parkinson’s disease,

A. Favaro, L. Moro-Vela´zquez, A. Butala, C. Motley, T. Cao, R. D. Stevens, J. Villalba, and N. Dehak, “Multilingual evalua- tion of interpretable biomarkers to represent language and speech patterns in parkinson’s disease,” Frontiers in Neurology, vol. 14, p. 1142642, 2023

work page 2023

Show all 35 references

[9]

Applied machine learning techniques to diagnose voice-affecting conditions and disorders: Systematic literature review,

A. Idrisoglu, A. L. Dallora, P. Anderberg, and J. S. Berglund, “Applied machine learning techniques to diagnose voice-affecting conditions and disorders: Systematic literature review,” J Med In- ternet Res, vol. 25, p. e46105, Jul 2023

2023
[10]

Clinical diagnostic accuracy of parkin- son’s disease: Where do we stand?

S. Virameteekul, T. Revesz, Z. Jaunmuktane, T. T. Warner, and E. De Pablo-Ferna´ndez, “Clinical diagnostic accuracy of parkin- son’s disease: Where do we stand?” Movement Disorders, vol. 38, no. 4, p. 558–566, Jan 2023

2023
[11]

Deep acoustic embeddings for identifying Parkinsonian Speech,

Z. S. Syed, S. Ali, and A. Latif, “Deep acoustic embeddings for identifying Parkinsonian Speech,” International Journal of Ad - vanced Computer Science and Applications, vol. 11, no. 10, 2020

2020
[12]

Multi- class voice disorder classification using OpenL3 -SVM,

X. Peng, H. Xu, J. Liu, J. Wang, and C. He, “Multi- class voice disorder classification using OpenL3 -SVM,” 2022, SSRN Electronic Journal. Last accessed: 2025-02-10. [Online]. Available: https://doi.org/10.2139/ssrn.4047840

2022 doi
[13]

Poster: Vggish embeddings based au- dio classifiers to improve parkinson’s disease diagnosis,

S. Kurada and A. Kurada, “Poster: Vggish embeddings based au- dio classifiers to improve parkinson’s disease diagnosis,” in 2020 IEEE/ACM International Conference on Connected Health: Ap - plications, Systems and Engineering Technologies (CHASE), Vir- tual Event, 2020, pp. 9–11

2020
[14]

Interpretable speech features vs. dnn embeddings: What to use in the automatic assessment of parkinson’s disease in multi-lingual scenarios,

A. Favaro, Y.-T. Tsai, A. Butala, T. Thebaud, J. Villalba, N. De- hak, and L. Moro -Vela´zquez, “Interpretable speech features vs. dnn embeddings: What to use in the automatic assessment of parkinson’s disease in multi-lingual scenarios,” Computers in Bi- ology and Medicine, v...

2023
[15]

Evaluating the Per- formance of wav2vec Embedding for Parkinson’s Disease Detec- tion,

O. Klemp´ıˇr, D. Pˇr´ıhoda, and R. Krupicˇka, “Evaluating the Per- formance of wav2vec Embedding for Parkinson’s Disease Detec- tion,” Measurement Science Review, vol. 23, no. 6, pp. 260–267, 2023

2023
[16]

Machine learning approaches to identify parkinson’s disease using voice signal features,

R. Alshammri, G. Alharbi, E. Alharbi, and I. Almubark, “Machine learning approaches to identify parkinson’s disease using voice signal features,” Frontiers in Artificial Intelligence, vol. 6, 2023

2023
[17]

Gender -Related Pat - terns of Dysprosody in Parkinson Disease and Correlation Be - tween Speech Variables and Motor Symptoms,

S. Skodda, W. Visser, and U. Schlegel, “Gender -Related Pat - terns of Dysprosody in Parkinson Disease and Correlation Be - tween Speech Variables and Motor Symptoms,” Journal of Voice, vol. 25, no. 1, pp. 76–82, 2011

2011
[18]

Speech as a biomarker for disease detection,

C. Botelho, A. Abad, T. Schultz, and I. Trancoso, “Speech as a biomarker for disease detection,” IEEE Access , vol. 12, pp. 184 487–184 508, 2024

2024
[19]

Parkinson’s disease and parkinsonism in a longitudinal study: two-fold higher incidence in men,

M. Baldereschi, A. Di Carlo, W. A. Rocca, P. Vanni, S. Maggi, E. Perissinotto, F. Grigoletto, L. Amaducci, and D. Inzitari, “Parkinson’s disease and parkinsonism in a longitudinal study: two-fold higher incidence in men,” Neurology, vol. 55, no. 9, pp. 1358–1363, 2000

2000
[20]

Gender differences in parkinson’s disease,

C. A. Haaxma, B. R. Bloem, G. F. Borm, W. J. Oyen, K. L. Leenders, S. Eshuis, J. Booij, D. E. Dluzen, and M. W. Horstink, “Gender differences in parkinson’s disease,” Journal of Neurol- ogy, Neurosurgery & Psychiatry , vol. 78, no. 8, pp. 819 –824, 2007

2007
[21]

X-vectors: New quantitative biomarkers for early parkin - son’s disease detection from speech,

L. Jeancolas, D. Petrovska -Delacre´taz, G. Mangone, B. -E. Benkelfat, J.-C. Corvol, M. Vidailhet, S. Lehe´ricy, and H. Be - nali, “X-vectors: New quantitative biomarkers for early parkin - son’s disease detection from speech,” Frontiers in Neuroinformat- ics, vol. 15, Feb 2021

2021
[22]

Gender differences in parkinson’s disease: clinical characteristics and cognition,

I. N. Miller and A. Cronin -Golomb, “Gender differences in parkinson’s disease: clinical characteristics and cognition,” Move- ment Disorders, vol. 25 (16), p. 2695–2703, 2010

2010
[23]

Analysis of voice as an assisting tool for detection of parkinson’s disease and its subse - quent clinical interpretation,

G. Solana-Lavalle and R. Rosas-Romero, “Analysis of voice as an assisting tool for detection of parkinson’s disease and its subse - quent clinical interpretation,” Biomedical Signal Processing and Control, vol. 66, p. 102415, 2021

2021
[24]

NeuroVoz: a Castillian Spanish corpus of parkinsonian speech,

J. Mendes-Laureano, J. A. Go´mez-Garc´ıa, A. Guerrero-Lo´pez, E. Luque-Buzo, J. D. Arias-London˜o, F. J. Grandas-Pe´rez, and J. I. Godino-Llorente, “NeuroVoz: a Castillian Spanish corpus of parkinsonian speech,” Scientific Data (Nature), vol. 11, no. 1, p. 1367, 2024

2024
[25]

On open -set classification with l3 -net embed- dings for machine listening applications,

K. Wilkinghoff, “On open -set classification with l3 -net embed- dings for machine listening applications,” in EUSIPCO, 2021, pp. 800–804

2021
[26]

Audio - based deep learing frameworks for detecting covid-19,

D. Ngo, L. Pham, T. Hoang, S. Kolozali, and D. Jarchi, “Audio - based deep learing frameworks for detecting covid-19,” in Euro- pean Signal Processing Conference, Belgrade, Serbia, 2022, pp. 1–5, eUSIPCO

2022
[27]

Speech emotion recognition based on two-stream deep learning model using korean audio information,

A.-H. Jo and K.-C. Kwak, “Speech emotion recognition based on two-stream deep learning model using korean audio information,” Applied Sciences, vol. 13, no. 4, 2023

2023
[28]

How robust are audio embeddings for polyphonic sound event tagging?

J. Abeßer, S. Grollmisch, and M. Mu¨ller, “How robust are audio embeddings for polyphonic sound event tagging?” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 2658–2667, 2023

2023
[29]

MLS: A Large-Scale Multilingual Dataset for Speech Research,

V. Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert, “MLS: A Large-Scale Multilingual Dataset for Speech Research,” in Interspeech 2020 , ser. interspeech 2020. Shanghai, China: ISCA, Oct. 2020, pp. 2757–2761

2020
[30]

Common voice: A massively-multilingual speech corpus,

R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. Hen - retty, R. Morais, L. Saunders, F. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,” in Proceedings of the Twelfth Language Resources and Evaluation Conference . Marseille, France: Europea...

2020
[31]

Speech recog- nition and keyword spotting for low -resource languages: Babel project research at cued,

M. J. Gales, K. M. Knill, A. Ragni, and S. P. Rath, “Speech recog- nition and keyword spotting for low -resource languages: Babel project research at cued,” in SLTU. St. Petersburg, Russia: ISCA, 2014, pp. 16–23

2014
[32]

Analyz - ing the potential of pre-trained embeddings for audio classifica - tion tasks,

S. Grollmisch, E. Cano, C. Kehling, and M. Taenzer, “Analyz - ing the potential of pre-trained embeddings for audio classifica - tion tasks,” in 2020 28th European Signal Processing Conference (EUSIPCO), Amsterdam, The Netherlands, 2021, pp. 790–794

2020
[33]

Pro - gression of voice and speech impairment in the course of parkin- son’s disease: A longitudinal study,

S. Skodda, W. Gro¨nheit, N. Mancinelli, and U. Schlegel, “Pro - gression of voice and speech impairment in the course of parkin- son’s disease: A longitudinal study,” Parkinson’s Disease, vol. 2013, no. 1, p. 389195, 2013

2013
[34]

Parkinson disease prediction using intrinsic mode function based features from speech signal,

B. Karan, S. S. Sahu, and K. Mahto, “Parkinson disease prediction using intrinsic mode function based features from speech signal,” Biocybernetics and Biomedical Engineering , vol. 40, no. 1, pp. 249–264, 2020

2020
[35]

Automatic detection of laryngeal pathologies in records of sustained vowels by means of mel-frequency cepstral coefficient parameters and differentiation of patients by sex,

R. Fraile, N. Sa´enz-Lecho´n, J. Godino-Llorente, V. Osma-Ruiz, and C. Fredouille, “Automatic detection of laryngeal pathologies in records of sustained vowels by means of mel-frequency cepstral coefficient parameters and differentiation of patients by sex,” Fo- lia Phoniatric...

2009

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Introduction Recent research has established speech as a critical and cost - effective biomarker of Parkinson’s disease (PD) [1, 2, 3]. Given that speech impairments are common symptoms of PD and of- ten emerge early in the disease progression, researchers have increasingly utilized speech data in machine learning (ML) ap- proaches to develop models capab...

work page 2025

[2] [2]

pa -ta-ka

Methodology 2.1. Dataset We used the publicly available NeuroVoz1 dataset [19], which includes voice recordings and PD -related metadata from 108 Castilian Spanish speakers —53 with PD and 55 healthy con - trols (HC). Metadata includes gender, age, vocal and cephalic tremors, mandibular tremor, sialorrhea, dysphagia, hypophonic voice, time since diagnosis...

work page

[3] [3]

Models Performance To answer RQ1, we compared the performances of all three embedding model (OL3, VGG, W2V2) and classifier (SVM, KNN, ERT) combinations on both DDK and LR data

Results and Discussion 3.1. Models Performance To answer RQ1, we compared the performances of all three embedding model (OL3, VGG, W2V2) and classifier (SVM, KNN, ERT) combinations on both DDK and LR data. Ta- ble 1 presents the mean and standard deviation (SD) of the ACC and AUC scores across the five outer folds of the nested -CV. To evaluate potential ...

work page arXiv 2025

[4] [4]

Conclusions The findings of this study highlight the overall superior perfor- mance of OL3 pre-trained audio embeddings for PD classifica- tion using speech data from DDK and LR tasks from the Neu - roVoz dataset. Furthermore, despite OL3 and VGG achieving a high ACC, their low MCC agreement indicates that there is a clear speaker variance in PD data due ...

work page

[5] [5]

The authors of this work would also like to thank Prof

Acknowledgements This publication is part of the project Responsible AI for Voice Diagnostics (RAIVD) with file number NGF.1607.22.013 of the research programme NGF AiNed Fellowship Grants which is financed by the Dutch Research Council (NWO). The authors of this work would also like to thank Prof. Dr. Martha Larson for her insightful feedback regarding t...

work page

[6] [6]

Computerized analysis of speech and voice for parkinson’s disease: A systematic review,

Q. C. Ngo, M. A. Motin, N. D. Pah, P. Drota´r, P. Kempster, and D. Kumar, “Computerized analysis of speech and voice for parkinson’s disease: A systematic review,” Computer Methods and Programs in Biomedicine, vol. 226, p. 107133, 2022

work page 2022

[7] [7]

Innovative Speech - Based Deep Learning Approaches for Parkinson’s Disease Classification: A Systematic Review,

L. van Gelderen and C. Tejedor -Garcia, “Innovative Speech - Based Deep Learning Approaches for Parkinson’s Disease Classification: A Systematic Review,” Applied Sciences, vol. 14, no. 17, p. 7873, 2024

work page 2024

[8] [8]

Multilingual evalua- tion of interpretable biomarkers to represent language and speech patterns in parkinson’s disease,

A. Favaro, L. Moro-Vela´zquez, A. Butala, C. Motley, T. Cao, R. D. Stevens, J. Villalba, and N. Dehak, “Multilingual evalua- tion of interpretable biomarkers to represent language and speech patterns in parkinson’s disease,” Frontiers in Neurology, vol. 14, p. 1142642, 2023

work page 2023