pith. sign in

arxiv: 2602.14671 · v2 · pith:VDWE2ZIFnew · submitted 2026-02-16 · 📡 eess.AS

Data Augmentation for Pathological Speech Enhancement

Mingchi Hou , Enno Hermann , Ina Kodrasi This is my paper

Pith reviewed 2026-05-25 06:48 UTC · model grok-4.3

classification 📡 eess.AS
keywords data augmentationspeech enhancementpathological speechParkinson's diseasenoise augmentationtransformative augmentationgenerative modelspredictive models
0
0 comments X

The pith

Noise augmentation delivers the largest gains for speech enhancement of pathological speech, while generative augmentation can degrade results with increasing data volumes.

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

The paper systematically evaluates data augmentation techniques to enhance the performance of speech enhancement models on the atypical speech produced by individuals with Parkinson's disease. Results indicate that noise augmentation provides the most consistent benefits across objective metrics, followed by moderate gains from transformative methods, whereas generative augmentation offers limited help and may reduce performance when excessive synthetic samples are added. These effects are more pronounced in predictive models than in generative ones, although the overall quality for pathological speech still does not reach that of neurotypical speech. This suggests that data augmentation can mitigate some data scarcity issues but does not fully resolve the challenges posed by atypical acoustic characteristics.

Core claim

Experimental results show that noise augmentation consistently delivers the largest and most robust gains, transformative augmentations provide moderate improvements, while generative augmentation yields limited benefits and can harm performance as the amount of synthetic data increases. Furthermore, the effectiveness of DA varies depending on the SE model, with DA being more beneficial for predictive SE models. While our results demonstrate that DA improves SE performance for pathological speakers, a performance gap between neurotypical and pathological speech persists, highlighting the need for future research on targeted DA strategies for pathological speech.

What carries the argument

Categories of data augmentation (transformative, generative, noise) applied to predictive and generative speech enhancement models on pathological speech datasets.

If this is right

  • Noise augmentation is the most effective strategy among the tested categories.
  • Generative augmentation should be used cautiously as performance can decline with more synthetic data.
  • Data augmentation benefits predictive speech enhancement models more than generative ones.
  • Targeted strategies are required to address the remaining performance gap for pathological speech.

Where Pith is reading between the lines

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

  • If objective metrics align poorly with listener preferences for pathological voices, the ranking of augmentation methods could change.
  • Real-time applications like hearing aids for Parkinson's patients could incorporate noise-based augmentation to improve usability.
  • Future experiments might combine noise and limited generative augmentations to achieve better results than either alone.

Load-bearing premise

The chosen objective metrics for speech enhancement accurately proxy the perceptual quality experienced by listeners of pathological speech.

What would settle it

A controlled listening experiment in which human raters prefer the output of models trained with high amounts of generative augmentation over those trained with noise augmentation.

Figures

Figures reproduced from arXiv: 2602.14671 by Enno Hermann, Ina Kodrasi, Mingchi Hou.

Figure 1
Figure 1. Figure 1: ∆PESQ (left) and ∆fwSSNR (right) for pathological speakers using the CR model with strategies from three DA categories at different augmentation ratios. Within each category, separate plots correspond to individual strategies. For reference, the baseline performance without any DA strategy is also shown. trained on neurotypical speech and therefore do not closely resemble the reference pathological speech … view at source ↗
read the original abstract

The performance of state-of-the-art speech enhancement (SE) models considerably degrades for pathological speech due to atypical acoustic characteristics and limited data availability. This paper systematically investigates data augmentation (DA) strategies to improve SE performance for pathological speakers affected by Parkinson`s disease, evaluating both predictive and generative SE models. We examine three DA categories, i.e., transformative, generative, and noise augmentation, assessing their impact with objective SE metrics. Experimental results show that noise augmentation consistently delivers the largest and most robust gains, transformative augmentations provide moderate improvements, while generative augmentation yields limited benefits and can harm performance as the amount of synthetic data increases. Furthermore, we show that the effectiveness of DA varies depending on the SE model, with DA being more beneficial for predictive SE models. While our results demonstrate that DA improves SE performance for pathological speakers, a performance gap between neurotypical and pathological speech persists, highlighting the need for future research on targeted DA strategies for pathological speech.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper systematically evaluates three categories of data augmentation (transformative, generative, and noise) to improve speech enhancement performance on pathological speech from Parkinson's disease speakers. It reports that noise augmentation yields the largest and most robust gains on objective SE metrics, transformative methods provide moderate benefits, and generative augmentation offers limited gains that can degrade performance with increasing synthetic data volume. DA is more effective for predictive SE models than generative ones, though a performance gap versus neurotypical speech remains.

Significance. If the empirical ranking holds under appropriate validation, the work offers actionable guidance on DA choices for low-resource pathological SE, an area with clear clinical relevance. The systematic cross-category comparison and model-type interaction analysis are strengths; the explicit acknowledgment of the remaining neurotypical-pathological gap is also constructive.

major comments (1)
  1. [Results / Experimental evaluation] Results section: the central ranking (noise > transformative > generative) rests entirely on objective SE metrics without any reported correlation analysis, listening tests, or validation showing these metrics track perceptual quality for Parkinsonian speech exhibiting atypical prosody, tremor, or breathiness. This is load-bearing because the metrics were developed on neurotypical speech; if the correlation is weak or reversed, the reported ordering does not support the headline conclusion about augmentation effectiveness.
minor comments (1)
  1. [Abstract] Abstract and experimental description omit concrete dataset sizes, speaker counts, exact metric definitions, and statistical testing procedures; these details should be stated explicitly even if present in the full text.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the major comment on metric validation below.

read point-by-point responses

  1. Referee: [Results / Experimental evaluation] Results section: the central ranking (noise > transformative > generative) rests entirely on objective SE metrics without any reported correlation analysis, listening tests, or validation showing these metrics track perceptual quality for Parkinsonian speech exhibiting atypical prosody, tremor, or breathiness. This is load-bearing because the metrics were developed on neurotypical speech; if the correlation is weak or reversed, the reported ordering does not support the headline conclusion about augmentation effectiveness.

    Authors: We acknowledge that standard objective metrics (PESQ, STOI, etc.) were developed on neurotypical speech and that their correlation with perceptual quality for Parkinsonian speech (with atypical prosody, tremor, or breathiness) is not validated in our work. Our experiments demonstrate consistent trends across multiple objective metrics, and the manuscript already notes the persistent gap to neurotypical performance. We did not include correlation analysis or listening tests, as the study focused on objective evaluation of DA strategies. We will revise the discussion section to explicitly address this limitation, caution against over-interpreting the ranking for perceptual quality, and identify subjective validation as important future work. This will strengthen the paper without altering the reported objective results. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical comparison study

full rationale

The paper reports experimental results comparing data augmentation strategies (transformative, generative, noise) for speech enhancement on Parkinsonian speech, using objective SE metrics. No mathematical derivations, equations, or predictions are presented that could reduce claims to fitted inputs or self-citations by construction. All findings are direct empirical observations from model training and evaluation, making the work self-contained against external benchmarks with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Only the abstract is available, so the ledger is populated from stated premises in the abstract alone; no free parameters, invented entities, or non-standard axioms are visible.

axioms (2)
  • domain assumption Pathological speech (Parkinson's) exhibits atypical acoustic characteristics that degrade standard SE models
    Explicitly stated in the first sentence of the abstract as the motivation for the work.
  • domain assumption Limited data availability is the primary reason for poor SE performance on pathological speech
    Stated alongside atypical acoustics as the reason performance degrades.

pith-pipeline@v0.9.0 · 5688 in / 1245 out tokens · 35908 ms · 2026-05-25T06:48:47.862578+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages

  1. [1]

    Speech Enhancement—A Review of Modern Meth- ods,

    D. O’Shaughnessy, “Speech Enhancement—A Review of Modern Meth- ods,”IEEE Trans. Human-Mach. Syst., vol. 54, no. 1, pp. 110–120, 2024

    work page 2024

  2. [2]

    Supervised Speech Separation Based on Deep Learning: An Overview,

    D. Wang and J. Chen, “Supervised Speech Separation Based on Deep Learning: An Overview,”IEEE/ACM Trans. Audio, Speech, Language Process., vol. 26, no. 10, pp. 1702–1726, Oct. 2018

    work page 2018

  3. [3]

    Speech Enhancement and Dereverberation with Diffusion-Based Gen- erative Models,

    J. Richter, S. Welker, J.-M. Lemercier, B. Lay, and T. Gerkmann, “Speech Enhancement and Dereverberation with Diffusion-Based Gen- erative Models,”IEEE/ACM Trans. Audio, Speech, Language Process., vol. 31, pp. 2351–2364, 2023

    work page 2023

  4. [4]

    Schr ¨odinger Bridge for Generative Speech Enhancement,

    A. Juki ´c, R. Korostik, J. Balam, and B. Ginsburg, “Schr ¨odinger Bridge for Generative Speech Enhancement,” inProc. Interspeech, 2024, pp. 1175–1179

    work page 2024

  5. [5]

    Conditional Diffusion Probabilistic Model for Speech Enhancement,

    Y .-J. Lu et al., “Conditional Diffusion Probabilistic Model for Speech Enhancement,” inProc. IEEE Int. Conf. Acoust., Speech, Signal Pro- cess., 2022, pp. 7402–7406

    work page 2022

  6. [6]

    The Effect of Training Dataset Size on Discriminative and Diffusion-Based Speech Enhancement Systems,

    P. Gonzalez, Z.-H. Tan, J. Østergaard, J. Jensen, T. S. Alstrøm, and T. May, “The Effect of Training Dataset Size on Discriminative and Diffusion-Based Speech Enhancement Systems,”IEEE Signal Process- ing Letters, vol. 31, pp. 2225–2229, 2024

    work page 2024

  7. [7]

    Variational Autoencoder for Personalized Pathological Speech Enhancement,

    M. Hou and I. Kodrasi, “Variational Autoencoder for Personalized Pathological Speech Enhancement,” inProc. Eur . Signal Process. Conf., 2025, pp. 116–120

    work page 2025

  8. [8]

    Generalizability of Predictive and Generative Speech Enhancement Models to Pathological Speakers,

    M. Hou, A. Jukic, and I. Kodrasi, “Generalizability of Predictive and Generative Speech Enhancement Models to Pathological Speakers,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process., 2026

    work page 2026

  9. [9]

    Differential diagnostic patterns of dysarthria,

    F. L. Darley, A. E. Aronson, and J. R. Brown, “Differential diagnostic patterns of dysarthria,”J. Speech Hearing Research, vol. 12, no. 2, pp. 246–269, Jun. 1969

    work page 1969

  10. [10]

    Statistical Modeling of Speech Spectral Coefficients in Patients with Parkinson’s Disease,

    I. Kodrasi and H. Bourlard, “Statistical Modeling of Speech Spectral Coefficients in Patients with Parkinson’s Disease,” inProc. 13th ITG- Symposium in Speech Communication, 2018

    work page 2018

  11. [11]

    Spectro-Temporal Sparsity Characterization for Dysarthric Speech Detection,

    ——, “Spectro-Temporal Sparsity Characterization for Dysarthric Speech Detection,”IEEE/ACM Trans. Audio, Speech, Language Pro- cess., vol. 28, pp. 1210–1222, 2020

    work page 2020

  12. [12]

    Neurological disorders: public health challenges,

    WHO, “Neurological disorders: public health challenges,” 2006

    work page 2006

  13. [13]

    Dysarthric speech database for universal access research,

    H. Kim, M. Hasegawa-Johnson, A. Perlman, J. Gunderson, T. S. Huang, K. Watkin, and S. Frame, “Dysarthric speech database for universal access research,” inProc. Interspeech, 2008, pp. 1741–1744

    work page 2008

  14. [14]

    The TORGO database of acoustic and articulatory speech from speakers with dysarthria,

    F. Rudzicz, A. K. Namasivayam, and T. Wolff, “The TORGO database of acoustic and articulatory speech from speakers with dysarthria,” Language Resources and Evaluation, pp. 523–541, 2012

    work page 2012

  15. [15]

    New Spanish speech corpus database for the analysis of people suffering from Parkinson’s disease,

    J. R. Orozco-Arroyave, J. D. Arias-Londo ˜no, J. F. Vargas-Bonilla, M. C. Gonz´alez-R´ativa, and E. N¨oth, “New Spanish speech corpus database for the analysis of people suffering from Parkinson’s disease,” inProc. Int. Conf. Lang. Resour . Eval, 2014, pp. 342–347

    work page 2014

  16. [16]

    On Using the UA-Speech and Torgo Databases to Validate Automatic Dysarthric Speech Classification Approaches,

    G. Schu, P. Janbakhshi, and I. Kodrasi, “On Using the UA-Speech and Torgo Databases to Validate Automatic Dysarthric Speech Classification Approaches,” inProc. IEEE Int. Conf. Acoust., Speech, Signal Process., 2023

    work page 2023

  17. [17]

    Data augmentation techniques for Automatic Speech Recognition: Taxonomy, method analysis, challenges, and future research directions,

    M. A. Kermanshahi, A. A. Azirani, B. Nasersharif, and S. J. Kabudian, “Data augmentation techniques for Automatic Speech Recognition: Taxonomy, method analysis, challenges, and future research directions,” Computers and Electrical Engineering, vol. 130, p. 110851, 2026

    work page 2026

  18. [18]

    Data augmentation for speech separation,

    A. Alex, L. Wang, P. Gastaldo, and A. Cavallaro, “Data augmentation for speech separation,”Speech Communication, vol. 152, p. 102949, 2023

    work page 2023

  19. [19]

    A Comprehensive Analysis of Data Augmentation Methods for Speech Emotion Recognition,

    U. Avci, “A Comprehensive Analysis of Data Augmentation Methods for Speech Emotion Recognition,”IEEE Access, vol. 13, pp. 111 647– 111 669, 2025

    work page 2025

  20. [20]

    A comparison of data augmentation methods in voice pathology detection,

    F. Javanmardi, S. R. Kadiri, and P. Alku, “A comparison of data augmentation methods in voice pathology detection,”Computer Speech & Language, vol. 83, p. 101552, 2024

    work page 2024

  21. [21]

    Data Augmentation and Loss Normalization for Deep Noise Suppression,

    S. Braun and I. Tashev, “Data Augmentation and Loss Normalization for Deep Noise Suppression,” inProc. 22nd Int. Conf. Speech, Computer, 2020, pp. 79–86

    work page 2020

  22. [22]

    SpecMix : A Mixed Sample Data Aug- mentation method for Training with Time-Frequency Domain Features,

    G. Kim, D. K. Han, and H. Ko, “SpecMix : A Mixed Sample Data Aug- mentation method for Training with Time-Frequency Domain Features,” inProc. Interspeech, 2021

    work page 2021

  23. [23]

    Generative Data Augmentation Challenge: Zero-Shot Speech Synthesis for Personalized Speech Enhancement,

    J.-S. Bae, A. Kuznetsova, D. Manocha, J. Hershey, T. Kristjansson, and M. Kim, “Generative Data Augmentation Challenge: Zero-Shot Speech Synthesis for Personalized Speech Enhancement,” inProc. IEEE Int. Conf. Acoust., Speech, Signal Process., 2025

    work page 2025

  24. [24]

    XTTS: a Massively Multilingual Zero-Shot Text-to-Speech Model,

    E. Casanova, K. Davis, E. G ¨olge, G. G ¨oknar, I. Gulea, L. Hart, A. Aljafari, J. Meyer, R. Morais, S. Olayemi, and J. Weber, “XTTS: a Massively Multilingual Zero-Shot Text-to-Speech Model,” inProc. Interspeech, 2024, pp. 4978–4982

    work page 2024

  25. [25]

    YourTTS: Towards Zero-Shot Multi-Speaker TTS and Zero-Shot V oice Conversion for Everyone,

    E. Casanova, J. Weber, C. D. Shulby, A. C. Junior, E. G ¨olge, and M. A. Ponti, “YourTTS: Towards Zero-Shot Multi-Speaker TTS and Zero-Shot V oice Conversion for Everyone,” inProc. Int. Conf. Mach. Learn., 2022, pp. 2709–2720

    work page 2022

  26. [26]

    StoRM: A Diffusion-Based Stochastic Regeneration Model for Speech Enhance- ment and Dereverberation,

    J.-M. Lemercier, J. Richter, S. Welker, and T. Gerkmann, “StoRM: A Diffusion-Based Stochastic Regeneration Model for Speech Enhance- ment and Dereverberation,”IEEE/ACM Trans. Audio, Speech, Language Process., vol. 31, pp. 2724–2737, 2023

    work page 2023

  27. [27]

    Influence of Clean Speech Characteristics on Speech Enhancement Performance,

    M. Hou and I. Kodrasi, “Influence of Clean Speech Characteristics on Speech Enhancement Performance,” inProc. IEEE Int. Conf. Acoust., Speech, Signal Process., 2026

    work page 2026

  28. [28]

    Communication impairment in Parkin- son’s disease: Impact of motor and cognitive symptoms on speech and language,

    K. M. Smith and D. N. Caplan, “Communication impairment in Parkin- son’s disease: Impact of motor and cognitive symptoms on speech and language,”Brain and Language, pp. 38–46, 2018

    work page 2018

  29. [29]

    An investigation into the influences of age, pathology and cognition on speech production,

    A. Lowit, B. Brendel, C. Dobinson, and P. Howell, “An investigation into the influences of age, pathology and cognition on speech production,”J Med Speech Lang Pathol., vol. 14, pp. 253–262, 2006

    work page 2006

  30. [30]

    Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech,

    J. Kim, J. Kong, and J. Son, “Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech,” inProc. Int. Conf. on Mach. Learn., 2021, pp. 5530–5540

    work page 2021

  31. [31]

    HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis,

    J. Kong, J. Kim, and J. Bae, “HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis,” inProc. Neural Inf. Process. Syst., 2020, pp. 17 022–17 033

    work page 2020

  32. [32]

    The third ‘CHiME’ speech separation and recognition challenge: Dataset, task and base- lines,

    J. Barker, R. Marxer, E. Vincent, and S. Watanabe, “The third ‘CHiME’ speech separation and recognition challenge: Dataset, task and base- lines,” inProc. IEEE Autom. Speech Recognit. Understanding, 2015, pp. 504–511

    work page 2015

  33. [33]

    CSR-I (WSJ0) Complete LDC93S6A,

    Garofolo, John S.et al., “CSR-I (WSJ0) Complete LDC93S6A,” Philadelphia: Linguistic Data Consortium, 1993

    work page 1993

  34. [34]

    Generative modeling by estimating gradients of the data distribution,

    Y . Song and S. Ermon, “Generative modeling by estimating gradients of the data distribution,” inProc. Neural Inf. Process. Syst., 2019

    work page 2019

  35. [35]

    CML-TTS: A Multilingual Dataset for Speech Synthesis in Low-Resource Languages,

    F. S. Oliveira, E. Casanova, A. C. Junior, A. S. Soares, and A. R. Galv˜ao Filho, “CML-TTS: A Multilingual Dataset for Speech Synthesis in Low-Resource Languages,” inProc. Text, Speech, and Dialogue, 2023, pp. 188–199

    work page 2023

  36. [36]

    Common V oice: A Massively-Multilingual Speech Corpus,

    R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. Henretty, R. Morais, L. Saunders, F. Tyers, and G. Weber, “Common V oice: A Massively-Multilingual Speech Corpus,” inProc. Conf. Lang. Resour . Eval, 2020, pp. 4218–4222

    work page 2020

  37. [37]

    Perceptual evaluation of speech quality (PESQ)-a new method for speech quality assessment of telephone networks and codecs,

    A. W. Rix, J. G. Beerends, M. P. Hollier, and A. P. Hekstra, “Perceptual evaluation of speech quality (PESQ)-a new method for speech quality assessment of telephone networks and codecs,” inProc. IEEE Int. Conf. Acoust., Speech, Signal Process., 2001, pp. 749–752

    work page 2001

  38. [38]

    Evaluation of Objective Quality Measures for Speech Enhancement,

    Y . Hu and P. C. Loizou, “Evaluation of Objective Quality Measures for Speech Enhancement,”IEEE Trans. Audio, Speech, Language Process., vol. 16, no. 1, pp. 229–238, Jan. 2008

    work page 2008