Pith. sign in

REVIEW 2 major objections 4 minor 54 references

Regularized Federated Learning for Privacy-Preserving Dysarthric and Elderly Speech Recognition

T0 review · 2 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 Regularized federated learning, and especially a new loss-based regularizer, reduces word error rates on dysarthric and elderly speech by up to 0.55% absolute over standard FedAvg.

desk verdict Useful benchmark on FL for pathological speech, but the headline WER gains are likely inflated by test-set selection—still worth a referee. read the letter →

arxiv 2506.11069 v1 pith:NU2YC54I submitted 2025-06-02 eess.AS cs.AIcs.CLcs.SD

Tao Zhong , Mengzhe Geng , Shujie Hu , Guinan Li , Xunying Liu This is my paper
classification eess.AScs.AIcs.CLcs.SD
keywords federatedlearningdysarthricspeechrecognitionelderlyregularizationKLdivergenceFedAvgWERprivacy-preservingprocessing
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 argues that regularized federated learning can make privacy-preserving speech recognition work better for two hard populations: people with dysarthria and elderly speakers. Using a HuBERT-based ASR model trained with FedAvg, it compares three levels of regularization—on model parameters, on intermediate embeddings, and on output losses—and finds that all improve over plain FedAvg. The strongest and most consistent gains come from a new loss-based regularizer that aligns each client's predicted distribution with pseudo-logits produced by a frozen global model. On the UASpeech dysarthric corpus the best systems reduce WER by up to 0.54% absolute, and on DementiaBank Pitt by up to 0.55% absolute, with both improvements reported as statistically significant. The paper also shows that raising communication frequency to one exchange per batch narrows the gap to centralized training.

What carries the argument

The load-bearing object is the loss-based regularization term $R_{\mathrm{loss}}$ of Eq. (3), a KL divergence between a client's output distribution and the output distribution of the frozen global model evaluated on the client's own embeddings. It acts as a form of self-distillation from the global model into each local model, keeping local predictions aligned with the global consensus without transmitting any data or labels. Parameter-based regularization uses a squared L2 distance to the previous global parameters, and embedding-based regularization uses a squared L2 distance between local and aggregated embeddings; the paper treats these as lower-level alternatives. Regularization can be inserted after different Transformer blocks, and the best configurations combine the two best-performing positions or all three techniques.

What would settle it

Re-run the UASpeech and DementiaBank Pitt experiments with the regularization position and weight selection performed on a separate validation set, and check whether the reported 0.54% and 0.55% absolute WER reductions over FedAvg remain statistically significant at $\alpha = 0.05$.

Watch

Extended reading notes

Core claim

The central claim is that in federated fine-tuning of a large self-supervised speech model on dysarthric and elderly speech, regularization at the output level outperforms regularization at the parameter or embedding level. The proposed loss-based term $R_{\mathrm{loss}} = D_{\mathrm{KL}}(\hat{y}_i \| \tilde{y}_i^l)$ takes the local client's intermediate embedding, passes it through the frozen global model from the previous round to obtain pseudo-logits, and penalizes the KL divergence between those pseudo-logits and the client's own output distribution. Across four intelligibility groups on UASpeech and across participant and investigator speakers on DementiaBank Pitt, the regularized systems beat the unregularized FedAvg baseline with statistically significant overall WER reductions, and combining parameter-, embedding-, and loss-based regularization gives the best results. The paper further claims that synchronizing after every local batch makes federated learning approach centralized-training performance while keeping data local.

Load-bearing premise

The headline gains assume that the regularization positions and penalty weights, which were chosen by looking at the same evaluation sets used to report the results, are not inflating the measured improvement.

Editorial extensions

If this is right

  • Loss-based regularization is a client-side-only addition, so it introduces no extra communication cost over standard FedAvg.
  • Regularized FL systems close part of the WER gap to centralized training, making privacy-preserving training more viable for clinical speech data.
  • Combining parameter-, embedding-, and loss-based regularization yields the best overall WER, suggesting the three target complementary sources of client drift.
  • Increasing communication frequency to one exchange per batch approaches centralized performance on both datasets, at a 9x training-time cost.
  • The loss-based regularizer's benefit is consistent across communication frequencies except the one-batch case, where unregularized and regularized systems converge.

Reading between the lines

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

  • The KL-to-global-pseudo-logits mechanism is close to teacher-student distillation; a natural extension is applying it to on-device personalization or to other federated medical tasks beyond speech.
  • Because the one-batch regime already approaches centralized performance, the practical bottleneck for healthcare FL may be communication cost rather than algorithmic drift; reducing that cost is a promising axis.
  • If the gains are partly tuning artifacts, the technique still likely helps, but the magnitude of the reported headline numbers would need to be confirmed with an independent validation split.
  • The method could be tested on other self-supervised speech models, such as wav2vec 2.0, to see whether the loss-based regularizer transfers across model families.
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

2 major / 4 minor

Summary. The paper investigates regularized federated learning for privacy-preserving ASR on dysarthric and elderly speech. It compares three regularization techniques applied to a HuBERT-based FedAvg system: parameter-based regularization (FedProx-style), embedding-based regularization at intermediate Transformer layers, and a newly proposed loss-based regularization that aligns local output distributions with global-model pseudo-logits via KL divergence. Experiments on UASpeech (dysarthric) and DementiaBank Pitt (elderly) report consistent WER reductions over unregularized FedAvg, with statistically significant gains up to 0.54% absolute on UASpeech and 0.55% absolute on DementiaBank Pitt, and show that increasing communication frequency to one exchange per batch approaches centralized training performance.

Significance. If the reported gains are unbiased, this paper provides a useful systematic comparison of regularization techniques for FL-based ASR in an under-studied healthcare domain, and the proposed loss-based regularization is a plausible contribution. The paper's strengths include the use of public benchmarks (UASpeech, DementiaBank Pitt), a standard HuBERT model, statistical significance testing, and ablations over regularization positions, combinations, and communication frequency. However, the headline gain magnitudes and significance flags are vulnerable to a test-set selection bias: regularization positions and combination weights were chosen using the same evaluation sets on which the gains are reported. The qualitative direction of the result is plausible and consistently positive, but the exact magnitudes and the associated statistical significance claims need to be re-established under a protocol that separates model selection from evaluation.

major comments (2)
  1. [Section 4.3, Table 1 footnotes 6-7; Section 4.2] The regularization positions and combination weights are selected on the same evaluation sets used for reporting. Footnote 6 states that Sys.7 and Sys.13 integrate the 'two best-performing positions' from Sys.3-6 and Sys.9-12, respectively, and footnote 7 says the combined-system weights are set empirically; Section 4.2 similarly states that penalty weights are empirically set. On UASpeech, B2 serves as both the selection criterion and the test set, so Sys.7, Sys.13, Sys.14, and Sys.15 are maxima over a small family of configurations rather than unbiased estimates of the regularization effect. The MAPSSWE p-values in Tables 1-2 are computed for a fixed system and do not account for this selection step or for multiple comparisons. This inflates the headline 0.54-0.55% absolute WER gains and the dagger significance flags. Please re-run the selection on a held-out validation set (e.g., DementiaBank dev, or an internal split of the UASpeech training data) and report both selected and validation/test results, or otherwise adjust the significance procedure to account for selection.
  2. [Section 4.2] All results are single-run, and no code, random seed information, or variance estimates are provided. Since the reported improvements are small (about 0.2-0.5% absolute WER), it is impossible to assess from the paper alone whether the differences are stable across training runs. Please provide multi-seed means and standard deviations, or at least release code and exact seed configurations to allow reproducibility; this is particularly important because the significance claims in Tables 1-2 rely on differences of this magnitude.
minor comments (4)
  1. [Equation (2)] The notation in Eq. (2) appears to use the same symbol for the local embedding and the aggregated reference embedding; the second term should be written with a bar or a distinct symbol (e.g., \bar{e}_i^l) to match the surrounding text and Figure 2.
  2. [Section 4.1] There is a typo in 'B1 and B3 are used as for training'; please correct to 'used for training'.
  3. [Figure 4] The legend in Figure 4 lists 'FedAvg' twice, which makes it difficult to distinguish the grey lines; please relabel the entries to identify each regularization method clearly.
  4. [Section 4.4] The claim that loss-based regularization provides a 'consistent and statistically significant' improvement across all communication frequencies except the 1-batch case should be supported by the same selection-corrected significance analysis requested above, since the figure appears to compare selected configurations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the regularization losses are standard, and the headline WER comparisons are independent evaluations on public corpora.

full rationale

The paper's central claim is empirical: regularized FL variants compared against a FedAvg baseline on the UASpeech and DementiaBank Pitt corpora. The regularizers in Eqs. (1)-(3) are standard L2-norm and KL-divergence losses; no equation is defined in terms of the reported WER, and no fitted parameter is renamed as a prediction. The baseline and comparison systems are trained and evaluated on external public benchmarks, not derived from the authors' prior results. Self-citations such as [3,4,9,10,12] provide background and prior context, but the comparative WER result does not reduce to any of them; there is no uniqueness theorem or load-bearing self-citation chain. The footnote about 'two best-performing positions' (Table 1, notes 6 and 7) indicates some hyperparameter selection was performed using evaluation-set results, which is a legitimate concern about selection bias and about the reliability of the reported significance levels. However, it is not a circular derivation: the reported WERs are not equal to the selection criterion by construction, and the qualitative conclusion that regularization helps is supported across multiple independent configurations. Therefore no circular step meets the evidentiary standard of the hard rules.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard FL and ASR machinery plus a few empirical choices not justified upstream: test-set-tuned penalty weights and positions, a single pretrained backbone, a single Pitt partition, and a self-referential teacher for the loss-based regularization. No new entities are postulated. The free parameters are the regularization weights and the discrete position choices; the axioms are the transferability of HuBERT, the validity of MAPSSWE for the significance claims, the representativeness of the client splits, and the usefulness of the global model as a teacher.

free parameters (5)
  • parameter-based penalty weight = 0.01
    Empirically set in Section 4.2 for Eq. (1); no held-out validation protocol is described.
  • embedding-based penalty weight = 0.001
    Empirically set in Section 4.2 for Eq. (2).
  • loss-based penalty weight = 0.01
    Empirically set in Section 4.2 for Eq. (3).
  • combined-system penalty weights = 0.1, 0.1, 1.0
    Set for Sys.15 (Table 1 footnote 7), chosen with reference to test-set performance.
  • regularization position selection = {6,12,18,24}; best two positions per test set
    Sys.7 and Sys.13 combine the two best-performing positions among Sys.3-6 and Sys.9-12 respectively (Table 1 footnote 6), a test-set-based discrete hyperparameter choice.
assumptions (4)
  • domain assumption HuBERT-large self-supervised features fine-tuned on 960 hours of LibriSpeech transfer sufficiently to dysarthric and elderly speech when fine-tuned with CTC.
    Section 4.2 model configuration: the entire comparison uses one pretrained backbone; the claimed regularization gains are demonstrated only within this transfer setup.
  • standard math MAPSSWE paired significance test at alpha=0.05 is a valid and sufficient tool for the reported significance claims.
    Section 4.2. The test assesses paired test-set differences only; it does not capture training-run stochasticity, so 'significant' does not imply stability across random seeds.
  • domain assumption The client partitioning scheme (one client per UASpeech speaker; one random 10-way split of 244 Pitt conversation pairs) is representative of real federated healthcare deployments.
    Section 4.2 client partitioning. The Pitt split is a single random draw, never repeated, and the results could shift with a different grouping.
  • domain assumption The previous-round global model is a useful frozen teacher for pseudo-logit generation in loss-based regularization.
    Section 3.3 loss-based regularization. The teacher is an aggregate of the clients' own previous updates; no ablation isolates the teacher's quality, so part of the observed gain may depend on it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Regularized Federated Learning for Privacy-Preserving Dysarthric and Elderly Speech Recognition." pith.science (2026). https://pith.science/paper/NU2YC54I

@misc{pith2026250611069,
  author       = {Pith},
  title        = {Pith review of: Regularized Federated Learning for Privacy-Preserving Dysarthric and Elderly Speech Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NU2YC54I}},
  note         = {Machine review of arXiv:2506.11069}
}
read the original abstract

Accurate recognition of dysarthric and elderly speech remains challenging to date. While privacy concerns have driven a shift from centralized approaches to federated learning (FL) to ensure data confidentiality, this further exacerbates the challenges of data scarcity, imbalanced data distribution and speaker heterogeneity. To this end, this paper conducts a systematic investigation of regularized FL techniques for privacy-preserving dysarthric and elderly speech recognition, addressing different levels of the FL process by 1) parameter-based, 2) embedding-based and 3) novel loss-based regularization. Experiments on the benchmark UASpeech dysarthric and DementiaBank Pitt elderly speech corpora suggest that regularized FL systems consistently outperform the baseline FedAvg system by statistically significant WER reductions of up to 0.55\% absolute (2.13\% relative). Further increasing communication frequency to one exchange per batch approaches centralized training performance.

Figures

Figures reproduced from arXiv: 2506.11069 by the authors.

Figure 2
Figure 2. Schematic of (a) parameter-based regularization (up￾per right) and (b) embedding-based regularization (lower right) applied to the FedAvg-based HuBERT ASR system. Compo￾nents with darker colors are on the central server side, while lighter-colored components are on the local client side. 3.2. Embedding-based regularization Building on [38,39], during each communication round, we extract latent embeddings from the lo… view at source ↗
Figure 1
Figure 1. Illustration of federated learning based HuBERT [44] ASR system. Following FedAvg [16], each communication round involves: (a) aggregating the parameters of locally trained client-side models into a global model using a weighted sum, and (b) redistributing the global model to clients. The widely used aggregation strategy for FL-based ASR is FedAvg [16], which computes global parameters as a weighted sum of client pa… view at source ↗
Figure 3
Figure 3. Schematic of the proposed loss-based regularization applied to the FedAvg-based HuBERT ASR system. For simplic￾ity, one client is shown as an example. Darker-colored compo￾nents represent the global model, while lighter-colored compo￾nents correspond to the local model. The embedding-based regularization and loss-based regu￾larization can be applied at multiple positions in the model. Furthermore, as these three reg… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Impact of communication frequency on the perfor￾mance (WER%) of federated learning. “bt.” and “ep.” stand for batch and epoch. × denotes a stat. significant improve￾ment (α = 0.05) over the baseline FedAvg system. The dashed line represents the performance of centraliz…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 51 canonical work pages

  1. [12]

    Speaker Adaptation Using Spectro- Temporal Deep Features for Dysarthric and Elderly Speech Recognition,

    M. Geng, X. Xieet al., “Speaker Adaptation Using Spectro- Temporal Deep Features for Dysarthric and Elderly Speech Recognition,”IEEE/ACM T-ASLP, 2022

    work page 2022

  2. [1]

    Introduction While automatic speech recognition (ASR) technologies tar- geting normal speech have advanced rapidly over the past decades [1, 2], accurate speech recognition in the healthcare domain, particularly for dysarthric and elderly speakers, re- mains highly challenging to date [3–13]. Dysarthric and el- derly speech introduces fundamental challeng...

    work page Pith review arXiv 2025

  3. [2]

    However, such a design inherently results in heterogeneous data distributions across clients

    Federated Learning Based ASR Unlike traditional distributed training [43] which first aggre- gates data from multiple sources and then redistributes the data among learners, federated learning (FL) preserves data locality to comply with privacy constraints. However, such a design inherently results in heterogeneous data distributions across clients. As il...

    work page

  4. [3]

    ##=2$%(45&||75&!) !

    Regularized Federated Learning Federated learning inherently introduces data heterogeneity among clients, which is further compounded by the challenges of data scarcity, data imbalance, and speaker diversity widely observed in dysarthric and elderly speech. To mitigate this, three regularization techniques at different levels are investi- 1The impact of c...

    work page

  5. [4]

    very low

    Experiments and Results 4.1. Task description The English UASpeechcorpus [41] is the largest publicly available and widely used dataset for dysarthric speech recog- nition. It comprises an isolated word recognition task with ap- proximately 103 hours of speech data from 29 speakers, among whom 16 are dysarthric speakers and 13 are healthy control speakers...

    work page

  6. [5]

    Conclusions This paper systematically investigates regularized FL tech- niques for privacy-preserving dysarthric and elderly speech recognition, i.e., parameter-, embedding-, and novel loss-based regularizations. Experiments on UASpeech and DementiaBank Pitt show that regularized FL systems consistently outperform FedAvg, while increasing communication fr...

    work page

  7. [6]

    14200220, 14200021, 14200324 and Innovation Technology Fund grant No

    Acknowledgements This research is supported by Hong Kong RGC GRF grant No. 14200220, 14200021, 14200324 and Innovation Technology Fund grant No. ITS/218/21

    work page

  8. [7]

    Speech-Transformer: A No-Recurrence Sequence-to-Sequence Model for Speech Recognition,

    L. Dong, S. Xuet al., “Speech-Transformer: A No-Recurrence Sequence-to-Sequence Model for Speech Recognition,” in ICASSP, 2018

    work page 2018

Show all 54 references
  1. [8]

    Conformer: Convolution-augmented Transformer for Speech Recognition,

    A. Gulati, J. Qinet al., “Conformer: Convolution-augmented Transformer for Speech Recognition,” inINTERSPEECH, 2020

    2020

  2. [9]

    Development of the CUHK Elderly Speech Recognition System for Neurocognitive Disorder Detection Using the Dementiabank Corpus,

    Z. Ye, S. Huet al., “Development of the CUHK Elderly Speech Recognition System for Neurocognitive Disorder Detection Using the Dementiabank Corpus,” inICASSP, 2021

    2021

  3. [10]

    Recent Progress in the CUHK Dysarthric Speech Recognition System,

    S. Liuet al., “Recent Progress in the CUHK Dysarthric Speech Recognition System,”IEEE/ACM T-ASLP, 2021

    2021

  4. [11]

    Conformer Based Elderly Speech Recog- nition System for Alzheimer’s Disease Detection,

    T. Wang, J. Denget al., “Conformer Based Elderly Speech Recog- nition System for Alzheimer’s Disease Detection,” inINTER- SPEECH, 2022

    2022

  5. [13]

    Acoustic modelling from raw source and filter components for dysarthric speech recognition,

    Z. Yue, E. Loweimiet al., “Acoustic modelling from raw source and filter components for dysarthric speech recognition,” IEEE/ACM T-ASLP, 2022

    2022

  6. [14]

    Speaker adaptation for Wav2vec2 based dysarthric ASR,

    M. K. Baskar, T. Herziget al., “Speaker adaptation for Wav2vec2 based dysarthric ASR,” inINTERPSEECH, 2022

    2022

  7. [15]

    Self-Supervised ASR Models and Features for Dysarthric and Elderly Speech Recognition,

    S. Hu, X. Xieet al., “Self-Supervised ASR Models and Features for Dysarthric and Elderly Speech Recognition,”IEEE/ACM T- ASLP, 2024

    2024

  8. [16]

    Personalized Adversarial Data Augmenta- tion for Dysarthric and Elderly Speech Recognition,

    Z. Jin, M. Genget al., “Personalized Adversarial Data Augmenta- tion for Dysarthric and Elderly Speech Recognition,”IEEE/ACM T-ASLP, 2023

    2023

  9. [17]

    Enhancing Pre-trained ASR System Fine- tuning for Dysarthric Speech Recognition using Adversarial Data Augmentation,

    H. Wang, Z. Jinet al., “Enhancing Pre-trained ASR System Fine- tuning for Dysarthric Speech Recognition using Adversarial Data Augmentation,” inICASSP, 2024

    2024

  10. [18]

    A Cluster-based Personalized Fed- erated Learning Strategy for End-to-End ASR of Dementia Pa- tients,

    W.-T. Hsu, C.-P. Chenet al., “A Cluster-based Personalized Fed- erated Learning Strategy for End-to-End ASR of Dementia Pa- tients,” inINTERSPEECH, 2024

    2024

  11. [19]

    Phone-purity Guided Discrete Tokens for Dysarthric Speech Recognition,

    H. Wang, X. Xieet al., “Phone-purity Guided Discrete Tokens for Dysarthric Speech Recognition,”arXiv preprint arXiv:2501.04379, 2025

    2025 arXiv

  12. [20]

    Spectro-temporal sparsity characteri- zation for dysarthric speech detection,

    I. Kodrasi and H. Bourlard, “Spectro-temporal sparsity characteri- zation for dysarthric speech detection,”IEEE/ACM T-ASLP, 2020

    2020

  13. [21]

    Balancing privacy and progress: a review of privacy challenges, systemic oversight, and patient perceptions in AI-driven healthcare,

    S. M. Williamson and V . Prybutok, “Balancing privacy and progress: a review of privacy challenges, systemic oversight, and patient perceptions in AI-driven healthcare,”Applied Sciences, 2024

    2024

  14. [22]

    Communication-Efficient Learn- ing of Deep Networks from Decentralized Data,

    B. McMahan, E. Mooreet al., “Communication-Efficient Learn- ing of Deep Networks from Decentralized Data,” inAISTATS, 2017

    2017

  15. [23]

    Federated learning for keyword spot- ting,

    D. Leroy, A. Couckeet al., “Federated learning for keyword spot- ting,” inICASSP, 2019

    2019

  16. [24]

    Training keyword spotting mod- els on non-iid data with federated learning,

    A. Hard, K. Partridgeet al., “Training keyword spotting mod- els on non-iid data with federated learning,” inINTERSPEECH, 2020

    2020

  17. [25]

    Improving on-device speaker verification using federated learning with privacy,

    F. Granqvist, M. Seigelet al., “Improving on-device speaker verification using federated learning with privacy,” inINTER- SPEECH, 2020

    2020

  18. [26]

    Stealthy backdoor attack towards feder- ated automatic speaker verification,

    L. Zhang, L. Liuet al., “Stealthy backdoor attack towards feder- ated automatic speaker verification,” inICASSP, 2024

    2024

  19. [27]

    Federated learning for speech emotion recognition applications,

    S. Latif, S. Khalifaet al., “Federated learning for speech emotion recognition applications,” inIPSN, 2020

    2020

  20. [28]

    Semi-fedSER: Semi-supervised learning for speech emotion recognition on federated learning us- ing multiview pseudo-labeling,

    T. Feng and S. Narayanan, “Semi-fedSER: Semi-supervised learning for speech emotion recognition on federated learning us- ing multiview pseudo-labeling,” inINTERSPEECH, 2022

    2022

  21. [29]

    A federated ap- proach in training acoustic models

    D. Dimitriadis, R. G. Ken’ichi Kumataniet al., “A federated ap- proach in training acoustic models.” inINTERSPEECH, 2020

    2020

  22. [30]

    Training speech recognition models with federated learning: A quality/cost framework,

    D. Guliani, F. Beaufayset al., “Training speech recognition models with federated learning: A quality/cost framework,” in ICASSP, 2021

    2021

  23. [31]

    Federated learning in ASR: Not as easy as you think,

    W. Yu, J. Freiwaldet al., “Federated learning in ASR: Not as easy as you think,” inITG SpeechCom, 2021

    2021

  24. [32]

    Federated acoustic modeling for automatic speech recognition,

    X. Cui, S. Luet al., “Federated acoustic modeling for automatic speech recognition,” inICASSP, 2021

    2021

  25. [33]

    Cross-silo federated training in the cloud with diversity scaling and semi-supervised learning,

    K. Nandury, A. Mohanet al., “Cross-silo federated training in the cloud with diversity scaling and semi-supervised learning,” in ICASSP, 2021

    2021

  26. [34]

    End-to-end speech recognition from federated acoustic models,

    Y . Gao, T. Parcolletet al., “End-to-end speech recognition from federated acoustic models,” inICASSP, 2022

    2022

  27. [35]

    Importance of Smoothness Induced by Optimizers in FL4ASR: Towards Understanding Fed- erated Learning for End-to-End ASR,

    S. S. Azam, T. Likhomanenkoet al., “Importance of Smoothness Induced by Optimizers in FL4ASR: Towards Understanding Fed- erated Learning for End-to-End ASR,” inASRU, 2023

    2023

  28. [36]

    Federated Learning for Speech Recognition: Revisiting Current Trends Towards Large-Scale ASR,

    S. S. Azam, M. Pelikanet al., “Federated Learning for Speech Recognition: Revisiting Current Trends Towards Large-Scale ASR,” inNeurIPS, 2023

    2023

  29. [37]

    Federated learning with differ- ential privacy for end-to-end speech recognition,

    M. Pelikan, S. S. Azamet al., “Federated learning with differ- ential privacy for end-to-end speech recognition,”arXiv preprint arXiv:2310.00098, 2023

    2023

  30. [38]

    Parameter-Efficient Transfer Learning under Federated Learning for Automatic Speech Recognition,

    X. Kan, Y . Xiaoet al., “Parameter-Efficient Transfer Learning under Federated Learning for Automatic Speech Recognition,” arXiv preprint arXiv:2408.11873, 2024

    2024 arXiv

  31. [39]

    Communication-Efficient Personalized Federated Learning for Speech-to-Text Tasks,

    Y . Du, Z. Zhanget al., “Communication-Efficient Personalized Federated Learning for Speech-to-Text Tasks,” inICASSP, 2024

    2024

  32. [40]

    Federated Optimization in Heteroge- neous Networks,

    T. Li, A. K. Sahuet al., “Federated Optimization in Heteroge- neous Networks,” inMLSys, 2020

    2020

  33. [41]

    Fair and Privacy-Preserving Alzheimer’s Disease Diagnosis Based on Spontaneous Speech Analysis via Federated Learning,

    S. I. A. Meerza, Z. Liet al., “Fair and Privacy-Preserving Alzheimer’s Disease Diagnosis Based on Spontaneous Speech Analysis via Federated Learning,” inEMBC, 2022

    2022

  34. [42]

    Federated learning for secure development of AI models for Parkinson’s disease detec- tion using speech from different languages,

    S. T. Arasteh, C. D. Rios-Urregoet al., “Federated learning for secure development of AI models for Parkinson’s disease detec- tion using speech from different languages,” inINTERSPEECH, 2023

    2023

  35. [43]

    A Comparative Analysis of Federated Learning for Speech-Based Cognitive De- cline Detection,

    S. Kalabakov, M. Gonzalez-Machorroet al., “A Comparative Analysis of Federated Learning for Speech-Based Cognitive De- cline Detection,” inINTERSPEECH, 2024

    2024

  36. [44]

    Federated learning with interme- diate representation regularization,

    Y . L. Tun, C. M. Thwalet al., “Federated learning with interme- diate representation regularization,” inBigComp, 2023

    2023

  37. [45]

    Sparse Training for Federated Learning With Regularized Error Correction,

    R. Greidi and K. Cohen, “Sparse Training for Federated Learning With Regularized Error Correction,”JSTSP, 2024

    2024

  38. [46]

    On information and sufficiency,

    S. Kullback and R. A. Leibler, “On information and sufficiency,” Ann. Math. Stat., 1951

    1951

  39. [47]

    Dysarthric speech database for universal access research

    H. Kim, M. Hasegawa-Johnsonet al., “Dysarthric speech database for universal access research.” inINTERSPEECH, 2008

    2008

  40. [48]

    The natural history of alzheimer’s disease: description of study cohort and accuracy of diagnosis,

    J. T. Becker, F. Boileret al., “The natural history of alzheimer’s disease: description of study cohort and accuracy of diagnosis,” Arch. Neurol., 1994

    1994

  41. [49]

    A survey on distributed machine learning,

    J. Verbraeken, M. Woltinget al., “A survey on distributed machine learning,”CSUR, 2020

    2020

  42. [50]

    HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units,

    W.-N. Hsu, B. Bolteet al., “HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units,” IEEE/ACM T-ASLP, 2021

    2021

  43. [51]

    Alzheimer’s Dementia Recognition through Spontaneous Speech: The ADReSS Challenge,

    S. Luz, F. Haideret al., “Alzheimer’s Dementia Recognition through Spontaneous Speech: The ADReSS Challenge,” inIN- TERSPEECH, 2020

    2020

  44. [52]

    Librispeech: An ASR corpus based on public domain audio books,

    V . Panayotov, G. Chenet al., “Librispeech: An ASR corpus based on public domain audio books,” inICASSP, 2015

    2015

  45. [53]

    Connectionist temporal classifica- tion: labelling unsegmented sequence data with recurrent neural networks,

    A. Graves, S. Fern ´andezet al., “Connectionist temporal classifica- tion: labelling unsegmented sequence data with recurrent neural networks,” inICML, 2006

    2006

  46. [54]

    Bootstrap estimates for confidence inter- vals in asr performance evaluation,

    M. Bisani and H. Ney, “Bootstrap estimates for confidence inter- vals in asr performance evaluation,” inICASSP, 2004

    2004

Pith tools

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