Pith. sign in

REVIEW 2 major objections 6 minor 57 references

On-the-fly Routing for Zero-shot MoE Speaker Adaptation of Speech Foundation Models for Dysarthric Speech Recognition

T0 review · 2 major / 6 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 On-the-fly router cuts dysarthric speech errors by up to 6.4%

desk verdict Well-ablated efficient MoE adaptation for dysarthric ASR produces real WER gains, but the zero-shot claim is undermined by train/test speaker overlap in the backbone and expert initialization. read the letter →

arxiv 2505.22072 v1 pith:HBY5VAVV submitted 2025-05-28 cs.SD eess.AS

Shujie HU , Xurong Xie , Mengzhe Geng , Jiajun Deng , Huimeng Wang , Guinan Li , Chengxi Deng , Tianzi Wang
show 3 more authors
Mingyu Cui Helen Meng Xunying Liu
This is my paper
classification cs.SDeess.AS
keywords dysarthricspeechrecognitionzero-shotspeakeradaptationmixtureofexpertsfoundationmodelson-the-flyroutingreal-timeUASpeech
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 aims to show that a speech-recognition foundation model can be adapted to a speaker with dysarthria instantly, without retraining or waiting for a batch of that speaker's data. The proposed method uses a small routing network that watches each incoming utterance and predicts how to blend a small set of specialized adapter experts, where each expert was initialized to handle a particular severity or gender group. On the UASpeech dysarthric benchmark, this on-the-fly adaptation reduces word error rate by up to 1.34% absolute (6.36% relative) over the unadapted HuBERT and WavLM baselines, and it matches batch-mode adaptation while running up to 7 times faster. If the result holds, users with dysarthria could get personalized, real-time speech recognition from a single utterance, without collecting or labeling large amounts of their own speech.

What carries the argument

The central object is a feature-driven routing network that predicts speaker-level routing parameters $\mathbf{r}_s$ for a mixture-of-experts adapter inserted into the second transformer block of a HuBERT or WavLM backbone. The experts are residual adapter blocks initialized from severity- and gender-conditioned adaptive-training parameters; a KL-divergence loss penalizes similarity between expert outputs, and an auxiliary classification task uses severity and gender labels. At test time the router consumes attentive statistics-pooled hidden states of the current utterance and outputs weights immediately, so adaptation is zero-shot and real-time. The mechanism carrying the argument is the regression of the router to the speaker-dependent weights obtained through supervised speaker adaptive training, which is what transfers adaptation ability to unseen speakers.

What would settle it

Apply the trained on-the-fly router, with no adaptation, to a held-out group of dysarthric speakers who contributed no training data, for example speakers from a different dysarthric corpus; if the router does not reduce WER relative to the unadapted baseline on that group, the zero-shot generalization claim is refuted.

Watch

Extended reading notes

Core claim

Zero-shot, real-time speaker adaptation for dysarthric speech recognition can be achieved by predicting the mixture-of-experts router weights directly from the acoustic input, rather than optimizing them per speaker at test time. During training, a supervised speaker-adaptive step learns speaker-dependent routing parameters; a separate routing network is then trained with mean-squared error to reproduce those parameters from utterance-level features produced by attentive statistics pooling. At test time the backbone and experts are frozen, the routing network produces weights on the fly, and the experts -- initialized from severity- and gender-conditioned adapters and regularized by a KL-divergence diversity loss -- combine to give the adapted output. The paper reports statistically significant WER reductions up to 1.34% absolute (6.36% relative) over unadapted HuBERT and WavLM baselines, WER reductions up to 2.55% absolute (11.44% relative) over batch-mode adaptation, real-time-factor speedups up to 7 times, and, after cross-system rescoring, the lowest published WER of 16.35% on UASpeech.

Load-bearing premise

The load-bearing premise is that a router trained on the training speakers' data will predict routing weights that transfer to dysarthric speakers it has never heard, even though the paper evaluates only on the same 16 UASpeech test speakers.

Editorial extensions

If this is right

  • A new dysarthric speaker can receive adapted recognition from the first utterance, with no fine-tuning and no accumulation of adaptation data.
  • On-the-fly MoE adaptation matches batch-mode MoE accuracy while using far fewer speaker-dependent parameters and operating about 7 times faster in terms of real-time factor.
  • Combined with existing adapted systems via cross-system multi-pass rescoring, the MoE-adapted HuBERT and WavLM systems reach a UASpeech word error rate of 16.35%, the lowest published at the time.
  • The on-the-fly predicted routing parameters cluster by speech-impairment severity in the paper's visualization, suggesting the router is capturing clinically meaningful speaker structure rather than only speaker identity.
  • With just one utterance, on-the-fly adaptation yields WER reductions of up to 2.55% absolute over batch-mode RAB and MoE approaches, removing the data-accumulation delay.

Reading between the lines

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

  • If this zero-shot mechanism generalizes beyond UASpeech's 16 dysarthric speakers, the same routing architecture could be reused for other low-resource, high-variability speaker populations such as elderly or child speech.
  • The severity-structured routing weights could be repurposed as a continuous intelligibility or severity estimator, since their pattern tracks the four severity groups used to initialize the experts.
  • A testable extension is to train the router on simulated or augmented dysarthric speech from additional speakers and measure whether gains scale with population diversity; that would separate learning of general impairment structure from memorizing the 16 UASpeech speakers.
  • The router's weights could also be used to warm-start a small amount of per-speaker fine-tuning when more data later becomes available, combining the zero-shot speedup with batch-mode refinement.
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 / 6 minor

Summary. This paper proposes a mixture-of-experts (MoE) based speaker adaptation method for self-supervised speech foundation models (SFMs) applied to dysarthric speech recognition. The method adds an MoE module with residual-adapter experts into an SFM and uses a feature-driven routing network to predict speaker-dependent routing parameters on the fly, avoiding the two-stage batch-mode adaptation used in prior work. Domain knowledge (dysarthria severity and gender) is incorporated by initializing experts from adaptive training and by an auxiliary classification loss; a KL-divergence loss encourages expert diversity. Experiments on UASpeech compare on-the-fly MoE adaptation with SI baselines, i-vector/x-vector adaptation, and batch-mode RAB/MoE systems, reporting statistically significant WER reductions (up to 1.34% absolute over SI), comparable WER to batch-mode with 7x RTF speedup, and the lowest published UASpeech WER of 16.35% when combined with other systems via rescoring.

Significance. The core idea of amortizing speaker-dependent routing parameters into a lightweight router is practically appealing for real-time dysarthric ASR, and the paper provides carefully controlled comparisons, MAPSSWE significance testing, and ablations of expert initialization, KL loss, classification task, domain knowledge, and attentive pooling. The main weakness is that the 'zero-shot' claim is not cleanly evaluated: the round-robin protocol in Sec. 4.4 still allows the frozen backbone and expert initializations to have seen the target speaker's voice, so the observed gains may reflect speaker identification rather than transfer to unseen dysarthric speakers. With that caveat, the method still demonstrates a useful real-time adaptation mechanism with WER reductions over SI baselines and large RTF savings relative to batch-mode adaptation.

major comments (2)
  1. [Sec. 4.4, Table 4] The round-robin experiment does not establish zero-shot speaker adaptation as claimed in the abstract. For each test speaker s, the authors 'exclude their data from the training set before speaker adaptation,' but the SI backbone, the expert initializations from adaptive training [10], and the SAT-derived routing targets are all produced from a training set that contains s's speech in Blocks 1 and 3. Since the routing network operates on features extracted from a backbone that has already encoded s's voice, the router could simply be interpolating among the 15 training speakers' routing vectors rather than predicting a routing vector for a genuinely unseen speaker. The authors' parenthetical in Sec. 4.4 ('even though the SI model is trained on data containing speakers from the test set') acknowledges this but does not mitigate it. To support the zero-shot claim, the authors should either (i) retrain the SI backbone and SAT targets with the target speaker's data fully excluded (all blocks), or (ii) evaluate on dysarthric speakers from a different corpus, or (iii) re-frame the claim as 'adaptation without test-time data from the target block' and state the limitation explicitly.
  2. [Sec. 4.4 and Abstract] The zero-shot claim also lacks evidence of generalization beyond the 16 UASpeech dysarthric speakers, who constitute the entire test set. Since the UASpeech training set includes those same speakers' speech from other blocks, the evaluation does not address how the router would behave for a dysarthric speaker whose voice was never seen in any form during training. A cross-corpus evaluation (e.g., TORGO) or a clear statement that the claim is limited to UASpeech-style speakers is needed.
minor comments (6)
  1. [Abstract] The phrase 'speech foundation models based dysarthric speech recognition' is grammatically awkward; consider 'dysarthric speech recognition based on speech foundation models.'
  2. [Sec. 2] The phrase 'unseen speakers's data' should be 'unseen speakers' data.'
  3. [Fig. 4] The label 'MOUTH' and the 'Gumbel Softmax' / 'Spk Weight' annotations in the right sub-figure are not explained in the caption or text; the figure appears to contain leftover elements from another version and should be cleaned up.
  4. [Fig. 2] The line charts for on-the-fly and batch-mode routing parameters are too small to read; please enlarge or separate them so the convergence behavior is visible.
  5. [Secs. 2 and 3] The loss weights alpha, beta, and gamma are stated to be 'empirically set' to 5, 0.1, and 0.5, respectively, without any sensitivity analysis or justification; adding a small sweep or a comment on how these values were chosen would strengthen reproducibility.
  6. [Sec. 3] Eq. 5 uses theta_hat_S as the training targets for the MSE loss, but theta_hat_S is not explicitly defined in Sec. 3 (it is the SAT-derived routing parameter set from Eq. 3); please define it at first use.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the on-the-fly router is a learned amortization of SAT routing parameters, and the WER gains are evaluated on held-out Block 2 test data, so the central claim does not reduce to its training target by construction.

full rationale

The main potential circular step is the MSE training of the routing network (Eq. 5) against SAT-derived speaker-dependent routing parameters (Sec. 3). This is a self-distillation / amortized-inference scheme rather than a definitional identity: at test time the routing parameters are produced by a learned function of the test utterance's hidden features, and the paper's headline WER numbers are measured on UASpeech Block 2, which is not used to obtain the router's training targets. The router could in principle fail to match the SAT targets; the fact that it achieves comparable WER is an empirical result, not a tautology. The round-robin zero-shot experiment (Sec. 4.4) has a genuine external-validity caveat: the paper explicitly notes that 'even though the SI model is trained on data containing speakers from the test set,' so the frozen backbone and expert initializations have seen the target speakers' other blocks. This weakens the 'unseen speaker' interpretation and is a validity/leakage concern, but it does not make the WER reduction equivalent to the router's training input by construction; the test utterances themselves are held out. Self-citations to [10] for RAB expert settings and to prior CUHK systems in the system combination are not load-bearing: the current paper re-runs its own SAT (Eq. 3) and compares against the SI baseline and batch-mode adaptation on a public benchmark. No uniqueness theorem or ansatz is imported from the authors' prior work to force the MoE design. Accordingly, no specific circular step can be exhibited, and the score is 0.

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

The central claim rests on a small set of empirically chosen hyperparameters, a domain-knowledge conditioning assumption about severity/gender labels, and a self-referential training target where the router imitates the SAT router. No new physical or architectural entities are introduced beyond the standard MoE adapter and routing network components.

free parameters (5)
  • KL loss weight alpha = 5
    Empirically set without sensitivity analysis (Sec. 2).
  • CE loss weight beta = 0.1
    Empirically set without sensitivity analysis (Sec. 2).
  • MSE loss weight gamma = 0.5
    Empirically set without sensitivity analysis (Sec. 3).
  • Number of experts N = 5 or 10
    Design choice; 10 for severity/gender experts and 5 for severity-only experts, with no systematic study of N (Sec. 4.3, Table 3).
  • MoE placement in transformer block = 2nd block
    Chosen without ablation; alternative placements are not tested (Sec. 2).
assumptions (5)
  • domain assumption KL divergence between softmaxed expert outputs encourages expert diversity and generalization to unseen speakers.
    The paper assumes diversity of experts improves coverage on unseen speakers; this is a heuristic, not proven (Eq. 1).
  • domain assumption Severity/gender labels are available and accurate for training, and the automatic predictions used at test time are sufficiently reliable.
    Training uses ground-truth labels; test-time batch adaptation uses predicted labels from an external classifier whose accuracy is not reported (Sec. 2, Test-Time Adaptation).
  • domain assumption Attentive statistics pooling captures speaker characteristics from hidden outputs.
    Adopted from speaker verification (ref [35]); assumed to transfer to dysarthric speaker routing (Eq. 4).
  • domain assumption Pre-trained HuBERT/WavLM features are appropriate for dysarthric speech despite being pre-trained mostly on healthy speech.
    The paper relies on prior work showing usefulness of these SSL models for dysarthric ASR (Sec. 4.1 and refs [5,9]).
  • ad hoc to paper SAT-derived routing parameters used as MSE targets are a good supervisory signal for the on-the-fly router.
    The router is trained to regress to the SAT routing parameters, which are themselves outputs of the same pipeline; this is a self-referential training target (Sec. 3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of On-the-fly Routing for Zero-shot MoE Speaker Adaptation of Speech Foundation Models for Dysarthric Speech Recognition." pith.science (2026). https://pith.science/paper/HBY5VAVV

@misc{pith2026250522072,
  author       = {Pith},
  title        = {Pith review of: On-the-fly Routing for Zero-shot MoE Speaker Adaptation of Speech Foundation Models for Dysarthric Speech Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HBY5VAVV}},
  note         = {Machine review of arXiv:2505.22072}
}
read the original abstract

This paper proposes a novel MoE-based speaker adaptation framework for foundation models based dysarthric speech recognition. This approach enables zero-shot adaptation and real-time processing while incorporating domain knowledge. Speech impairment severity and gender conditioned adapter experts are dynamically combined using on-the-fly predicted speaker-dependent routing parameters. KL-divergence is used to further enforce diversity among experts and their generalization to unseen speakers. Experimental results on the UASpeech corpus suggest that on-the-fly MoE-based adaptation produces statistically significant WER reductions of up to 1.34% absolute (6.36% relative) over the unadapted baseline HuBERT/WavLM models. Consistent WER reductions of up to 2.55% absolute (11.44% relative) and RTF speedups of up to 7 times are obtained over batch-mode adaptation across varying speaker-level data quantities. The lowest published WER of 16.35% (46.77% on very low intelligibility) is obtained.

Figures

Figures reproduced from arXiv: 2505.22072 by the authors.

Figure 1
Figure 1. Architecture of MoE-based adaptation on SFM, where the routing parameters are derived from either a) speaker￾dependent parameters in batch mode; or b) an on-the-fly rout￾ing network. “LN”, “MHSA”, “DP” and “FFN” are layer￾norm, multi-head self-attention, dropout and feedforward. 2) Performance: Experimental results on the UASpeech [33] dysarthric corpus suggest that i) the proposed on-the-fly MoE￾based adaptation ap… view at source ↗
Figure 2
Figure 2. Examples of on-the-fly (a) & b)) and batch-mode (a) & c)) MoE-based speaker adaptation on the SFM. Routing parameters r s in a) serve as SD parameters, while experts are shared by all speakers. The line charts in b) and c) illustrate the variation in a specific expert’s routing parameters as a function of utterance count. α s,k t = exp(v T f(Whˆ s,k t + b) + c) PT s,k t=1 exp(vT f(Whˆ s,k t + b) + c) (4) where W, b,… view at source ↗
Figure 3
Figure 3. T-SNE visualization of the on-the-fly MoE-based, i￾vector, and x-vector adaptation. The determinants of their co￾variance matrices are shown in each bracket. speaker features produced by on-the-fly MoE in T-SNE visual￾ization compared to i-vectors and x-vectors, as shown in [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: WER and cosine similarity on HuBERT systems with varying amounts of speaker adaptation data. Severity Label 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Spk_id 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Spk_id 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Spk_id (a) (b) (c) VL: {0, 1, 2, …
Figure 5
Figure 5. Figure 5: Heatmap visualization of routing parameters under varying settings on HuBERT: a) batch-mode without domain knowledge, b) batch-mode and c) on-the-fly with domain knowl￾edge. “Severity Label: {Spk ids}” is given at the top of Figure. Low Processing Latency. The left par…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 51 canonical work pages

  1. [42]

    Parameter-Efficient Conformers via Sharing Sparsely-Gated Experts for End-to-End Speech Recognition,

    Y . Bai et al. , “Parameter-Efficient Conformers via Sharing Sparsely-Gated Experts for End-to-End Speech Recognition,” in INTERSPEECH, 2022, pp. 1676–1680

    work page 2022

  2. [10]

    6 & 7 vs

    Comparison with RAB-based6 methods: Both batch-mode and on-the-fly MoE-based adaptation achieve comparable aver- age WERs compared to batch-mode RAB-based method [10], while producing lower WERs on the “VL” group (Sys. 6 & 7 vs. 4). Notably, the MoE-based methods require only 1/25000 of the SD parameters (Sys. 6 vs. 4) and operate at 1/7 of the RTF (Sys. ...

    work page

  3. [1]

    Introduction Despite the rapid progress of ASR technologies targeting nor- mal and healthy users, their application to those suffering from speech disorders, such as dysarthria, remains a challenging task to date [1–10]. Dysarthric speech brings challenges on all fronts to current deep learning based ASR technologies predominantly targeting healthy users:...

    work page

  4. [2]

    On-the-fly Routing for Zero-shot MoE Speaker Adaptation of Speech Foundation Models for Dysarthric Speech Recognition

    Novelty: To the best of our knowledge, this paper is the first to investigate on-the-fly MoE-based speaker adaptation for dysarthric speech recognition, whereas prior efforts have pri- marily focused on typical speech [32]. Our method addresses the three major challenges outlined earlier: a) while previous methods lack the ability to adapt to unseen speak...

    work page Pith review arXiv 2025

  5. [3]

    Performance: Experimental results on the UASpeech [33] dysarthric corpus suggest that i) the proposed on-the-fly MoE- based adaptation approach produces statistically significant word error rate (WER) reductions of up to 1.34% absolute (6.36% relative) over the baseline SI HuBERT and WavLM models. ii) Consistent WER reductions of up to 2.55% abso- lute (1...

    work page

  6. [4]

    Analysis: Heatmap visualization intuitively reveals that the on-the-fly predicted SD routing parameters exhibit more con- sistent and interpretable speech impairment severity centric features than those obtained without domain knowledge

    work page

  7. [5]

    For example, HuBERT contains three main components: 1) a multi-layer CNN-based feature encoder; 2) an L-layer transformer-based context network with a projection layer; and

    Batch-Mode MoE Speaker Adaptation Backbone Speech Foundation Models (SFMs): SSL speech foundation models such as Wav2vec2.0 [12], HuBERT [14], and WavLM [13] share similar Transformer-based backbones. For example, HuBERT contains three main components: 1) a multi-layer CNN-based feature encoder; 2) an L-layer transformer-based context network with a proje...

    work page

  8. [6]

    In this paper, we fine-tune the pre-trained HuBERT and WavLM with a CTC decoder

    a k-means quantization module. In this paper, we fine-tune the pre-trained HuBERT and WavLM with a CTC decoder. MoE Architecture: As shown in Fig. 1, the MoE module is in- tegrated into the 2nd Transformer block, positioned between the feedforward layer and the dropout module. Residual Adapter Blocks (RAB) [10] act as expert network modules. All speakers ...

    work page

Show all 57 references
  1. [7]

    !ℒ!%𝑪𝒔𝒀𝒔 BackboneSFM 🔥 a) SAT 🔥Router Requiresdata accumula9on Experts Router 🔥Experts Softmax(𝒂𝟒𝒔)Softmax(𝒂𝟏𝒔)…ℒ12 U?erance of speaker 𝑠 c) Batch-modeℒ!

    On-the-fly MoE Speaker Adaptation Routing Network Architecture: To achieve zero-shot speaker adaptation and reduce latency, a feature-driven routing network is designed to generate homogeneous SD routing parameters (shown in the line chart of Fig. 2(b)) on the fly, enabling ef...

  2. [8]

    Task Description and Experimental Setup UASpeech [33] is the largest publicly available dysarthric speech dataset containing 16 dysarthric and 13 control speak- ers

    Experiments 4.1. Task Description and Experimental Setup UASpeech [33] is the largest publicly available dysarthric speech dataset containing 16 dysarthric and 13 control speak- ers. It includes 155 common and 300 uncommon words and is further divided into three subset blocks ...

  3. [9]

    +” represents score interpolation, while “X →Y

    Comparison with the SI baseline, i-vector and x-vector adaptation: The proposed on-the-fly MoE-based speaker adap- tation consistently outperforms these systems with statistically significant WER reductions of up to 1.34% absolute (6.36% relative) on HuBERT and WavLM (Sys. 7 v...

  4. [11]

    while operating approximately 7 times faster in terms of RTF (Sys

    Comparison with batch-model MoE-based adaptation: The on-the-fly MoE-based speaker adaptation achieves compa- rable performance to offline batch-mode MoE-based method, 6The RAB-based method can be considered as a special case of MoE, with a single expert and SD parameters as t...

  5. [12]

    Know. ”) integration and KL loss, as well as on-the-fly MoE-based speaker adaptation with and with- out attentive pooling (“Atten. Pool

    Best performing system: By combining multiple adapted systems, including LHUC-SAT TDNN, RAB-based adapted SFMs, and the proposed MoE-based adapted SFMs via cross- system rescoring, the lowest published overall WER of16.35% (46.77% on very low intelligibility, Sys. 13) is obtai...

    2022

  6. [13]

    Feature- driven routing networks are designed to produce homogeneous SD routing parameters on the fly, thereby facilitating zero-shot and real-time speaker adaptation

    Conclusion This paper presents a novel on-the-fly MoE-based speaker adap- tation for SSL pre-trained SFMs on dysarthric speech. Feature- driven routing networks are designed to produce homogeneous SD routing parameters on the fly, thereby facilitating zero-shot and real-time s...

  7. [14]

    14200220, 14200021, 14200324, TRS T45-407/19N, Innova- tion Technology Fund grant No

    Acknowledgements This research is supported by Hong Kong RGC GRF grant No. 14200220, 14200021, 14200324, TRS T45-407/19N, Innova- tion Technology Fund grant No. ITS/218/21, the project of China Disabled Persons Federation (CDPF2023KF00002), Ba- sic Research Project of ISCAS (I...

  8. [15]

    Model adaptation and adaptive training for the recognition of dysarthric speech,

    S. Sehgal et al., “Model adaptation and adaptive training for the recognition of dysarthric speech,” in SLPAT, 2015

    2015

  9. [16]

    Deep learning of articulatory-based representa- tions and applications for improving dysarthric speech recogni- tion,

    F. Xiong et al., “Deep learning of articulatory-based representa- tions and applications for improving dysarthric speech recogni- tion,” in ITG-Symposium. VDE, 2018, pp. 1–5

    2018

  10. [17]

    Recent Progress in the CUHK Dysarthric Speech Recognition System,

    S. Liu et al., “Recent Progress in the CUHK Dysarthric Speech Recognition System,” TASLP, vol. 29, pp. 2267–2281, 2021

    2021

  11. [18]

    Speaker adaptation using spectro-temporal deep features for dysarthric and elderly speech recognition,

    M. Geng et al., “Speaker adaptation using spectro-temporal deep features for dysarthric and elderly speech recognition,” TASLP, vol. 30, pp. 2597–2611, 2022

    2022

  12. [19]

    Exploring Self-supervised Pre-trained ASR Mod- els For Dysarthric and Elderly Speech Recognition,

    S. Hu et al., “Exploring Self-supervised Pre-trained ASR Mod- els For Dysarthric and Elderly Speech Recognition,” in ICASSP. IEEE, 2023, pp. 1–5

    2023

  13. [20]

    Acoustic Modelling From Raw Source and Filter Components for Dysarthric Speech Recognition,

    Z. Yue et al., “Acoustic Modelling From Raw Source and Filter Components for Dysarthric Speech Recognition,”TASLP, vol. 30, pp. 2968–2980, 2022

    2022

  14. [21]

    Exploiting Cross-Domain And Cross-Lingual Ul- trasound Tongue Imaging Features For Elderly And Dysarthric Speech Recognition,

    S. Hu et al., “Exploiting Cross-Domain And Cross-Lingual Ul- trasound Tongue Imaging Features For Elderly And Dysarthric Speech Recognition,” in INTERSPEECH, 2023, pp. 2313–2317

    2023

  15. [22]

    Exploiting Cross Domain Acoustic-to- articulatory Inverted Features for Disordered Speech Recogni- tion,

    S. Hu, S. Liu et al. , “Exploiting Cross Domain Acoustic-to- articulatory Inverted Features for Disordered Speech Recogni- tion,” in ICASSP. IEEE, 2022, pp. 6747–6751

    2022

  16. [23]

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

    S. Hu et al. , “Self-Supervised ASR Models and Features for Dysarthric and Elderly Speech Recognition,” TASLP, vol. 32, pp. 3561–3575, 2024

    2024

  17. [24]

    Structured Speaker-Deficiency Adaptation of Foundation Models for Dysarthric and Elderly Speech Recog- nition,

    S. Hu, X. Xie et al., “Structured Speaker-Deficiency Adaptation of Foundation Models for Dysarthric and Elderly Speech Recog- nition,” arXiv preprint arXiv:2412.18832, 2024

    2024 arXiv

  18. [25]

    Temporal characteristics of the speech of nor- mal elderly adults,

    B. L. Smith et al., “Temporal characteristics of the speech of nor- mal elderly adults,” JSLHR, vol. 30, pp. 522–529, 1987

    1987

  19. [26]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski et al., “wav2vec 2.0: A framework for self-supervised learning of speech representations,” in Advances in NeuralIPS , 2020, pp. 12 449–12 460

    2020

  20. [27]

    Wavlm: Large-scale self-supervised pre-training for full stack speech processing,

    S. Chen et al., “Wavlm: Large-scale self-supervised pre-training for full stack speech processing,” JSTSP, vol. 16, pp. 1505–1518, 2022

    2022

  21. [28]

    Hubert: Self-supervised speech representation learning by masked prediction of hidden units,

    W.-N. Hsu et al., “Hubert: Self-supervised speech representation learning by masked prediction of hidden units,” TASLP, vol. 29, pp. 3451–3460, 2021

    2021

  22. [29]

    Speaker adaptation for Wav2vec2 based dysarthric ASR,

    M. K. Baskar et al. , “Speaker adaptation for Wav2vec2 based dysarthric ASR,” in INTERSPEECH, 2022, pp. 3403–3407

    2022

  23. [30]

    Perceiver-Prompt: Flexible Speaker Adaptation in Whisper for Chinese Disordered Speech Recogni- tion,

    Y . Jiang, T. Wang et al. , “Perceiver-Prompt: Flexible Speaker Adaptation in Whisper for Chinese Disordered Speech Recogni- tion,” in INTERSPEECH, 2024, pp. 2025–2029

    2024

  24. [31]

    Efficient adapter transfer of self-supervised speech models for automatic speech recognition,

    B. Thomas et al. , “Efficient adapter transfer of self-supervised speech models for automatic speech recognition,” in ICASSP. IEEE, 2022, pp. 7102–7106

    2022

  25. [32]

    Exploring Efficient-Tuning Methods in Self- Supervised Speech Models,

    Z.-C. Chen et al., “Exploring Efficient-Tuning Methods in Self- Supervised Speech Models,” in SLT, 2023, pp. 1120–1127

    2023

  26. [33]

    Efficient Domain Adaptation for Speech Foundation Models,

    B. Li et al., “Efficient Domain Adaptation for Speech Foundation Models,” in ICASSP, 2023, pp. 1–5

    2023

  27. [34]

    Chapter: Exploiting Convolutional Neural Net- work Adapters for Self-Supervised Speech Models,

    Z.-C. Chen et al., “Chapter: Exploiting Convolutional Neural Net- work Adapters for Self-Supervised Speech Models,” inICASSPW, 2023, pp. 1–5

    2023

  28. [35]

    Adaptive mixtures of local experts,

    R. A. Jacobs et al., “Adaptive mixtures of local experts,” Neural computation, vol. 3, pp. 79–87, 1991

    1991

  29. [36]

    Hierarchical mixtures of experts and the EM algorithm,

    M. I. Jordan et al., “Hierarchical mixtures of experts and the EM algorithm,” Neural computation, vol. 6, pp. 181–214, 1994

    1994

  30. [37]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,

    N. Shazeer et al. , “Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,” arXiv preprint arXiv:1701.06538, 2017

    2017 arXiv

  31. [38]

    Mixtral of experts,

    A. Q. Jiang et al. , “Mixtral of experts,” arXiv preprint arXiv:2401.04088, 2024

    2024 arXiv

  32. [39]

    Deepseekmoe: Towards ultimate expert special- ization in mixture-of-experts language models,

    D. Dai et al., “Deepseekmoe: Towards ultimate expert special- ization in mixture-of-experts language models,” arXiv preprint arXiv:2401.06066, 2024

    2024 arXiv

  33. [40]

    Hunyuan-large: An open-source moe model with 52 billion activated parameters by tencent,

    X. Sun et al. , “Hunyuan-large: An open-source moe model with 52 billion activated parameters by tencent,” arXiv preprint arXiv:2411.02265, 2024

    2024 arXiv

  34. [41]

    SpeechMoE: Scaling to Large Acoustic Models with Dynamic Routing Mixture of Experts,

    Z. You et al. , “SpeechMoE: Scaling to Large Acoustic Models with Dynamic Routing Mixture of Experts,” in INTERSPEECH, 2021, pp. 2077–2081

    2021

  35. [43]

    BA-MoE: Boundary-Aware Mixture-of-Experts Adapter for Code-Switching Speech Recognition,

    P. Chen et al. , “BA-MoE: Boundary-Aware Mixture-of-Experts Adapter for Code-Switching Speech Recognition,” in ASRU, 2023, pp. 1–7

    2023

  36. [44]

    MoLE : Mixture Of Language Experts For Multi- Lingual Automatic Speech Recognition,

    Y . Kwon et al., “MoLE : Mixture Of Language Experts For Multi- Lingual Automatic Speech Recognition,” in ICASSP, 2023, pp. 1–5

    2023

  37. [45]

    Aphasic Speech Recognition Using a Mixture of Speech Intelligibility Experts,

    M. Perez et al., “Aphasic Speech Recognition Using a Mixture of Speech Intelligibility Experts,” in INTERSPEECH, 2020, pp. 4986–4990

    2020

  38. [46]

    SAML: Speaker Adaptive Mixture of LoRA Ex- perts for End-to-End ASR,

    Q. Zhao et al., “SAML: Speaker Adaptive Mixture of LoRA Ex- perts for End-to-End ASR,” in INTERSPEECH, 2024, pp. 777– 781

    2024

  39. [47]

    Dysarthric speech database for universal access research,

    H. Kim et al., “Dysarthric speech database for universal access research,” in INTERSPEECH, 2008, pp. 1741–1744

    2008

  40. [48]

    Use of Speech Impairment Severity for Dysarthric Speech Recognition,

    M. Geng et al. , “Use of Speech Impairment Severity for Dysarthric Speech Recognition,” in INTERSPEECH, 2023, pp. 2328–2332

    2023

  41. [49]

    Attentive statistics pooling for deep speaker em- bedding,

    K. Okabe et al., “Attentive statistics pooling for deep speaker em- bedding,” in INTERSPEECH, 2018, pp. 2252–2256

    2018

  42. [50]

    Investigation of Data Augmentation Techniques for Disordered Speech Recognition

    M. Geng et al., “Investigation of Data Augmentation Techniques for Disordered Speech Recognition.” in INTERSPEECH, 2020, pp. 696–700

    2020

  43. [51]

    Some statistical issues in the comparison of speech recognition algorithms,

    L. Gillick and S. J. Cox, “Some statistical issues in the comparison of speech recognition algorithms,” in ICASSP. IEEE, 1989, pp. 532–535

    1989

  44. [52]

    Investigating Self-supervised Pretraining Frameworks for Pathological Speech Recognition,

    L. P. Violeta, W. C. Huang, and T. Toda, “Investigating Self-supervised Pretraining Frameworks for Pathological Speech Recognition,” in INTERSPEECH, 2022, pp. 41–45

    2022

  45. [53]

    Cross-lingual Self-Supervised Speech Rep- resentations for Improved Dysarthric Speech Recognition,

    A. Hernandez et al., “Cross-lingual Self-Supervised Speech Rep- resentations for Improved Dysarthric Speech Recognition,” inIN- TERSPEECH, 2022, pp. 51–55

    2022

  46. [54]

    DuTa-VC: A Duration-aware Typical-to-atypical V oice Conversion Approach with Diffusion Probabilistic Model,

    H. Wang et al., “DuTa-VC: A Duration-aware Typical-to-atypical V oice Conversion Approach with Diffusion Probabilistic Model,” in INTERSPEECH, 2023, pp. 1548–1552

    2023

  47. [55]

    Enhancing Pre-Trained ASR System Fine-Tuning for Dysarthric Speech Recognition Using Adversar- ial Data Augmentation,

    H. Wang, X. Xie et al. , “Enhancing Pre-Trained ASR System Fine-Tuning for Dysarthric Speech Recognition Using Adversar- ial Data Augmentation,” in ICASSP, 2024, pp. 12 311–12 315

    2024

  48. [56]

    Homogeneous Speaker Features for On-the- Fly Dysarthric and Elderly Speaker Adaptation,

    M. Geng et al. , “Homogeneous Speaker Features for On-the- Fly Dysarthric and Elderly Speaker Adaptation,” arXiv preprint arXiv:2407.06310, 2024

    2024 arXiv

  49. [57]

    Improving Acoustic Models in TORGO Dysarthric Speech Database,

    N. M. Joy and S. Umesh, “Improving Acoustic Models in TORGO Dysarthric Speech Database,” IEEE TNSRE, vol. 26, no. 3, pp. 637–645, 2018

    2018

Pith tools

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