Pith. sign in

REVIEW 3 major objections 6 minor 48 references

MOPSA: Mixture of Prompt-Experts Based Speaker Adaptation for Elderly Speech Recognition

T0 review · 3 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 MOPSA adapts Whisper to unseen elderly speakers in a single online pass, with statistically significant error cuts and up to 16x speedup over batch adaptation.

desk verdict A solid zero-shot adaptation entry for Whisper on elderly speech, with a new prompt-expert router; the headline numbers overstate, and a missing fixed-prompt control leaves the speaker-specific claim under-supported. read the letter →

arxiv 2505.24224 v1 pith:4CUNJIMA submitted 2025-05-30 eess.AS

Chengxi Deng , Xurong Xie , Shujie Hu , Mengzhe Geng , Yicong Jiang , Jiankun Zhao , Jiajun Deng , Guinan Li
show 5 more authors
Youjun Chen Huimeng Wang Haoning Xu Mingyu Cui Xunying Liu
This is my paper
classification eess.AS
keywords elderlyspeechrecognitionspeakeradaptationpromptlearningmixtureofexpertsWhisperzero-shotK-meansclusteringreal-time
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

MOPSA is a proposed method for adapting Whisper to elderly speech without hearing a speaker's voice in training and without waiting for test-time optimization. The paper's central claim is that clustering per-speaker prompts from speaker adaptive training into prototype 'prompt-experts', then having a small router combine these experts on the fly for each input, yields statistically significant word error rate and character error rate reductions over a speaker-independent Whisper baseline on two elderly speech corpora. It also claims this online routing is up to 16.12 times faster than offline batch-mode prompt adaptation. If these claims hold, MOPSA offers an inexpensive, low-latency path to personalized speech recognition for elderly users, including speakers with cognitive decline.

What carries the argument

The central machinery is a mixture of prompt-experts built from K-means clustering. In speaker adaptive training, each training speaker gets encoder-side prompts $P^i_e$ and decoder-side prompts $P^i_d$, optimized jointly with shared LoRA parameters. K-means then clusters the prompt vectors at each position into $C$ clusters, and each cluster centroid $E^l_c$ becomes an expert. A router network, made of multi-head attention layers and a downsampling CNN followed by linear layers, outputs weights $W \in \mathbb{R}^{C \times L}$ for each prompt position; the online prompt is computed as $Z^l = \sum_{j=1}^{C} w^l_j E^l_j$ followed by $M = Z + \phi(Z)$. The router is trained with a multi-task loss $\mathcal{L}_{Router} = \mathcal{L}_{ASR} + \alpha \mathcal{L}_{Spkr} + \beta \mathcal{L}_{MSE}$, combining the ASR cross-entropy loss, a speaker-identity classification loss, and an MSE loss that aligns online prompts with the original SAT prompts.

What would settle it

Run MOPSA on a held-out elderly corpus whose dialect, accent, or cognitive profile is not represented among the training speakers: if the gains over the speaker-independent baseline disappear or reverse under the paper's significance test, the claimed zero-shot generalization does not hold.

Watch

Extended reading notes

Core claim

On English DementiaBank Pitt and Cantonese JCCOCC MoCA, online MOPSA adaptation outperforms the speaker-independent Whisper baseline with word error rate reductions of 0.86% absolute (4.21% relative) and character error rate reductions of 1.47% absolute (5.40% relative), and it achieves real-time factor speed-ups of up to 16.12 times over offline batch-mode prompt adaptation. The improvements are reported as statistically significant under a matched-pair test at the 0.05 level. The method works by turning per-speaker prompts learned during speaker adaptive training into a small set of cluster centroids, then using a learned router to combine those centroids dynamically for each incoming utterance, with separate prompts for the encoder and decoder of Whisper.

Load-bearing premise

The whole zero-shot gain rests on the router's ability to generalize from the training speakers whose prompts were clustered to elderly speakers it has never heard.

Editorial extensions

If this is right

  • Zero-shot adaptation becomes possible for unseen speakers: the router can produce a speaker-tailored prompt from the first utterance, with no pseudo-label generation or gradient updates.
  • The single-pass routing mechanism removes the latency of batch-mode adaptation, which matters for real-time assistive communication and cognitive assessment interviews.
  • Separate encoder and decoder prompts let the system model both acoustic deterioration and language-level degradation such as repetitive phrasing and disfluencies, without changing Whisper's frozen parameters.
  • Because the backbone stays frozen and only small prompt experts and a router are trained, the same recipe should transfer to other prompt-conditioned foundation models.
  • The reported gains are statistically significant on two languages, suggesting the effect is not a single-corpus artifact.

Reading between the lines

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

  • The paper leaves implicit that the router's zero-shot guarantee is only as broad as the population coverage of the training clusters; a natural stress test is to measure gains as a function of how far a test speaker's acoustic profile sits from the nearest training cluster centroid.
  • The reported t-SNE separation of decoder prompts by Alzheimer's disease versus non-AD labels suggests the prompt-expert representations could be repurposed as features for cognitive screening, though the paper does not train such a detector.
  • The cluster-and-route mechanism is not tied to Whisper; any conditional model that accepts soft prompts could in principle use the same design, with the same latency argument.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. MOPSA proposes a zero-shot, online speaker-adaptation method for Whisper on elderly speech. Speaker adaptive training learns per-speaker encoder and decoder prompts together with LoRA weights; K-means clustering converts the training speakers' SAT prompts into C prompt experts; a router network with a multi-task loss (ASR, speaker-ID, and MSE alignment to the SAT prompts) then combines these experts in a single forward pass. Experiments on DementiaBank Pitt (English) and JCCOCC MoCA (Cantonese) compare online MOPSA with the speaker-independent baseline, batch-mode prompt adaptation, LHUC/RAB, and i/x/ECAPA embeddings, reporting WER/CER reductions and an RTF speed-up of up to 16.12 times over batch-mode adaptation.

Significance. If the router's speaker-specific contribution were established, MOPSA would be a practical contribution: it is, to my knowledge, the first mixture-of-prompt-experts adaptation method for elderly speech, it operates in a single pass with low latency, and it models both acoustic and language-level variability through separate encoder and decoder prompts. The paper reports statistical significance tests, uses two languages, and evaluates on held-out speakers. However, the central interpretive claim of zero-shot speaker-specific adaptation lacks a decisive control, and the headline numbers are split-selected; both issues need to be resolved before the contribution is fully supported.

major comments (3)
  1. [Abstract; Section 5, Table 1] The abstract states that online MOPSA outperforms the SI model by statistically significant WER/CER reductions of 0.86% and 1.47% absolute. In Table 1, the 0.86% WER number is the gain on the Investigator (Inv) split of DementiaBank (Sys.1 Inv=20.43 vs Sys.12 Inv=19.57). On the Eval split the improvement is only 0.24% absolute and is not marked significant, and the Participant (Par) split is also not marked significant. The claim should either name the split explicitly or use the least favorable split in the headline, since the current wording overstates the consistency of the improvement.
  2. [Section 4, Eq. (7); Section 5, Table 2] The central claim of zero-shot speaker-specific adaptation is not isolated by any baseline. All online variants (Sys.O1-O4 and Sys.10-12) use the router; the i/x/vector and ECAPA conditions are different input features and do not control for a generic prompt benefit. If the router collapses to an input-independent prompt, Whisper's WER can still improve over the SI baseline because any learned prompt can help. Please add controls: (a) a single fixed prompt of the same length trained with the same multi-task losses, (b) uniform weights over the cluster centroids, and (c) a router whose outputs are shuffled or independent of the input. Without these, the numerical gains are compatible with a generic-prompt effect rather than speaker-specific adaptation.
  3. [Section 5, Figure 2] The t-SNE visualization in Fig. 2 shows only two speakers and is anecdotal. The text claims that MOPSA speaker representations are 'more consistent' than i/x-vector and ECAPA representations, but no quantitative criterion or full-evaluation statistics are provided. Report per-speaker embedding consistency (e.g., mean intra-speaker distance normalized by inter-speaker distance) or speaker-ID accuracy on the evaluation speakers.
minor comments (6)
  1. [Abstract; Section 4] The abstract mentions 'Top-K most distinctive speaker prompt clusters', but Section 4 defines C clusters via K-means and does not describe any top-K selection or 'distinctiveness' criterion; the terminology should be aligned.
  2. [Table 1] The column header is difficult to parse because the 'Par' and 'Inv' labels for DementiaBank and JCCOCC are not visually separated; consider separate tables or clearer column groupings.
  3. [Equation (6)] The K-means formulation does not specify the distance metric, initialization, or convergence criterion; state these details for reproducibility.
  4. [Section 5, Experimental setup] The RTF measurements are not described in enough detail: hardware, batch size, number of decoder passes, and whether batch-mode RTF includes the full test-time adaptation iterations are missing. Adding this information is necessary for the speed-up claim to be reproducible.
  5. [Figure 3] The legend entries such as 'Batch-Mode Enc Len4' are not explained in the caption; specify which line and marker correspond to each system and what the x-axis units denote.
  6. [Section 5, Table 1] The paper reports only asterisks for statistical significance; giving p-values or confidence intervals for the key comparisons would strengthen the significance claims.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; MOPSA's online prompt generation is a learned mapping evaluated on disjoint speakers.

full rationale

The derivation chain is not circular. SAT prompts are estimated on training speakers via Eq. (5); K-means experts are centroids of those prompts via Eq. (6); and the router output in Eq. (7) is a learned function of the input speech and the fixed experts. The MSE term in L_Router does align online prompts to the same SAT prompts used to build the experts, but this is a training regularizer, not a test-time target, and the evaluation speakers are disjoint from training: "No elderly speakers in the training set overlap with those in the development or evaluation sets for both corpora." The reported WER/CER reductions are therefore empirical outcomes rather than identities. The only mild self-referential element is that the cluster count, prompt lengths, and loss weights were selected on the same DementiaBank dev/eval columns reported in Table 1 ("The above settings are used for all the main experiments in Table 1"), which weakens the independence of that dataset's headline number but does not force the result by construction; the JCCOCC results use the same settings transferred to a separate dataset. Missing controls, such as a fixed global prompt or uniform expert weighting, are external-validity concerns rather than circularity, and self-citations in the references are baselines or prior related work, not load-bearing premises for the MOPSA construction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central method depends on a handful of empirically chosen hyperparameters (loss weights, prompt lengths, cluster count) and on the domain assumption that prompts learned from training speakers transfer to unseen elderly speakers. No new physical entities are introduced.

free parameters (6)
  • alpha = 0.1
    Interpolation weight for speaker-recognition loss in the router multi-task training (Section 4 footnote).
  • beta_encoder = 0.02
    MSE alignment weight for encoder-side prompts in the router multi-task training (Section 4 footnote).
  • beta_decoder = 0.2
    MSE alignment weight for decoder-side prompts for the Enc&Dec configuration in Table 1 Sys.12 (Section 4 footnote).
  • prompt_length_encoder = 4
    Encoder prompt length selected empirically from ablation (Table 2, B3).
  • prompt_length_decoder = 1
    Decoder prompt length selected empirically from ablation (Table 2, B5).
  • cluster_count_C = 50
    Number of K-means prompt clusters per position, chosen empirically from ablation (Table 2, O9).
assumptions (4)
  • domain assumption K-means clusters of training speaker prompts retain enough speaker variability to represent unseen elderly speakers
    Section 4; the entire online adaptation relies on cluster centroids being transferable.
  • domain assumption The router network trained on training speakers' SAT prompts generalizes to unseen speakers
    Section 4 and Fig. 1(b); zero-shot claim rests on this generalization.
  • domain assumption Whisper-medium with LoRA fine-tuning is a sufficiently strong speaker-independent baseline
    Section 5; all gains are measured against this baseline, not against other modeling choices.
  • standard math MAPSSWE significance testing with alpha=0.05 applies correctly to the paired test-set comparisons
    Section 5; the test is standard in ASR but the paper does not describe its implementation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MOPSA: Mixture of Prompt-Experts Based Speaker Adaptation for Elderly Speech Recognition." pith.science (2026). https://pith.science/paper/4CUNJIMA

@misc{pith2026250524224,
  author       = {Pith},
  title        = {Pith review of: MOPSA: Mixture of Prompt-Experts Based Speaker Adaptation for Elderly Speech Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4CUNJIMA}},
  note         = {Machine review of arXiv:2505.24224}
}
read the original abstract

This paper proposes a novel Mixture of Prompt-Experts based Speaker Adaptation approach (MOPSA) for elderly speech recognition. It allows zero-shot, real-time adaptation to unseen speakers, and leverages domain knowledge tailored to elderly speakers. Top-K most distinctive speaker prompt clusters derived using K-means serve as experts. A router network is trained to dynamically combine clustered prompt-experts. Acoustic and language level variability among elderly speakers are modelled using separate encoder and decoder prompts for Whisper. Experiments on the English DementiaBank Pitt and Cantonese JCCOCC MoCA elderly speech datasets suggest that online MOPSA adaptation outperforms the speaker-independent (SI) model by statistically significant word error rate (WER) or character error rate (CER) reductions of 0.86% and 1.47% absolute (4.21% and 5.40% relative). Real-time factor (RTF) speed-up ratios of up to 16.12 times are obtained over offline batch-mode adaptation.

Figures

Figures reproduced from arXiv: 2505.24224 by the authors.

Figure 1
Figure 1. Examples of (a) the speaker adaptive training, (b) the online mixture of prompt-experts adaptation, and (c) the batch-mode speaker prompts test-time adaptation performed on the encoder of Whisper. enabling efficient online adaptation; 3) In addition to encoder prompts, we also use decoder-level speaker prompts to capture non-acoustic, language-level diversity among elderly speakers. The main contributions of this wo… view at source ↗
Figure 3
Figure 3. WER% of adapted Whisper w.r.t. varying amounts of speaker data on the DementiaBank (Dev+Eval) data. Encoder Prompt Decoder Prompt [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. T-SNE visualization of batch-mode estimated encoder prompts (left, Sys 4, Tab. 1) and decoder prompts (right, Sys 5, Tab. 1) for 48 elderly speakers of DementiaBank Eval set. Speaker-level AD/non-AD labels in orange/blue. Ablation study: As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 43 canonical work pages

  1. [1]

    Elderly speech is often characterized by both imprecise articulations stemming from weakened neuromotor control and linguistic degradation asso- ciated with cognitive decline [1]

    Introduction In the current aging society, ensuring effective communication for the elderly becomes increasingly vital for maintaining their social engagement and quality of life. Elderly speech is often characterized by both imprecise articulations stemming from weakened neuromotor control and linguistic degradation asso- ciated with cognitive decline [1...

    work page

  2. [2]

    The advent of large-scale foundation models [2–5] has further intensified these challenges as their massive parameters require extensive *Corresponding author

    speaker heterogeneity[8, 9] among elderly speakers, where typical sources of variability in speech, such as accent and gen- der, are further compounded by varying degrees of phonetic de- terioration and linguistic expression degradation. The advent of large-scale foundation models [2–5] has further intensified these challenges as their massive parameters ...

    arXiv 2025

  3. [3]

    <|PREV|>,decoder prompt,<|SOT|>,<|LANGUAGE|>,<|TRANSCRIBE|>, <|NO-TIMESTAMP|>

    Large-Scale Foundation Model Whisper Whisper [2] is a transformer-based model capable of processing multiple languages and handling various speech-related tasks. The input to Whisper is log-Mel spectrogramX∈R D×T , whereDandTrespectively denote the feature dimension and length of the input. The convolutional block downsamples the input by a factor of 1/2,...

    work page

  4. [4]

    This process introduces latency and is highly dependent on the quantity and quality of the pseudo-labels

    Batch-Mode Speaker Prompt Adaptation Adaptation Data Accumulation:In batch-mode adaptation, speaker prompts are estimated using first-pass decoding ˆYas pseudo-labels generated by the SI system, which requires sub- stantial data from the testing speaker. This process introduces latency and is highly dependent on the quantity and quality of the pseudo-labe...

    work page

  5. [5]

    query”, “key

    Experiments Task description:The EnglishDementiaBank Pitt[1] corpus is the most widely used publicly available elderly speech cor- pus for speech-based Alzheimer’s Disease (AD) diagnosis. It contains 33 hours of audio recordings from 292 interviews be- tween elderly participants and clinical investigators during AD assessments. The training set consists o...

    work page

  6. [6]

    1(b), K-means clustering is applied to the speaker prompts from SAT, pro- ducing robust and generalizable prompt-experts

    Online Mixture of Prompt-Experts Speaker Adaptation MoE using speaker prompts:To enable online speaker adap- tation for unseen speakers, as shown in Fig. 1(b), K-means clustering is applied to the speaker prompts from SAT, pro- ducing robust and generalizable prompt-experts. ForItrain- ing speakers, the encoder-side prompts for each speaker are represente...

    work page

  7. [7]

    14200220, 14200021, 14200324, Innovation Technology Fund grant No

    Acknowledgements This research is supported by Hong Kong RGC GRF grant No. 14200220, 14200021, 14200324, Innovation Technology Fund grant No. ITS/218/21, China Disabled Persons Federa- tion (CDPF2023KF00002), Basic Research Project of ISCAS (ISCAS-JCMS-202306), Youth Innovation Promotion Associ- ation CAS Grant (2023119), and Guangzhou CASTF project (2022MZK02)

    work page

  8. [8]

    Cluster-based prompt-experts and dynamic router networks allow zero-shot, real-time adaptation to unseen speak- ers

    Conclusion This paper proposes a novel Mixture of Prompt-Experts based Speaker Adaptation approach (MOPSA) for elderly speech recognition. Cluster-based prompt-experts and dynamic router networks allow zero-shot, real-time adaptation to unseen speak- ers. Experiments on DementiaBank Pitt and JCCOCC MoCA elderly speech datasets suggest that our method outp...

    work page

Show all 48 references
  1. [9]

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

    M. Geng, X. Xie, Z. Ye, T. Wang, G. Li, S. Hu, X. Liu, and H. Meng, “Speaker adaptation using spectro-temporal deep fea- tures for dysarthric and elderly speech recognition,”IEEE/ACM T-ASLP, 2022

    2022

  2. [10]

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

    J. T. Becker, F. Boiler, O. L. Lopez, J. Saxton, and K. L. McGo- nigle, “The natural history of alzheimer’s disease: description of study cohort and accuracy of diagnosis,”Archives of neurology, vol. 51, no. 6, pp. 585–594, 1994

    1994

  3. [11]

    Robust Speech Recognition via Large-Scale Weak Supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust Speech Recognition via Large-Scale Weak Supervision,” inICML, 2023

    2023

  4. [12]

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

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,”Advances in neural information processing systems, 2020

    2020

  5. [13]

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

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,”IEEE/ACM T-ASLP, 2021

    2021

  6. [14]

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

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiaoet al., “Wavlm: Large-scale self- supervised pre-training for full stack speech processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  7. [15]

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

    S. Hu, X. Xie, Z. Jin, M. Geng, Y . Wang, M. Cui, J. Deng, X. Liu, and H. Meng, “Exploring Self-supervised Pre-trained ASR Mod- els For Dysarthric and Elderly Speech Recognition,” inICASSP, 2023

    2023

  8. [16]

    Self-supervised asr models and fea- tures for dysarthric and elderly speech recognition,

    S. Hu, X. Xie, M. Geng, Z. Jin, J. Deng, G. Li, Y . Wang, M. Cui, T. Wang, and H. Meng, “Self-supervised asr models and fea- tures for dysarthric and elderly speech recognition,”IEEE/ACM T-ASLP, 2024

    2024

  9. [17]

    Homogeneous speaker features for on- the-fly dysarthric and elderly speaker adaptation,

    M. Geng, X. Xie, J. Deng, Z. Jin, G. Li, T. Wang, S. Hu, Z. Li, H. Meng, and X. Liu, “Homogeneous speaker features for on- the-fly dysarthric and elderly speaker adaptation,”arXiv preprint arXiv:2407.06310, 2024

    2024 arXiv

  10. [18]

    Speaker adaptation for end-to-end ctc models,

    K. Li, J. Li, Y . Zhao, K. Kumar, and Y . Gong, “Speaker adaptation for end-to-end ctc models,” inSLT, 2018

    2018

  11. [19]

    Structured speaker-deficiency adaptation of foundation models for dysarthric and elderly speech recognition,

    S. Hu, X. Xie, M. Geng, J. Deng, Z. Jin, T. Wang, M. Cui, G. Li, Z. Li, H. Menget al., “Structured speaker-deficiency adaptation of foundation models for dysarthric and elderly speech recognition,” arXiv preprint arXiv:2412.18832, 2024

    2024 arXiv

  12. [20]

    Development of the CUHK Elderly Speech Recog- nition System for Neurocognitive Disorder Detection Using the Dementiabank Corpus,

    Z. Ye, S. Hu, J. Li, X. Xie, M. Geng, J. Yu, J. Xu, B. Xue, S. Liu, X. Liuet al., “Development of the CUHK Elderly Speech Recog- nition System for Neurocognitive Disorder Detection Using the Dementiabank Corpus,” inICASSP, 2021

    2021

  13. [21]

    Lin- ear hidden transformations for adaptation of hybrid ANN/HMM models,

    R. Gemello, F. Mana, S. Scanzio, P. Laface, and R. De Mori, “Lin- ear hidden transformations for adaptation of hybrid ANN/HMM models,”Speech Communication, 2007

    2007

  14. [22]

    Comparison of discriminative input and output transformations for speaker adaptation in the hybrid NN/HMM systems,

    B. Li and K. C. Sim, “Comparison of discriminative input and output transformations for speaker adaptation in the hybrid NN/HMM systems,” inINTERSPEECH, 2010

    2010

  15. [23]

    Cluster adaptive training for deep neural network based acoustic model,

    T. Tan, Y . Qian, and K. Yu, “Cluster adaptive training for deep neural network based acoustic model,”IEEE/ACM T-ASLP, 2015

    2015

  16. [24]

    DNN speaker adaptation using parameterised sigmoid and ReLU hidden activation functions,

    C. Zhang and P. C. Woodland, “DNN speaker adaptation using parameterised sigmoid and ReLU hidden activation functions,” in ICASSP, 2016

    2016

  17. [25]

    Low-rank plus diagonal adaptation for deep neural networks,

    Y . Zhao, J. Li, and Y . Gong, “Low-rank plus diagonal adaptation for deep neural networks,” inICASSP, 2016

    2016

  18. [26]

    Subspace lhuc for fast adapta- tion of deep neural network acoustic models

    L. Samarakoon and K. C. Sim, “Subspace lhuc for fast adapta- tion of deep neural network acoustic models.” inINTERSPEECH, 2016

    2016

  19. [27]

    LoRA: Low-Rank Adaptation of Large Language Models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-Rank Adaptation of Large Language Models,” inICLR, 2022

    2022

  20. [28]

    Fast dnn acoustic model speaker adaptation by learning hidden unit contribution features

    X. Xie, X. Liu, T. Lee, and L. Wang, “Fast dnn acoustic model speaker adaptation by learning hidden unit contribution features.” inINTERSPEECH, 2019

    2019

  21. [29]

    Speaker-aware speech- transformer,

    Z. Fan, J. Li, S. Zhou, and B. Xu, “Speaker-aware speech- transformer,” inASRU, 2019

    2019

  22. [30]

    Rapid speaker adaptation for conformer transducer: Attention and bias are all you need

    Y . Huang, G. Ye, J. Li, and Y . Gong, “Rapid speaker adaptation for conformer transducer: Attention and bias are all you need.” in INTERSPEECH, 2021

    2021

  23. [31]

    Robust continuous on-device personalization for au- tomatic speech recognition

    K. C. Sim, A. Chandorkar, F. Gao, M. Chua, T. Munkhdalai, and F. Beaufays, “Robust continuous on-device personalization for au- tomatic speech recognition.” inINTERSPEECH, 2021

    2021

  24. [32]

    Inv. ” and “Par

    process. ForItraining speakers, the speaker prompts are represented asP i ={P i e ,P i d}wherei∈ {1,2, ..., I}indexes the training speaker. LetYdenote the ground truth transcripts andrrepresent the LoRA parameters shared among all speak- ers. The optimization process can be fo...

  25. [33]

    Confidence score based speaker adaptation of conformer speech recognition systems,

    J. Deng, X. Xie, T. Wang, M. Cui, B. Xue, Z. Jin, G. Li, S. Hu, and X. Liu, “Confidence score based speaker adaptation of conformer speech recognition systems,”IEEE/ACM TASLP, 2023

    2023

  26. [34]

    Towards High-Performance and Low-Latency Feature-Based Speaker Adaptation of Conformer Speech Recog- nition Systems,

    J. Deng, X. Xie, G. Li, M. Cui, M. Geng, Z. Jin, T. Wang, S. Hu, Z. Li, and X. Liu, “Towards High-Performance and Low-Latency Feature-Based Speaker Adaptation of Conformer Speech Recog- nition Systems,” inICASSP, 2024

    2024

  27. [35]

    Adap- tive mixtures of local experts,

    R. A. Jacobs, M. I. Jordan, S. J. Nowlan, and G. E. Hinton, “Adap- tive mixtures of local experts,”Neural computation, 1991

    1991

  28. [36]

    Hierarchical mixtures of experts and the EM algorithm,

    M. I. Jordan and R. A. Jacobs, “Hierarchical mixtures of experts and the EM algorithm,”Neural computation, 1994

    1994

  29. [37]

    Parameter-efficient transfer learning for NLP,

    N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly, “Parameter-efficient transfer learning for NLP,” inICML, 2019

    2019

  30. [38]

    Saml: Speaker adaptive mixture of lora experts for end-to-end asr,

    Q. Zhao, G. Sun, C. Zhang, M. Xu, and T. F. Zheng, “Saml: Speaker adaptive mixture of lora experts for end-to-end asr,” in INTERSPEECH, 2024

    2024

  31. [39]

    GPT Understands, Too,

    X. Liu, Y . Zheng, Z. Du, M. Ding, Y . Qian, Z. Yang, and J. Tang, “GPT Understands, Too,”arXiv:2103.10385, 2021

    2021 arXiv

  32. [40]

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

    Y . Jiang, T. Wang, X. Xie, J. Liu, W. Sun, N. Yan, H. Chen, L. Wang, X. Liu, and F. Tian, “Perceiver-Prompt: Flexible Speaker Adaptation in Whisper for Chinese Disordered Speech Recognition,” inINTERSPEECH, 2024

    2024

  33. [41]

    A compact model for speaker-adaptive training,

    T. Anastasakos, J. McDonough, R. Schwartz, and J. Makhoul, “A compact model for speaker-adaptive training,” inICSLP, 1996

    1996

  34. [42]

    Speaker turn aware similarity scoring for diarization of speech- based cognitive assessments,

    S. S. Xu, M.-W. Mak, K. H. Wong, H. Meng, and T. C. Kwok, “Speaker turn aware similarity scoring for diarization of speech- based cognitive assessments,” inAPSIPA ASC, 2021

    2021

  35. [43]

    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,” inICASSP, 1989

    1989

  36. [44]

    Advancing speaker embedding learning: Wespeaker toolkit for research and produc- tion,

    S. Wang, Z. Chen, B. Han, H. Wang, C. Liang, B. Zhang, X. Xi- ang, W. Ding, J. Rohdin, A. Silnovaet al., “Advancing speaker embedding learning: Wespeaker toolkit for research and produc- tion,”Speech Communication, 2024

    2024

  37. [45]

    Attention is All you Need,

    A. Waswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. Gomez, L. Kaiser, and I. Polosukhin, “Attention is All you Need,” inNeurIPS, 2017

    2017

  38. [46]

    Exploiting prompt learning with pre-trained language models for alzheimer’s disease detection,

    Y . Wang, J. Deng, T. Wang, B. Zheng, S. Hu, X. Liu, and H. Meng, “Exploiting prompt learning with pre-trained language models for alzheimer’s disease detection,” inICASSP, 2023

    2023

  39. [47]

    A comparative study of acoustic and linguistic features classification for alzheimer’s disease detection,

    J. Li, J. Yu, Z. Ye, S. Wong, M. Mak, B. Mak, X. Liu, and H. Meng, “A comparative study of acoustic and linguistic features classification for alzheimer’s disease detection,” inICASSP, 2021

    2021

  40. [48]

    Conformer based elderly speech recogni- tion system for alzheimer’s disease detection,

    T. Wang, J. Deng, M. Geng, Z. Ye, S. Hu, Y . Wang, M. Cui, Z. Jin, X. Liu, and H. Meng, “Conformer based elderly speech recogni- tion system for alzheimer’s disease detection,” inINTERSPEECH, 2022

    2022

Pith tools

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