Pith. sign in

REVIEW 3 major objections 7 minor 57 references

Exploring Generative Error Correction for Dysarthric Speech Recognition

T0 review · 3 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper claims a two-stage pipeline—Whisper N-best hypotheses plus FlanT5 generative error correction—improves dysarthric speech recognition, with best results at 6.40% WER on development and 12.89% on TEST-2, while single-word…

desk verdict A well-executed challenge-system paper that shows real but modest gains from LLM-based error correction on dysarthric speech; the main weakness is that it never proves the correct transcript is present in the N-best list, which leaves the central mechanism under-supported. read the letter →

arxiv 2505.20163 v1 pith:6JCYMVPU submitted 2025-05-26 cs.CL eess.AS

classification cs.CLeess.AS
keywords dysarthricspeechrecognitiongenerativeerrorcorrectionWhisperFlanT5AccessibilityProjectN-besthypothesesLoRAfine-tuningsemanticscore
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

Using the Speech Accessibility Project Challenge data, the authors try to establish that a two-stage framework—a Whisper ASR model producing N-best hypotheses followed by a FlanT5-based generative error correction (GER) model—reliably improves transcription of dysarthric speech. They report that combining fine-tuned Whisper large-v2 with an 11B FlanT5 GER stage gives the best results (6.40% WER, 92.47 SemScore on development; 12.89% WER on the hidden TEST-2 set), better than ASR alone. The paper argues this works because general-purpose ASR captures enough acoustic evidence inside its top hypotheses, and the LLM can pick out the correct transcription by looking across diverse candidates. It also shows a clear boundary: single-word utterances remain at 63.08% WER with no improvement from GER, indicating the framework handles structured and spontaneous speech but not isolated words.

What carries the argument

The central mechanism is the N-best-to-text error-correction readout. Stage 1 uses Whisper (large-v3 zero-shot or fine-tuned large-v2) to generate 20 beam-search hypotheses, then a diversity-based selector keeps five hypotheses that maximize minimum normalized edit distance, preserving the top-1 hypothesis. Stage 2 feeds these five hypotheses to FlanT5 (3B or 11B parameters) through a structured prompt, fine-tuned with LoRA, to output the corrected transcription. This design isolates acoustic modeling from linguistic correction, letting the authors attribute gains to the LLM's ability to identify a consistent, contextually coherent interpretation across the candidates.

What would settle it

Compare the oracle WER of the 20-best hypotheses on the development set with the 6.40% WER of the full pipeline; if the oracle WER is not substantially below the top-1 WER, then the GER stage cannot be recovering acoustic evidence, and if it is much lower than 6.40%, the pipeline is not fully exploiting its N-best list. Either measurement would settle the central claim.

Watch

Extended reading notes

Core claim

On the Speech Accessibility Project data, the authors find that general-purpose ASR models (Whisper large-v3 zero-shot) already transcribe dysarthric speech to 11.60% WER on development, and that a fine-tuned Whisper large-v2 reaches 7.17% WER. Adding a generative error-correction stage based on FlanT5, prompted with five diverse N-best hypotheses, further lowers WER to 6.40% and raises SemScore to 92.47, and gives the best TEST-2 result at 12.89% WER. The authors interpret this as evidence for two complementary claims: the ASR model retains the necessary acoustic information in its ranked hypotheses, and an LLM can select the correct transcription by reasoning over those hypotheses. They also report a sharp limitation: single-word utterances sit at 63.08% WER, unchanged by GER, because the system biases isolated words toward short phrases.

Load-bearing premise

The correct transcription must appear among the ASR N-best hypotheses often enough for the GER model to recover it; the paper never reports oracle N-best WER, so this premise is unmeasured.

Editorial extensions

If this is right

  • Fine-tuning the acoustic model and adding GER are complementary: the best configuration pairs fine-tuned Whisper large-v2 with an 11B FlanT5 GER stage, reaching 6.40% WER on development and 12.89% on TEST-2.
  • GER improves all speaking styles except single words: WER drops by 0.86% absolute on digital assistant commands, 0.53% on novel sentences, and 0.55% on spontaneous speech, with semantic score gains on every category except single words.
  • Single-word recognition is the unsolved boundary of the framework, at 63.08% WER with no GER improvement, indicating a bias toward full utterances.
  • Scaling the GER model from 3B to 11B parameters yields only small gains on test sets, so computational budget may be better spent elsewhere.
  • The full ASR+GER pipeline runs at 0.69s per sample on a single A100, versus 0.55s for ASR alone, keeping the method practical.

Reading between the lines

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

  • The paper does not report oracle N-best WER, so the true ceiling of the approach is unknown; if the oracle WER on the development 20-best list is much lower than the pipeline's 6.40%, a stronger hypothesis-selection or reranking method could close the gap without a larger LLM.
  • The single-word failure suggests a decoding-level bias rather than an acoustic one; a GER prompt that explicitly constrains output to a single word, or a constrained decoding head, is a testable extension the paper does not explore.
  • The same two-stage recipe should transfer to other atypical speech varieties and to any ASR that exposes N-best lists, since the diversity-based selection algorithm and the LLM prompt are model-agnostic.
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 / 7 minor

Summary. This paper presents a two-stage framework for dysarthric speech recognition evaluated on the Speech Accessibility Project Challenge. Stage one uses Whisper (zero-shot or fine-tuned) to generate an N-best list of hypotheses; a diversity-based selection algorithm reduces the 20-best list to 5 hypotheses; stage two uses a LoRA-tuned FlanT5 model to perform generative error correction (GER) over the selected hypotheses. Experiments compare zero-shot and fine-tuned ASR with and without GER on development, TEST-1, and TEST-2, and report the best result for fine-tuned Whisper large-v2 with GER (6.40% WER on development, 12.89% on TEST-2). The paper also breaks down performance by utterance type, showing gains for commands, novel sentences, and spontaneous speech, and no improvement for isolated single words.

Significance. If the reported results hold, the paper offers a practical recipe for dysarthric ASR within a challenge setting, with a public code link, consistent improvements from GER over zero-shot ASR on all three evaluation sets, and a candid analysis of the single-word failure mode. The separation of acoustic and linguistic components is a useful framework, and the inference-time numbers suggest the pipeline is feasible. However, the central mechanistic claim—that correct transcriptions are present in the ASR N-best list and recoverable by the LLM—is not directly evaluated, and the marginal GER gains over fine-tuned ASR on the test sets lack any significance testing. These gaps limit the strength of the 'complementary roles' conclusion.

major comments (3)
  1. [Section 4.2, Table 2] The N-best analysis does not actually test whether the correct transcription is present in the hypothesis list. Table 2 reports the WER of the single top-ranked hypothesis for different beam widths N (the N=20 row matches the no-GER Lv3 baseline in Table 1), not the oracle WER of the best hypothesis in the N-best list. The claim in Section 1 that 'more correct alternatives may appear among lower-ranked candidates' and Research Question 1 both require an oracle measurement. Without this, the GER gains in Tables 1 and 3 could be produced by the LLM generating plausible text from linguistic priors rather than selecting acoustic evidence from lower-ranked hypotheses. Please report oracle WER (and oracle SemScore) for the Lv3 zero-shot and Lv2 fine-tuned ASR configurations at N=5, 10, and 20 on the development set.
  2. [Section 4.1, Table 1] The 'complementary benefits' claim is supported by small absolute differences on the test sets without any measure of variability or statistical significance. The GER gain over fine-tuned Whisper large-v2 is 0.77 WER on development, but only 0.23 on TEST-1 and 0.15 on TEST-2. Since no confidence intervals or significance tests are reported, these marginal gains may be within run-to-run variation. Please provide paired significance tests (e.g., bootstrap or McNemar) or multiple runs with standard deviations for the key comparisons, or clearly state if the test-set differences are not statistically significant.
  3. [Section 3.3, Table 1] The effect of fine-tuning is not cleanly quantified. The paper states 'Fine-tuning Whisper LARGE-V2 provides substantial gains' (Section 4.1) but Table 1 does not include a Whisper LARGE-V2 zero-shot row; the implicit baseline is LARGE-V3 zero-shot, which conflates model version with fine-tuning. The reported instability of LARGE-V3 fine-tuning (Section 3.3) is anecdotal and not shown. Please add a LARGE-V2 zero-shot row to Table 1, or explicitly frame the comparison as cross-model, and provide evidence (e.g., a learning curve or example outputs) for the claimed LARGE-V3 fine-tuning instability.
minor comments (7)
  1. [Abstract and Section 1] The abstract uses 'we proposed' while Section 1 uses 'We propose'; please make the tense consistent.
  2. [Section 2.1] The procedure for concatenating N-best hypotheses when long recordings are split into multiple segments is under-specified: it is unclear how variable numbers of segments are handled and how the 20-best list is reconstructed for long-form audio. Please provide a precise procedure or a reference.
  3. [Section 3.4] The SemScore definition says it combines BERTScore, phonetic distance, and natural language inference with 'different weights', but the weights are not given; if SemScore is a challenge-specific metric, please cite the exact source or provide the formula.
  4. [Table 3] The w/ GER columns show entries like '5.61−0.86' but the caption does not explain that these are absolute differences from the no-GER column; please clarify the notation.
  5. [Section 4.3] The sentence 'single words (SW) transcriptions reveals a critical limitation' has a subject-verb agreement error; it should be 'single-word transcriptions reveal'.
  6. [Section 2.2] The GER model is essentially the authors' FlanEC [25]; the connection between the Figure 2 prompt and FlanEC should be made explicit in the text.
  7. [Table 2] The caption of Table 2 should clarify that the reported WER is the top-1 hypothesis after beam search with beam width N, not the oracle over the N-best list, to avoid confusion with the N-best generation described in Section 2.1.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the two-stage pipeline is evaluated on held-out data, and the only self-citation is non-load-bearing.

full rationale

The paper's derivation chain is a two-stage pipeline: Whisper generates a 20-best list, a diversity selector keeps five hypotheses, and FlanT5 fine-tuned with LoRA on training-set (N-best, ground-truth) pairs produces the final transcription. The central claim, that fine-tuning plus GER yields the best held-out WER and SemScore (Table 1), is supported by measurements on the development set and on hidden TEST-2, not by recycling fitted constants. No parameter is fitted to a subset and then reported as a prediction of a closely related quantity. The recoverability premise, namely that the correct transcription can appear among lower-ranked N-best hypotheses, is indeed load-bearing, but it is a factual premise rather than a circular one: the paper offers one qualitative example (Table 4) and an N-best size ablation (Table 2), yet it never reports oracle N-best WER. Table 2 reports the ASR stage's own selected output at different beam widths, and its N=20 row coincides with the no-GER baseline, so it does not establish how often the reference is present in the list. This is an evidence gap, not a definitional reduction. The only self-citation with methodological content is [25] (FlanEC), cited to justify FlanT5 as a post-ASR error corrector; FlanEC is an externally published system with its own benchmark results, and the present paper's gains are independently measured on held-out data, so the citation is not load-bearing. No equation equals its own input by construction, and no fitted parameter is renamed as a prediction. Therefore, no circular step can be exhibited.

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

The central claim depends mainly on domain assumptions about N-best oracle availability and the transferability of augmentation and LLM correction to dysarthric speech. No new entities are introduced, and the only hand-set numbers are standard training hyperparameters that were not swept; they are not fitted to the evaluation targets.

free parameters (5)
  • N-best pool size N = 20
    Hand-chosen pool size for ASR hypotheses in Stage 1; Table 2 shows most gains come from the first 5 candidates, so 20 is a computational design choice.
  • Selected hypotheses k = 5
    Number of diverse hypotheses passed to GER; fixed by the diversity selection algorithm and used in all main experiments.
  • LoRA rank and alpha = r=16, alpha=32
    LoRA configuration for FlanT5 fine-tuning; standard values chosen without sensitivity analysis.
  • Data augmentation probabilities = 50% noise placement, 1% pure noise, 25% time stretch or SpecAugment
    Probabilities from the CrisperWhisper-inspired pipeline, hand-set and not ablated.
  • Training epochs and learning rates = 5 epochs/5e-6 (ASR), 10 epochs/1e-4 (GER)
    Hyperparameters fixed during training; no sweep or early stopping reported.
assumptions (5)
  • domain assumption The correct transcription occurs in the ASR N-best list frequently enough for GER to recover it.
    Central motivation for Stage 2 (Section 2.1); Table 4 gives one supportive example, but no oracle N-best WER is reported (Section 4.2).
  • domain assumption Whisper's acoustic representations preserve information about dysarthric speech even when the top-1 hypothesis is wrong.
    Research question 1 (Section 1) assumes acoustic capacity; supported only indirectly by N-best gains.
  • domain assumption FlanT5 can learn to map noisy hypotheses to verbatim transcriptions containing disfluencies and parenthetical markers.
    GER is trained on SAPC pairs; assumes the target format with round brackets is learnable and that LoRA suffices (Section 3.3).
  • domain assumption CrisperWhisper-style augmentation (noise placement, noise-only samples, time stretch, SpecAugment) transfers to dysarthric speech.
    Section 3.1 adopts these augmentations without dysarthria-specific validation.
  • standard math The challenge evaluation protocol, including reference selection by lower edit distance and SemScore composition, is a faithful measure of transcription quality.
    Section 3.4: metrics and reference selection follow challenge guidelines; their validity is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Generative Error Correction for Dysarthric Speech Recognition." pith.science (2026). https://pith.science/paper/6JCYMVPU

@misc{pith2026250520163,
  author       = {Pith},
  title        = {Pith review of: Exploring Generative Error Correction for Dysarthric Speech Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6JCYMVPU}},
  note         = {Machine review of arXiv:2505.20163}
}
read the original abstract

Despite the remarkable progress in end-to-end Automatic Speech Recognition (ASR) engines, accurately transcribing dysarthric speech remains a major challenge. In this work, we proposed a two-stage framework for the Speech Accessibility Project Challenge at INTERSPEECH 2025, which combines cutting-edge speech recognition models with LLM-based generative error correction (GER). We assess different configurations of model scales and training strategies, incorporating specific hypothesis selection to improve transcription accuracy. Experiments on the Speech Accessibility Project dataset demonstrate the strength of our approach on structured and spontaneous speech, while highlighting challenges in single-word recognition. Through comprehensive analysis, we provide insights into the complementary roles of acoustic and linguistic modeling in dysarthric speech recognition

Figures

Figures reproduced from arXiv: 2505.20163 by the authors.

Figure 1
Figure 1. Overview of our two-stage framework for dysarthric speech recognition. Stage 1 uses the ASR model to generate 20-best hypotheses from the input audio. Stage 2 selects diverse hypotheses and employs the GER model to analyze them collectively, producing a refined final transcription. different interpretations of unclear or challenging speech seg￾ments. For long recordings, WhisperX handles segmentation using integrate… view at source ↗
Figure 2
Figure 2. GER model prompt. We train the GER model using LoRA [26] to maintain efficiency while specializing in transcription correction. This approach allows the model to leverage patterns across hypotheses and its own language understanding to resolve ambiguities in the ASR output. The GER stage serves multiple purposes: (i) ana￾lyzes patterns across multiple hypotheses to identify consistent elements, (ii) leverages lingui… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 50 canonical work pages

  1. [1]

    V oice-controlled systems now support tasks ranging from simple commands to complex conversations

    Introduction Speech recognition technology has become essential in daily life, enabling natural interaction with devices and applications [1–4]. V oice-controlled systems now support tasks ranging from simple commands to complex conversations. However, these advances have not benefited all users equally [5–7]. Indi- viduals with dysarthric speech face sig...

  2. [2]

    Do general-purpose ASR models possess sufficient acoustic modeling capacity to capture dysarthric speech patterns, even if they struggle with producing correct transcriptions?

  3. [3]

    Our approach first uses ASR models to generate multiple transcription hypotheses, captur- ing different interpretations of the acoustic input

    Can large language models leverage their linguistic knowl- edge to identify correct transcriptions by analyzing patterns across multiple ASR hypotheses while ensuring grammatical and contextual coherence? We propose a two-stage framework that combines ASR with generative error correction. Our approach first uses ASR models to generate multiple transcripti...

  4. [4]

    Exploring Generative Error Correction for Dysarthric Speech Recognition

    Methodology Our work explores (i) whether general ASR systems can cap- ture useful information about dysarthric speech in their N-best hypotheses, and (ii) if language models can effectively distill this information. To investigate these research questions, we develop a two-stage framework that combines ASR with Gen- erative Error Correction (GER) (Figure...

  5. [6]

    My favorite hobbies are to go dancing and carry a key

  6. [7]

    A con- trastive learning approach to mitigate bias in speech models,

    A. Koudounas, F. Giobergia, E. Pastor, and E. Baralis, “A con- trastive learning approach to mitigate bias in speech models,” in Proc. INTERSPEECH 2024, 2024, pp. 827–831

  7. [8]

    My favourite hobby is to go dancing and karaoke

  8. [9]

    Preferred of Bobby is school dancing and car hockey

Show all 57 references
  1. [10]

    Given the following n-best list of hypotheses from ASR, provide the correct transcription: My favourite hobby is to go dancing and karaoke

    My favorite hobbies are to go dancing and carry a key. . . . Given the following n-best list of hypotheses from ASR, provide the correct transcription: My favourite hobby is to go dancing and karaoke. My favorite hobbies are school dancing and karaoke.GER Model Stage 1Stage 2 ...

  2. [11]

    Retain the top-scoring hypothesis to preserve the highest con- fidence transcription

  3. [12]

    Calculate normalized edit distances between all remaining hypotheses

  4. [13]

    Iteratively select hypotheses that maximize the minimum dis- tance to previously selected ones This approach is designed to select hypotheses that represent genuinely different interpretations rather than minor variations of the same transcription. 2.2. Generative Error Correc...

  5. [14]

    Experimental Setup Our experimental evaluation primarily uses the Speech Accessi- bility Project Challenge (SAPC) dataset [18], containing 105.76 hours of dysarthric speech in the training set and 39.56 hours in the development set. The dataset consists of multiple utterance t...

  6. [15]

    Following CrisperWhisper’s positional augmentation strat- egy, we inject background noise from MUSAN dataset [30] at either the beginning or end of each audio segment with 50% Table 1: Performance comparison across ASR and GER configurations. ASR Model FT GER Dev TEST-1 TEST-2...

  7. [16]

    This technique aims to mitigate hallucinations, a common issue in generative ASR methods like Whisper

    We introduce pure noise samples from MUSAN [30] during training with 1% probability and require empty predictions as output. This technique aims to mitigate hallucinations, a common issue in generative ASR methods like Whisper

  8. [17]

    (che- che-) checkout

    With 25% probability, we apply either time stretching (0.85x- 1.15x speed range) to address variable speaking rates in dysarthric speech, or SpecAugment [31] with frequency and time masking to enhance spectral robustness. 3.2. Transcription Format The SAPC dataset provides det...

  9. [18]

    Our analysis focuses on understanding how acoustic and lin- guistic modeling capabilities contribute to dysarthric speech recognition

    Results and Analysis We evaluate our framework through a series of experiments examining both overall performance and specific components. Our analysis focuses on understanding how acoustic and lin- guistic modeling capabilities contribute to dysarthric speech recognition. 4.1...

  10. [19]

    My favorite play is the one that’s set on Monday

  11. [20]

    ASR only GER Pipeline

    Preferred of Bobby is school dancing and car hockey. . . . ASR only GER Pipeline

  12. [21]

    My favorite pet is the one that sits on my lap

  13. [22]

    My favorite player is the one that’s in Orlando

  14. [23]

    My favorite play is the ones that sit on the

  15. [24]

    football

    My favorite pick is the one that said ”Wonder.” which represent the majority of the dataset, GER reduces WER by 0.86% absolute while improving semantic scores by 0.51 points. Similar improvements appear in novel sentences (SN), with a 0.53% WER reduction and a 0.45-point incre...

  16. [25]

    Conclusions Our investigation through the Speech Accessibility Project Challenge reveals both opportunities and challenges in dysarthric speech recognition. The strong performance of zero- shot Whisper (11.60% WER) answers our first research ques- tion, demonstrating that gene...

  17. [26]

    Acknowledgements This work has been partially supported by the ”D.A.R.E. – Dig- ital Lifelong Prevention” project (code: PNC0000002, CUP: B53C22006450001), co-funded by the Italian Complemen- tary National Plan PNC-I.1 Research initiatives for innovative technologies and pathw...

  18. [27]

    Superb: Speech processing universal perfor- mance benchmark,

    S. wen Yang et al., “Superb: Speech processing universal perfor- mance benchmark,” in Interspeech 2021, 2021, pp. 1194–1198

  19. [28]

    Italic: An italian intent classification dataset,

    A. Koudounas, M. La Quatra, L. Vaiani, L. Colomba, G. Attana- sio, E. Pastor, L. Cagliero, and E. Baralis, “Italic: An italian intent classification dataset,” inInterspeech 2023, 2023, pp. 2153–2157

  20. [29]

    Speech-massive: A multilingual speech dataset for slu and be- yond,

    B. Lee, I. Calapodescu, M. Gaido, M. Negri, and L. Besacier, “Speech-massive: A multilingual speech dataset for slu and be- yond,” Interspeech 2024, 2024

  21. [30]

    Benchmarking representations for speech, music, and acoustic events,

    M. La Quatra, A. Koudounas, L. Vaiani, E. Baralis, L. Cagliero, P. Garza, and S. M. Siniscalchi, “Benchmarking representations for speech, music, and acoustic events,” in 2024 IEEE Interna- tional Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). IEEE...

  22. [31]

    Exploring subgroup performance in end-to-end speech models,

    A. Koudounas, E. Pastor, G. Attanasio, V . Mazzia, M. Giollo, T. Gueudre, L. Cagliero, L. de Alfaro, E. Baralis, and D. Amberti, “Exploring subgroup performance in end-to-end speech models,” in ICASSP 2023 - 2023 IEEE International Conference on Acous- tics, Speech and Signal ...

  23. [32]

    Towards comprehensive subgroup performance analysis in speech models,

    A. Koudounas, E. Pastor, G. Attanasio, V . Mazzia, M. Giollo, T. Gueudre, E. Reale, L. Cagliero, S. Cumani, L. de Alfaro, E. Baralis, and D. Amberti, “Towards comprehensive subgroup performance analysis in speech models,”IEEE/ACM Transactions on Audio, Speech, and Language Pro...

  24. [33]

    Training data augmentation for dysarthric automatic speech recognition by text- to-dysarthric-speech synthesis,

    W.-Z. Leung, M. Cross, A. Ragni, and S. Goetze, “Training data augmentation for dysarthric automatic speech recognition by text- to-dysarthric-speech synthesis,” in Interspeech, 2024

  25. [34]

    Dysarthric speech recognition: A comparative study,

    D. Shah, V . Lal, Z. Zhong, Q. Wang, and S. R. Shahamiri, “Dysarthric speech recognition: A comparative study,” in 2023 International Conference on Speech Technology and Human- Computer Dialogue (SpeD), 2023, pp. 89–94

  26. [35]

    New spanish speech cor- pus database for the analysis of people suffering from parkinson’s disease

    J. R. Orozco-Arroyave, J. D. Arias-Londo ˜no, J. F. Vargas-Bonilla, M. C. Gonzalez-R ´ativa, and E. N ¨oth, “New spanish speech cor- pus database for the analysis of people suffering from parkinson’s disease.” in Lrec, 2014, pp. 342–347

  27. [36]

    V oice disorder analysis: a transformer-based approach,

    A. Koudounas, G. Ciravegna, M. Fantini, E. Crosetti, G. Succo, T. Cerquitelli, and E. Baralis, “V oice disorder analysis: a transformer-based approach,” Interspeech 2024, 2024

  28. [37]

    Exploiting foundation models and speech enhancement for parkinson’s disease detection from speech in real-world operative conditions,

    M. L. Quatra, M. F. Turco, T. Svendsen, G. Salvi, J. R. Orozco- Arroyave, and S. M. Siniscalchi, “Exploiting foundation models and speech enhancement for parkinson’s disease detection from speech in real-world operative conditions,” Interspeech 2024 , 2024

  29. [38]

    Bilingual dual-head deep model for parkinson’s disease detection from speech,

    M. La Quatra, J. R. Orozco-Arroyave, and M. S. Siniscalchi, “Bilingual dual-head deep model for parkinson’s disease detection from speech,” inICASSP 2025 - 2025 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

  30. [39]

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

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,” Advances in neural information processing systems , vol. 33, pp. 12 449–12 460, 2020

  31. [40]

    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 transactions on audio, speech, and language processing, 2021

  32. [41]

    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,” in International conference on machine learning . PMLR, 2023, pp. 28 492–28 518

  33. [42]

    voc2vec: A foundation model for non-verbal vocalization,

    A. Koudounas, M. La Quatra, S. M. Siniscalchi, and E. Baralis, “voc2vec: A foundation model for non-verbal vocalization,” in ICASSP 2025 - 2025 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP), 2025

  34. [43]

    Community-supported shared in- frastructure in support of speech accessibility,

    M. Hasegawa-Johnson et al., “Community-supported shared in- frastructure in support of speech accessibility,”Journal of Speech, Language, and Hearing Research, 2024

  35. [44]

    Enhancing pre-trained asr system fine-tuning for dysarthric speech recognition using adversarial data augmenta- tion,

    H. Wang, Z. Jin, M. Geng, S. Hu, G. Li, T. Wang, H. Xu, and X. Liu, “Enhancing pre-trained asr system fine-tuning for dysarthric speech recognition using adversarial data augmenta- tion,” in ICASSP, 2024

  36. [45]

    Dysarthric speech recognition using curriculum learning and articulatory feature embedding,

    I.-T. Hsieh and C.-H. Wu, “Dysarthric speech recognition using curriculum learning and articulatory feature embedding,” inInter- speech 2024, 2024, pp. 1300–1304

  37. [46]

    Whisperx: Time- accurate speech transcription of long-form audio,

    M. Bain, J. Huh, T. Han, and A. Zisserman, “Whisperx: Time- accurate speech transcription of long-form audio,” in Interspeech 2023, 2023, pp. 4489–4493

  38. [47]

    Optimization of rnn-based speech activity detection,

    G. Gelly and J.-L. Gauvain, “Optimization of rnn-based speech activity detection,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 26, no. 3, pp. 646–656, 2018

  39. [48]

    Pyannote.audio: Neural building blocks for speaker diarization,

    H. Bredin, R. Yin, J. M. Coria, G. Gelly, P. Korshunov, M. Lavechin, D. Fustes, H. Titeux, W. Bouaziz, and M.-P. Gill, “Pyannote.audio: Neural building blocks for speaker diarization,” in ICASSP, 2020

  40. [49]

    Scaling instruction-finetuned language mod- els,

    H. W. Chung et al., “Scaling instruction-finetuned language mod- els,” Journal of Machine Learning Research, vol. 25, no. 70, pp. 1–53, 2024

  41. [50]

    Flanec: Exploring flan-t5 for post-asr error correction,

    M. La Quatra, V . M. Salerno, Y . Tsao, and S. M. Siniscalchi, “Flanec: Exploring flan-t5 for post-asr error correction,” in 2024 IEEE Spoken Language Technology Workshop (SLT) . IEEE, 2024, pp. 608–615

  42. [51]

    LoRA: Low-rank adaptation of large lan- guage models,

    E. J. Hu, yelong shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large lan- guage models,” in International Conference on Learning Repre- sentations, 2022

  43. [52]

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

    F. Rudzicz, A. K. Namasivayam, and T. Wolff, “The torgo database of acoustic and articulatory speech from speakers with dysarthria,” Language resources and evaluation, vol. 46, pp. 523– 541, 2012

  44. [53]

    V oxpopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation,

    C. Wang, M. Riviere, A. Lee, A. Wu, C. Talnikar, D. Haz- iza, M. Williamson, J. Pino, and E. Dupoux, “V oxpopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation,” 2021

  45. [54]

    Crisperwhisper: Accu- rate timestamps on verbatim speech transcriptions,

    M. Zusag, L. Wagner, and B. Thallinger, “Crisperwhisper: Accu- rate timestamps on verbatim speech transcriptions,” Interspeech 2024, 2024

  46. [55]

    Musan: A music, speech, and noise corpus,

    D. Snyder, G. Chen, and D. Povey, “Musan: A music, speech, and noise corpus,” arXiv preprint arXiv:1510.08484, 2015

  47. [56]

    Specaugment: A simple data augmentation method for automatic speech recognition,

    D. S. Park, W. Chan, Y . Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V . Le, “Specaugment: A simple data augmentation method for automatic speech recognition,” in Interspeech, 2019

  48. [57]

    Decoupled weight decay regulariza- tion,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regulariza- tion,” in International Conference on Learning Representations , 2019

  49. [58]

    Bertscore: Evaluating text generation with bert,

    T. Zhang*, V . Kishore*, F. Wu*, K. Q. Weinberger, and Y . Artzi, “Bertscore: Evaluating text generation with bert,” inInternational Conference on Learning Representations, 2020

Pith tools

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