Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

Analysis and Evaluation of Synthetic Data Generation in Speech Dysfluency Detection

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that prompting LLMs to write dysfluent text and synthesizing it with TTS produces a 12,790-hour corpus, and that a detector trained on it beats prior results on real stuttered speech.

desk verdict The dataset is a genuinely useful resource; the claimed SOTA on real speech is not supported by the evaluation. read the letter →

arxiv 2505.22029 v2 pith:RQXFBFHS submitted 2025-05-28 eess.AS cs.AIcs.SD

Jinming Zhang , Xuanru Zhou , Jiachen Lian , Shuhe Li , William Li , Zoe Ezzes , Rian Bogley , Lisa Wauters
show 5 more authors
Zachary Miller Jet Vonk Brittany Morin Maria Gorno-Tempini Gopala Anumanchipalli
This is my paper · ORCID
classification eess.AScs.AIcs.SD
keywords speechdysfluencydetectionsyntheticdatagenerationlargelanguagemodelstext-to-speechstutteringcorpusWhisperSEP-28k
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

Speech dysfluency detection is held back by scarce, inconsistently annotated real stutter data. The paper's central claim is that synthetic data can close that gap when the text simulations are written by a large language model rather than hand-built rules: it introduces LLM-Dys, a corpus of over 12,790 hours covering 11 dysfluency types at word and phoneme levels, which it reports sounds as natural as fluent speech and more natural than prior simulated corpora. Fine-tuning a token-based Whisper detector on this corpus yields state-of-the-art F1 scores on real stuttered speech benchmarks, SEP-28k and UCLASS, while the authors also document a scaling plateau and an optimal fluent-to-disfluent training ratio.

What carries the argument

The load-bearing object is the paired dysfluent-text-plus-label generator, where the LLM supplies authentic dysfluency patterns from its learned language model and its output tokens double as training labels, eliminating manual annotation. The TTS stage uses VITS's duration-prediction matrix to realize pauses and prolongations at precise timestamps, and E2-TTS for filler-word insertions, which is what keeps the synthesized audio natural rather than rule-pasted. On the detection side, Whisper-large-v3-turbo is fine-tuned to transcribe dysfluency tokens inserted into clean text, outputting dysfluency type and location jointly for eleven categories.

What would settle it

Have independent annotators label the same SEP-28k subset using the dataset's original annotation scheme, then rerun the zero-shot evaluation; if the LLM-Dys model's F1 falls below the reported baselines (for example, word insertion 0.77, word repetition 0.64, word pause 0.62 from Wagner et al.), the state-of-the-art claim on real stuttered speech collapses.

Watch

Extended reading notes

Core claim

LLM-Dys rests on a two-stage pipeline: first a large language model rewrites clean sentences into dysfluent text, generating word-level insertions, repetitions, pauses, deletions, and substitutions plus phoneme-level versions from CMU/IPA transcriptions, while emitting matching token labels at the same time. These texts are synthesized with the VITS text-to-speech model, with pauses inserted as silent gaps of 0.3-3.5 seconds and prolongations produced by stretching phoneme durations in VITS's duration matrix, and E2-TTS is used for word-level insertions because VITS handles filler words poorly. The authors report that this pipeline produces higher Audiobox Aesthetics scores than VCTK-Token, VCTK, LibriTTS, and LibriStutter, and that their Whisper-large-v3-turbo transcriber reaches F1 0.91 for word insertion, 0.67 for word repetition, and 0.79 for word pause on a self-annotated SEP-28k test subset, beating prior published numbers, plus 0.977 word-level accuracy on UCLASS after fine-tuning on only 80 clips.

Load-bearing premise

The narrowest load-bearing premise is that a test set the authors built themselves, by hand-labeling 200 SEP-28k clips per level in the model's own token format, fairly measures real-world dysfluency detection.

Editorial extensions

If this is right

  • If correct, end-to-end dysfluency detection can be trained almost entirely on synthetic speech, with real data reserved for a small test set.
  • The eleven-type, word-and-phoneme unified annotation scheme gives one detector format for tasks that previously used binary, event-level, and ASR-filler labels separately.
  • The scaling plateau at 3,000-4,000 samples per type implies that far less than the full 12,790 hours may suffice, allowing cheap subsampling for fast iteration.
  • The roughly 0.05 optimal fluent-to-disfluent ratio tells practitioners to keep fluent speech in the training mix, or synthetic-only models will over-predict dysfluencies.
  • The UCLASS result suggests that synthetic pretraining transfers to binary stuttering classification with minimal real fine-tuning.

Reading between the lines

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

  • Inference: if synthetic dysfluent speech keeps closing the quality gap with real recordings, the field can benchmark detectors on procedurally generated patients, dialects, and speaking styles, making evaluation fair for groups underrepresented in SEP-28k and UCLASS.
  • Inference: the SEP-28k evaluation relies on a self-annotated 200-sample subset labeled in the model's own token format; a natural next test is the original SEP-28k annotations used by the compared baseline, and if the advantage shrinks there, the state-of-the-art claim is benchmark-specific.
  • Inference: the same LLM-prompt-to-TTS recipe could generate dysfluency data for other languages or for non-stutter disorders such as apraxia and aphasia, provided a TTS with sufficient phonetic control exists.
  • Inference: the duration-matrix manipulation for pauses and prolongations suggests TTS acoustic models can be instrumented to emit event-level timing supervision, a technique that might transfer to other timestamp-annotated audio tasks.
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

4 major / 4 minor

Summary. The manuscript introduces LLM-Dys, a synthetic dysfluent speech corpus generated by prompting a large language model (Claude 3.5 Sonnet) to produce dysfluent text with token-level labels, then synthesizing speech with VITS and E2-TTS. Pauses and prolongations are inserted post-hoc at specified durations. The corpus spans 11 dysfluency types at word and phoneme levels and amounts to over 12,000 hours across 109 speakers. The authors train a Whisper-based transcription model on selected subsets and report results on (i) held-out synthetic data, (ii) a self-annotated subset of SEP-28k (200 samples per level), and (iii) a self-defined binary split of UCLASS. They claim state-of-the-art performance on real-world dysfluency detection and discuss the effect of training set size and fluent-to-disfluent ratio.

Significance. If validated, LLM-Dys would be a valuable contribution: it is one of the first efforts to use LLM-generated text diversity for dysfluency simulation, and the open-sourced dataset and code are commendable. The idea of using LLMs to generate more naturalistic dysfluent text is a plausible step beyond rule-based text insertion. However, the central empirical claim—that the resulting model achieves state-of-the-art performance on real stuttered speech—is not currently supported by the evaluation protocol: the SEP-28k test set is re-annotated in the authors' own format, baselines are cited from other papers without rerunning on the same test set, and no error bars are provided. The near-perfect synthetic test scores are expected given shared generation pipelines. The dataset itself may still be useful to the community, but the paper's headline claims require substantially stronger validation.

major comments (4)
  1. [Section 4.1 and Table 3] The SEP-28k evaluation is not conducted on the official SEP-28k labels. The authors manually annotate 200 samples per level 'following our model's output format with dysfluency tokens added to clean text,' and in Section 4.3.2 they state that block-level scores from [11] are 'applied to both word-level and phoneme-level metrics.' Consequently, the numbers in Tables 3-4 are not directly comparable to Wagner et al. [11] or Yolo-Stutter [14], which were evaluated on official SEP-28k labels. The claimed state-of-the-art performance on real stuttered speech is therefore not established. To support the claim, the authors should evaluate on the standard SEP-28k benchmark (or at least the same test clips and label schema as prior work), or rerun the baselines on their new test set, and report inter-annotator agreement for the new annotations.
  2. [Section 4.3.1 and Table 2] The near-perfect results on LLM-Dys are expected because the test set is drawn from the same LLM prompt/TTS pipeline used for training; the authors' own explanation ('consistent patterns in LLM-generated dysfluencies and our standardized TTS pipeline') confirms that this evaluation cannot measure generalization to real dysfluent speech. These results should be presented as a sanity check, not as evidence of system capability.
  3. [Section 4.1 and Table 5] The UCLASS evaluation uses a self-defined 200-sample split (80 fine-tuning, 120 testing) with binary labels, and the comparison to StutterNet is made against previously published results. It is unclear whether the splits and label protocols are matched; no variance or confidence intervals are reported anywhere. The authors should report results over multiple random splits or runs and, if possible, use the same evaluation protocol as the baseline.
  4. [Section 2.4 and Fig. 2] The claim that LLM-Dys has 'superior synthesis quality' and is 'comparable to real fluent speech' rests on the Meta Audiobox Aesthetics model, a generic audio quality predictor. There is no evidence that this model is sensitive to the naturalness of dysfluency patterns specifically, and it may simply reward pleasant prosody. A human perceptual study with dysfluent speakers or clinicians would be needed to support the naturalness claim.
minor comments (4)
  1. [Section 2.2] The duration ranges for pauses (0.8-3.5s word-level, 0.3-1.5s phoneme-level) and prolongations (0.17-0.8s) are stated without justification; please provide at least a brief explanation of how they were chosen, or a sensitivity analysis.
  2. [Section 4.3.3] 'Scaling Law' is an overstatement for a training-set-size ablation; consider renaming to 'Effect of Training Data Size'.
  3. [Table 1] The reported hours are computed by multiplying per-speaker durations by 109 speakers, which inflates the apparent size of the corpus. The number of unique LLM-generated utterances per type is the better measure of text diversity and should be highlighted.
  4. [Figures 2, 3, and 5] These figures are difficult to read in the current version; consider enlarging fonts, expanding abbreviations, and ensuring all axis labels are legible.

Circularity Check

1 steps flagged · score 4.0 of 10

SEP-28k evaluation is self-referential: the test set is annotated in the model's own output format and baseline scores are borrowed from a different label scheme, so the headline SOTA claim is not independently demonstrated; the rest of the pipeline is not circular.

  1. self definitional [Section 4.1 (Datasets, SEP-28k) and Section 4.3.2 (Evaluation on SEP-28k and UCLASS), Tables 3-4]
    "Due to poor segmentation quality, we created a test set by manually annotating 200 samples each for word/phoneme-level evaluation... Annotations follow our model's output format with dysfluency tokens added to clean text. ... Since the original SEP-28k annotations only contain block labels (without distinguishing between word-level and phoneme-level), we apply the block-level scores reported in [11] to both word-level and phoneme-level metrics."

    The abstract and conclusion claim 'state-of-the-art performance' on real stuttered speech. That claim is grounded in Tables 3-4, but the SEP-28k ground truth was created by the authors 'following our model's output format', i.e., the test labels are defined in the predictor's own output schema. Meanwhile the baseline numbers from Wagner et al. [11] and Yolo-Stutter [14] are not recomputed on this author-annotated test set; the same block-level score is even copied across both word-level and phoneme-level columns. Thus the SOTA comparison is not between systems on a common, independently labeled benchmark, but between the model on a self-defined test set and published numbers from another label scheme.

full rationale

No equation-level circularity or fitted-parameter-renamed-as-prediction was found. The synthetic corpus construction (LLM-generated dysfluent text, TTS synthesis, pause/prolongation post-processing) is a data-generation pipeline whose quality is assessed with an external automated scorer (Meta Audiobox Aesthetics), which is not circular. The UCLASS evaluation uses a real-stutter binary split and compares against an external baseline, giving some independent content. The main circularity concern is the SEP-28k evaluation supporting the headline SOTA claim: the test set is annotated by the authors in the model's own output format and baselines are borrowed from the official SEP-28k label scheme without being run on the same test set, so the comparison is self-referential by construction. This is a benchmark-validity problem rather than a derivation that reduces to its inputs, which is why the score is moderate rather than severe.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central derivation rests on several unverified modeling assumptions: the LLM's dysfluency patterns are realistic, the TTS preserves them, the automated aesthetic score measures naturalness, and the self-annotated SEP-28k test set is a valid benchmark. Only the last is used to support the SOTA claim, and it is not externally validated.

free parameters (6)
  • word-level pause duration range = 0.8-3.5s
    Chosen by hand for word-level pause insertion (Section 2.2).
  • phoneme-level pause duration range = 0.3-1.5s
    Chosen by hand for phoneme-level pause insertion (Section 2.2).
  • prolongation duration = 0.17-0.8s
    Chosen by hand to extend phoneme duration during VITS synthesis (Section 2.2).
  • fluent-to-disfluent ratio = 0.05
    Tuned on SEP-28k-like distribution to optimize F1 (Section 4.3.4).
  • phoneme:word pause supplement ratio = 3:7
    Chosen to balance phoneme-level pause samples (Section 3).
  • training set size per type = 3x4000 or 3x12000
    Selected based on scaling experiment (Section 4.3.3).
assumptions (5)
  • domain assumption Claude-3.5-Sonnet can generate realistic dysfluent text with accurate labels.
    The entire corpus depends on the LLM's understanding of dysfluency patterns (Section 2.1).
  • domain assumption VITS preserves inserted dysfluencies during synthesis and supports IPA input for phoneme-level simulation.
    This is asserted from the authors' experiments (Section 2.2).
  • domain assumption Meta Audiobox Aesthetics scores are a valid proxy for dysfluent speech naturalness and quality.
    Used for dataset quality assessment (Section 2.4).
  • ad hoc to paper The manually annotated SEP-28k test set, labeled in the model's output format, is a valid benchmark.
    No external validation of these annotations is provided (Section 4.1).
  • domain assumption The 11-category taxonomy covers the major dysfluency types of interest.
    The paper selects these categories based on prior work and clinical need (Section 1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Analysis and Evaluation of Synthetic Data Generation in Speech Dysfluency Detection." pith.science (2026). https://pith.science/paper/RQXFBFHS

@misc{pith2026250522029,
  author       = {Pith},
  title        = {Pith review of: Analysis and Evaluation of Synthetic Data Generation in Speech Dysfluency Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RQXFBFHS}},
  note         = {Machine review of arXiv:2505.22029}
}
read the original abstract

Speech dysfluency detection is crucial for clinical diagnosis and language assessment, but existing methods are limited by the scarcity of high-quality annotated data. Although recent advances in TTS model have enabled synthetic dysfluency generation, existing synthetic datasets suffer from unnatural prosody and limited contextual diversity. To address these limitations, we propose LLM-Dys -- the most comprehensive dysfluent speech corpus with LLM-enhanced dysfluency simulation. This dataset captures 11 dysfluency categories spanning both word and phoneme levels. Building upon this resource, we improve an end-to-end dysfluency detection framework. Experimental validation demonstrates state-of-the-art performance. All data, models, and code are open-sourced at https://github.com/Berkeley-Speech-Group/LLM-Dys.

Figures

Figures reproduced from arXiv: 2505.22029 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Comparision between differenct datasets, CE:Content Enjoyment, CU:Content Usefulness, PQ: Production Quality Rep Del Pau Ins Sub 0 10 (word-level) CE Degradation (%) Rep Del Pau Ins Sub 0 5 (word-level) CU Degradation (%) Rep Del Pau Ins Sub 0.0 2.5 (word-level) PQ Degradation (%) Rep Del Pro Sub Pau Ins 0 10 (phoneme-level) Rep Del Pro Sub Pau Ins 0 10 (phoneme-level) Rep Del Pro Sub Pau Ins 0.0 2.5 (phoneme-level)… view at source ↗
Figure 3
Figure 3. Comparative quality degradation analysis between LLM-Dys and VCTK-token n. 65.64% ( ) aux. 35.31% ( ) conj. 26. %42 ( ) pron. 16. 58% ( ) pron. 28.85% ( ) filler words 52. 72% ( ) other conj. prep. 17.57% ( ) other other prep. pron. adv. v. v. 27. 15 % ( ) adj. adv. other v. Substitution Deletion Repetition Insertion [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: POS analysis on four word level subsets of LLM-Dys mirrors natural speech patterns where speakers often omit these elements for efficiency; (3) Repetitions: LLMs frequently du￾plicate pronouns and prepositions, reflecting natural hesitation points; (4) Insertions: prim…
Figure 5
Figure 5. Figure 5: Left: Impact of dataset size on dysfluency detection performance. Right: Impact of Fluent-to-Disfluent speech ratio on model performance (P = Phoneme-level, W = Word-level, Libri = LibriTTS, Acc = Accuracy, Pre = Precision, Rec = Re￾call, wF1 = weighted F1 score comput…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Seamless Dysfluent Speech Text Alignment for Disordered Speech Analysis

    eess.AS 2025-06 conditional novelty 5.0 of 10

    Neural LCS uses learned phoneme and word similarity instead of exact matches to align dysfluent speech to intended text, and it outperforms DTW and Hard LCS on simulated benchmarks.

  2. Revisiting Rule-Based Stuttering Detection: A Comprehensive Analysis of Interpretable Models for Clinical Applications

    cs.AI 2025-08 reject novelty 4.0 of 10

    A hand-coded rule system with rate-normalized thresholds is reported to reach F1 0.86 on UCLASS for stuttering detection, but the supporting evaluation is largely unreproducible.

Reference graph

Works this paper leans on

46 extracted references · 23 canonical work pages · cited by 2 Pith papers

  1. [11]

    Self- supervised speech models for word-level stuttered speech de- tection,

    Y .-J. Shih, Z. Gkalitsiou, A. G. Dimakis, and D. Harwath, “Self- supervised speech models for word-level stuttered speech de- tection,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 937–944

    work page 2024

  2. [14]

    VCTK-TTS demonstrates significantly bet- ter intelligibility and naturalness than all previous datasets

    proposed VCTK-TTS, where dysfluencies are simulated only at the text level and then synthesized into speech using a TTS model [18]. VCTK-TTS demonstrates significantly bet- ter intelligibility and naturalness than all previous datasets. A subsequent study [19] introduced VCTK-Pro by incorporating co-dysfluencies. Libri-Dys [20] adopted the same technology...

    work page

  3. [1]

    For a long time, dysfluency or stutter detection has been treated simply as a classification problem, mostly bi- nary [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11], among others

    Introduction Speech dysfluency detection is an essential step for assisting in disordered speech diagnosis, language screening, or early prevention. For a long time, dysfluency or stutter detection has been treated simply as a classification problem, mostly bi- nary [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11], among others. However, to better serve clinical needs...

    work page

  4. [2]

    phoneme_before_<pause>

    Data Simulation 2.1. Dysfluent Text Generation We build upon the most recent and naturalistic simulated corpus VCTK-Token [23], leveraging Large Language Models instead of rule-based methods to generate authentic dysfluent texts. Through prompting, we obtain both dysfluent texts and corre- sponding labels, eliminating manual annotation needs (The la- bel ...

    arXiv 2025

  5. [3]

    We divide dysfluency detection into word and phoneme levels

    Token-based Dysfluency Detection We follow [23] to treat dysfluency detection as a token-based recognition problem and adopt Whisper-large-v3-turbo[32] as our base model. We divide dysfluency detection into word and phoneme levels. Based on the annotated dysfluency types in SEP-28K dataset, at word level, we train for insertion, pause, and repetition usin...

    work page

  6. [4]

    Datasets

    Experiments 4.1. Datasets

    work page

  7. [5]

    Conclusion and Future Work We introduce LLM-Dys, a large-scale dysfluency dataset span- ning 11 categories and 12,790 hours. Our method generates higher-quality synthetic speech than rule-based baselines while preserving dysfluency authenticity, and achieves state-of-the-art performance on real-world dysfluency detection. Experiments reveal optimal datase...

    work page

  8. [6]

    LLM-Dys: Our synthetic dataset contains 11 dysfluency types, totaling 12,790 hours. Details in Section 2.2) SEP- 28k[16]: Real-world dataset with 28,000 clips, labeled with blocks, prolongations, sound/word repetitions, and interjec- Table 2:Metrics on LLM-Dys Model Metrics Word Level Phoneme Level Ins Rep Pau Pau Rep Pro Ours (3*4000 samples) Recall 0.99...

    work page

Show all 46 references
  1. [7]

    Recall: Ratio of correctly identified to total actual disfluen- cies.2) Precision: Ratio of correctly identified to total pre- dicted disfluencies.3) F1-score: Harmonic mean of precision and recall.4) Accuracy (Acc): Model’s performance in identi- fying fluent speech and dysfl...

  2. [8]

    Sequence labeling to detect stuttering events in read speech,

    S. Alharbi, M. Hasan, A. J. Simons, S. Brumfitt, and P. Green, “Sequence labeling to detect stuttering events in read speech,” Computer Speech & Language, vol. 62, p. 101052, 2020

    2020

  3. [9]

    Acknowledgements Thanks for support from UC Noyce Initiative, Society of Hell- man Fellows, NIH/NIDCD, and the Schwab Innovation fund

  4. [10]

    Stutternet: stuttering disfluencies detection in syn- thetic speech signals via mel frequency cepstral coefficients fea- tures using deep learning,

    M. Abubakar, M. Mujahid, K. Kanwal, S. Iqbal, N. Asghar, and A. Alaulamie, “Stutternet: stuttering disfluencies detection in syn- thetic speech signals via mel frequency cepstral coefficients fea- tures using deep learning,”IEEE Access, 2024

    2024

  5. [12]

    Systematic review of ma- chine learning approaches for detecting developmental stutter- ing,

    L. Barrett, J. Hu, and P. Howell, “Systematic review of ma- chine learning approaches for detecting developmental stutter- ing,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 1160–1172, 2022

    2022

  6. [13]

    Dysfluency classification in stut- tered speech using deep learning for real-time applications,

    M. Jouaiti and K. Dautenhahn, “Dysfluency classification in stut- tered speech using deep learning for real-time applications,” in ICASSP. IEEE, 2022, pp. 6482–6486

    2022

  7. [15]

    Detect- ing dysfluencies in stuttering therapy using wav2vec 2.0,

    S. P. Bayerl, D. Wagner, E. N ¨oth, and K. Riedhammer, “Detect- ing dysfluencies in stuttering therapy using wav2vec 2.0,”Inter- speech, 2022

    2022

  8. [16]

    Disfluency detec- tion using a bidirectional lstm,

    V . Zayats, M. Ostendorf, and H. Hajishirzi, “Disfluency detec- tion using a bidirectional lstm,”arXiv preprint arXiv:1604.03209, 2016

    2016 arXiv

  9. [17]

    Automatic recognition of children’s read speech for stuttering application,

    S. Alharbi, A. J. Simons, S. Brumfitt, and P. D. Green, “Automatic recognition of children’s read speech for stuttering application,” in 6th. Workshop on Child Computer Interaction, 2017, pp. 1–6

    2017

  10. [18]

    Frame-level stut- ter detection,

    J. Harvill, M. Hasegawa-Johnson, and C. Yoo, “Frame-level stut- ter detection,” inInterspeech, 2022

    2022

  11. [19]

    Enhancing asr for stuttered speech with limited data using detect and pass,

    O. Shonibare, X. Tong, and V . Ravichandran, “Enhancing asr for stuttered speech with limited data using detect and pass,”arXiv preprint arXiv:2202.05396, 2022

    2022 arXiv

  12. [20]

    Large language models for dysfluency detection in stuttered speech,

    D. Wagner, S. P. Bayerl, I. Baumann, K. Riedhammer, E. N ¨oth, and T. Bocklet, “Large language models for dysfluency detection in stuttered speech,”Interspeech, 2024

    2024

  13. [21]

    Unconstrained dysfluency modeling for dysfluent speech transcription and de- tection,

    J. Lian, C. Feng, N. Farooqi, S. Li, A. Kashyap, C. J. Cho, P. Wu, R. Netzorg, T. Li, and G. K. Anumanchipalli, “Unconstrained dysfluency modeling for dysfluent speech transcription and de- tection,” in2023 IEEE Automatic Speech Recognition and Under- standing Workshop (ASRU),...

    2023

  14. [22]

    Towards hierarchical spoken lan- guage disfluency modeling,

    J. Lian and G. Anumanchipalli, “Towards hierarchical spoken lan- guage disfluency modeling,” inProceedings of the 18th Confer- ence of the European Chapter of the Association for Computa- tional Linguistics, Mar. 2024, pp. 539–551

    2024

  15. [23]

    Yolo-stutter: End-to-end region-wise speech dys- fluency detection,

    X. Zhou, A. Kashyap, S. Li, A. Sharma, B. Morin, D. Baquirin, J. V onk, Z. Ezzes, Z. Miller, M. Tempini, J. Lian, and G. Anu- manchipalli, “Yolo-stutter: End-to-end region-wise speech dys- fluency detection,” inInterspeech 2024, 2024, pp. 937–941

    2024

  16. [24]

    The university college london archive of stuttered speech (uclass),

    P. Howell, S. Davis, and J. Bartrip, “The university college london archive of stuttered speech (uclass),” 2009

    2009

  17. [25]

    Sep- 28k: A dataset for stuttering event detection from podcasts with people who stutter,

    C. Lea, V . Mitra, A. Joshi, S. Kajarekar, and J. P. Bigham, “Sep- 28k: A dataset for stuttering event detection from podcasts with people who stutter,” inICASSP, 2021

    2021

  18. [26]

    Fluentnet: End-to- end detection of stuttered speech disfluencies with deep learning,

    T. Kourkounakis, A. Hajavi, and A. Etemad, “Fluentnet: End-to- end detection of stuttered speech disfluencies with deep learning,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 29, pp. 2986–2999, 2021

    2021

  19. [27]

    Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,

    J. Kim, J. Kong, and J. Son, “Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,” inInter- national Conference on Machine Learning, 2021

    2021

  20. [28]

    Stutter-solver: End-to-end multi-lingual dysfluency detection,

    X. Zhou, C. J. Cho, A. Sharma, B. Morin, D. Baquirin, J. V onk, Z. Ezzes, Z. Miller, B. L. Tee, M. L. Gorno-Tempiniet al., “Stutter-solver: End-to-end multi-lingual dysfluency detection,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 1039–1046

    2024

  21. [29]

    Ssdm: Scalable speech dysfluency modeling,

    J. Lian, X. Zhou, Z. Ezzes, J. V onk, B. Morin, D. P. Baquirin, Z. Miller, M. L. Gorno Tempini, and G. Anumanchipalli, “Ssdm: Scalable speech dysfluency modeling,” inAdvances in Neural In- formation Processing Systems, vol. 37, 2024

    2024

  22. [30]

    Libritts: A corpus derived from librispeech for text- to-speech,

    H. Zen, V . Dang, R. Clark, Y . Zhang, R. J. Weiss, Y . Jia, Z. Chen, and Y . Wu, “Libritts: A corpus derived from librispeech for text- to-speech,”arXiv preprint arXiv:1904.02882, 2019

    1904 arXiv

  23. [31]

    Ssdm 2.0: Time-accurate speech rich transcription with non-fluencies,

    J. Lian, X. Zhou, Z. Ezzes, J. V onk, B. Morin, D. Baquirin, Z. Mille, M. L. G. Tempini, and G. K. Anumanchipalli, “Ssdm 2.0: Time-accurate speech rich transcription with non-fluencies,” arXiv preprint arXiv:2412.00265, 2024

    2024 arXiv

  24. [32]

    Time and tokens: Benchmarking end- to-end speech dysfluency detection,

    X. Zhou, J. Lian, C. J. Cho, J. Liu, Z. Ye, J. Zhang, B. Morin, D. Baquirin, J. V onk, Z. Ezzes, Z. Miller, M. L. G. Tempini, and G. Anumanchipalli, “Time and tokens: Benchmarking end- to-end speech dysfluency detection,” 2024. [Online]. Available: https://arxiv.org/abs/2409.13582

    2024 arXiv

  25. [33]

    Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit (ver- sion 0.92),

    J. Yamagishi, C. Veaux, K. MacDonaldet al., “Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit (ver- sion 0.92),”University of Edinburgh. The Centre for Speech Tech- nology Research (CSTR), pp. 271–350, 2019

    2019

  26. [34]

    Crisperwhisper: Accu- rate timestamps on verbatim speech transcriptions,

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

    2024

  27. [35]

    Meta audiobox aes- thetics: Unified automatic quality assessment for speech, music, and sound,

    A. Tjandra, Y .-C. Wu, B. Guo, J. Hoffman, B. Ellis, A. Vyas, B. Shi, S. Chen, M. Le, N. Zacharovet al., “Meta audiobox aes- thetics: Unified automatic quality assessment for speech, music, and sound,”arXiv preprint arXiv:2502.05139, 2025

    2025 arXiv

  28. [36]

    Scaling laws for neural language models,

    J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei, “Scaling laws for neural language models,”arXiv preprint arXiv:2001.08361, 2020

    2001 arXiv

  29. [37]

    Introducing the next generation of claude,

    Anthropic, “Introducing the next generation of claude,”

  30. [39]

    Phonemizer: Text to phones transcription for multiple languages in python,

    M. Bernard and H. Titeux, “Phonemizer: Text to phones transcription for multiple languages in python,”Journal of Open Source Software, vol. 6, no. 68, p. 3958, 2021. [Online]. Available: https://doi.org/10.21105/joss.03958

    2021 doi

  31. [40]

    F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,

    Y . Chen, Z. Niu, Z. Ma, K. Deng, C. Wang, J. Zhao, K. Yu, and X. Chen, “F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,”arXiv preprint arXiv:2410.06885, 2024

    2024 arXiv

  32. [41]

    E2 tts: Embarrassingly easy fully non-autoregressive zero-shot tts,

    S. E. Eskimez, X. Wang, M. Thakker, C. Li, C.-H. Tsai, Z. Xiao, H. Yang, Z. Zhu, M. Tang, X. Tan, Y . Liu, S. Zhao, and N. Kanda, “E2 tts: Embarrassingly easy fully non-autoregressive zero-shot tts,” in2024 IEEE Spoken Language Technology Workshop (SLT), 2024, pp. 682–689

    2024

  33. [42]

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

    2023

  34. [43]

    Coding speech through vocal tract kinematics,

    C. J. Cho, P. Wu, T. S. Prabhune, D. Agarwal, and G. K. Anu- manchipalli, “Coding speech through vocal tract kinematics,” in IEEE JSTSP, 2025

    2025

  35. [44]

    Deep Speech Synthesis from MRI-Based Articulatory Representations,

    P. Wu, T. Li, Y . Lu, Y . Zhang, J. Lian, A. W. Black, L. Goldstein, S. Watanabe, and G. K. Anumanchipalli, “Deep Speech Synthesis from MRI-Based Articulatory Representations,” inProc. INTER- SPEECH 2023, 2023, pp. 5132–5136

    2023

  36. [45]

    Deep Neural Convolutive Matrix Factorization for Articulatory Representation Decomposition,

    J. Lian, A. W. Black, L. Goldstein, and G. K. Anumanchipalli, “Deep Neural Convolutive Matrix Factorization for Articulatory Representation Decomposition,” inProc. Interspeech 2022, 2022, pp. 4686–4690

    2022

  37. [46]

    Articulatory representation learning via joint factor analysis and neural matrix factorization,

    J. Lian, A. W. Black, Y . Lu, L. Goldstein, S. Watanabe, and G. K. Anumanchipalli, “Articulatory representation learning via joint factor analysis and neural matrix factorization,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICA...

    2023

  38. [2024]

    Available: https://www.anthropic.com/news/ claude-3-family

    [Online]. Available: https://www.anthropic.com/news/ claude-3-family

Pith tools

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