Pith. sign in

REVIEW 5 major objections 6 minor 38 references

The paper claims that a fully interpretable rule-based detector—using rate-normalized thresholds and a hierarchical cascade—reaches competitive stuttering-detection performance (0.86 F1 on UCLASS), within about six points of neural models w

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

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.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A coherent rule-based framework with a useful mechanism test, but the headline evaluation is unauditable: undefined UCLASS splits, unspecified baselines, and thresholds tuned on the same test corpora. the 5 major comments →

arxiv 2508.16681 v1 pith:XZHKB6C2 submitted 2025-08-21 cs.AI cs.CL

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

Eric Zhang This is my paper
classification cs.AI cs.CL
keywords stuttering detectionrule-based detectioninterpretabilityspeaking-rate normalizationprolongation detectiondysfluencyacoustic featuresclinical speech assessment
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that a deliberately simple, hand-coded detector—using explicit acoustic thresholds rather than learned parameters—can perform nearly as well as modern neural models on stuttering-event detection while keeping every decision auditable. Across three corpora it reports F1 scores of 0.86 on UCLASS, 0.83 on FluencyBank, and 0.74 on SEP-28k, about six points below the strongest neural baselines, and 97–99% accuracy on prolongations. The central mechanism is speaking-rate normalization: duration thresholds scale with estimated syllable rate, so detections stay stable when speech is artificially slowed or sped up from 0.5× to 2.0×. The paper also reports a pilot with three speech-language pathologists who preferred the rule-based system's transparency and adjusted its thresholds per patient. If the results hold, interpretable detection is a practical clinical tool, not a compromise.

Core claim

On its own terms, the paper's discovery is that a fully interpretable rule cascade closes most of the gap to neural dysfluency detectors. The headline comparison, Table 1, shows the proposed rule-based system reaching 0.86 F1 on UCLASS, 0.83 on FluencyBank, and 0.74 on SEP-28k, against 0.91, 0.88, and 0.83 for YOLO-Stutter, the strongest neural system; the average gap is about 6 F1 points. Per dysfluency type on UCLASS, prolongations are detected at 0.98 F1, matching or exceeding neural systems, while blocks remain the weak spot (0.69 versus 0.79). The paper credits rate-normalized thresholds and a hierarchical decision cascade for the gain over earlier rules, and clinical value to complete

What carries the argument

The load-bearing mechanism is the rate-normalized prolongation detector (Algorithm 1/2). It sets a minimum segment duration T_min = 1.2 / SR, where SR is the estimated syllable rate, and flags a prolongation when consecutive MFCC frames correlate above 0.92 for longer than T_min, with additional gates on F0 variation (<15 Hz) and harmonic-to-noise ratio (>10 dB). Because the duration threshold shrinks or grows with speaking rate, the same rule remains stable across 0.5×–2.0× rate changes. Around this core sits a cascade: DTW-based sound-repetition detection, forced-alignment word-repetition detection, and block detection, with a fixed precedence order for overlapping events. The cascade plus

Load-bearing premise

Every reported number depends on the evaluation being real, comparable, and correctly instantiated: the paper says it used official train/test splits for UCLASS but does not name the split, and it does not document whether the neural baseline scores were produced under the same evaluation protocol; if this comparability fails, the 0.86 F1 and the six-point gap are not established.

What would settle it

Run the released rule code on UCLASS with a documented split and re-evaluate the neural baselines on the exact same event-level metric; if the rule system's F1 falls below 0.86 on UCLASS, or the gap to YOLO-Stutter exceeds the reported margin, the headline comparison is falsified. Checking the UCLASS archive for the claimed official splits would settle reproducibility immediately.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the results hold, clinics can deploy a real-time CPU-only stutter detector (0.02× real-time, under 50MB memory) that logs the acoustic evidence behind every flagged event, supporting progress tracking and audit.
  • Near-ceiling prolongation performance makes the rule module a natural proposal generator: it can pre-mark candidate regions for a neural model, pairing neural recall with rule explanations.
  • Rate normalization keeps the detector usable in therapy settings where patients deliberately slow or speed their speech, where fixed-threshold rules collapse.
  • The roughly six-point F1 gap becomes the measurable price of full interpretability; hybrid pipelines that attach rule constraints to neural outputs have a concrete target for closing it.
  • The reported 0.71 F1 zero-shot Mandarin score turns cross-linguistic transfer into a testable claim: acoustic dysfluency cues may generalize without retraining.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the results replicate, the biggest barrier to clinical use is evaluation transparency: the paper never specifies which UCLASS split it used or whether neural baseline scores were measured under the same protocol, so the exact size of the gap remains open.
  • A testable next step is patient-specific threshold adaptation: tune alpha and the similarity threshold from a few clinician-labeled utterances, within acoustically meaningful ranges, and measure whether F1 improves without losing explainability.
  • The rule cascade could serve as a cheap falsifier for synthetic stuttered speech: generated samples that fail to trigger the interpretable rules are unlikely to contain the target acoustic dysfluency, giving synthetic-data pipelines a validation filter.
  • Running the English-tuned rules on additional languages with existing stuttering corpora would directly test the zero-shot generalization claim beyond the single Mandarin result reported.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper argues that a rule-based, fully interpretable stuttering-detection system can closely match modern neural models. The proposed framework combines speaking-rate normalization (Algorithm 1: Tmin = α/SR with α = 1.2), MFCC frame-correlation gating for prolongations, DTW-based repetition detection, and a hierarchical post-processing cascade. The authors report F1 scores of 0.86 on UCLASS, 0.83 on FluencyBank, and 0.74 on SEP-28k, claim 97–99% accuracy for prolongation detection, and present a rate-robustness study showing stable performance for speaking-rate changes from 0.5× to 2.0×. They also discuss clinical validation, computational efficiency, zero-shot Mandarin transfer, and integration strategies with neural pipelines. The central claim is that the rule system is within about 6 F1 points of neural baselines while preserving complete decision auditability.

Significance. If the reported evaluation were trustworthy, this would be a useful contribution: an interpretable, lightweight, rate-normalized detector with competitive performance and transparent decision evidence would have practical value for clinical deployment, and the proposed integration ideas (rules as proposal generators or constraints) are sensible. The paper's strengths are the conceptual clarity of Algorithm 1 and the attempt at cross-corpus evaluation. However, every headline quantitative claim rests on an evaluation protocol that is not reproducible: no code is provided, the UCLASS split is not actually official or described, baseline conditions are unspecified, and the key thresholds are tuned on the same corpora whose results are reported. The significance is therefore conditional; at present the experimental foundation does not support the paper's central comparative claims.

major comments (5)
  1. [§4.1, Table 1] The UCLASS F1 of 0.86 is the paper's headline result. §4.1 states 'We use official train/test splits,' but the cited UCLASS archive (Howell et al., 2009) defines no official train/test splits. The manuscript never specifies which recordings were used for training or testing, how speaker overlap was handled, or how the split was generated. The number is therefore not reproducible or auditable. The authors must provide the exact split (or a released split file) and justify why it should be considered official.
  2. [§4.2, Table 1] Baseline comparisons are not specified enough to interpret. For H-UDM, SSDM, YOLO-Stutter, and Dysfluent-WFST, the paper does not state whether models were re-run locally, which checkpoints/configurations were used, or which evaluation protocol (event-based vs. interval-based, tolerance window, label mapping) was applied. Source papers use incompatible evaluation settings; e.g., SEP-28k and UCLASS protocols differ. Without this information, the claimed '6% gap' in Table 1 may be an artifact of protocol mismatch rather than a real performance difference.
  3. [§3.2.1, §B.2, Table 5] Thresholds are tuned on the same corpora used for the reported results. α=1.2 is described as 'empirically tuned' in §3.2.1, and the thresholds in Algorithm 2 (θ_mfcc=0.92, θ_f0, θ_hnr) come from the sensitivity analysis in §B.2. Table 5 explicitly reports F1=0.86 in the UCLASS→UCLASS diagonal after training thresholds on UCLASS. This is in-sample tuning, not held-out evaluation. The central claim that a fixed interpretable rule system is within 6% of neural models requires nested or cross-corpus threshold selection, with tuning performed only on training folds and validation used only for final assessment.
  4. [Tables 1–3, 5–8] No error bars, confidence intervals, or significance tests are reported for any quantitative result. Single-run F1 scores and the rate-robustness table (Table 3) cannot support comparative claims; the SLP agreement κ=0.74 in §4.4 is based on 3 clinicians and 50 sessions without reporting clinician-level variability or even the same evaluation units. At minimum, bootstrapped confidence intervals, per-speaker or per-clip breakdowns, and repeated or significance-tested comparisons are needed.
  5. [§1, contribution 5] Contribution 5 states 'We release our implementation with detailed parameter specifications,' but no repository, code, or precomputed detections are provided in the manuscript. Combined with the missing split and protocol details above, none of the empirical claims can currently be checked. A release of code, features, and exact evaluation scripts is essential for the paper's reproducibility claims.
minor comments (6)
  1. [Abstract, §4.3.2, Table 2] The abstract claims '97-99% accuracy' for prolongation detection, but Table 2 reports F1 scores, not accuracy. The corresponding number for 'Ours' is F1=0.98. Please use consistent metrics and avoid conflating accuracy with F1.
  2. [§5, zero-shot claim] Section 5 states the English-tuned system 'achieved 0.71 F1 on Mandarin stuttering data without modification,' but no dataset, annotation scheme, sample size, or evaluation protocol is described. This claim cannot be assessed and should be either fully specified or removed.
  3. [§B.2, Figure 1] Figure 1 appears as a placeholder ('Figure: F1 score vs. threshold values...') with no actual plot, axes, or data points. A real figure with axis labels and values is needed.
  4. [§4.4, clinical validation] The clinician trust survey ('4.2/5.0 for rule-based vs. 2.8/5.0 for neural') uses a non-validated instrument with no item list, no confidence intervals, and a very small sample. Please report the survey items, response scales, and variability, or frame this as anecdotal feasibility evidence.
  5. [§D.3, longitudinal case study] The longitudinal case study in Table 8 is a single patient, 12-session anecdote. It should not be described in terms that imply validation ('System tracking closely matched SLP assessment'); it can only illustrate potential utility.
  6. [Algorithm 3, §A.2] In Algorithm 3, the line 'i ← i + W' skips forward by the full window after a repetition is found, which prevents detection of overlapping or immediately consecutive repetition candidates. Please justify this design choice or compare against a sliding-window alternative.

Circularity Check

1 steps flagged

Headline UCLASS F1 is an in-sample threshold fit presented as prediction; the rule algorithm itself is not circular.

specific steps
  1. fitted input called prediction [Section 3.2.1 (Algorithm 1), Section B.2, Table 1]
    "Tmin ← α/SR {Adaptive threshold} ... The key innovation is speaking-rate normalization: Tmin = α/SR where α ≈ 1.2 (empirically tuned). ... We conducted extensive sensitivity analysis to determine optimal thresholds. ... Table 1: Ours (Rule) ... UCLASS 0.86"

    The headline UCLASS F1 is produced by thresholds that the paper says were empirically tuned/sensitivity-optimized. Section 4.1 claims 'official train/test splits' for UCLASS, but the cited UCLASS archive defines no official splits and the paper gives none. Table 5's caption confirms thresholds are 'trained' per corpus, and its UCLASS→UCLASS diagonal is 0.86—the same number Table 1 presents as the rule system's UCLASS performance. Thus the 0.86 F1 is obtained by fitting α and the related thresholds to UCLASS and then evaluating on UCLASS; it is a training-set score, not an independent held-out prediction. The 'within 6% of neural models' claim therefore reduces to a fitted comparison, not to a prediction the rules were able to make without seeing the evaluation corpus. Cross-corpus rows are

full rationale

The rule-based detector itself is not circular: Algorithm 1 deterministically maps acoustic features to prolongation segments, and the hierarchy in Section 3.2 is a definitional decision procedure, not a disguised reuse of the evaluation labels. The circularity is confined to the empirical claim. α is described as 'empirically tuned' and Section B.2 reports a sensitivity analysis to choose thresholds; Table 5 explicitly says thresholds were 'trained on one corpus and evaluated on others,' and the UCLASS→UCLASS diagonal equals the 0.86 headline. Combined with the false statement that UCLASS has official train/test splits, the reported UCLASS F1 cannot be distinguished from in-sample fitting. This is the fitted-input-called-prediction pattern. Other concerns—unreported neural baseline protocols, missing code, undefined θR in Eq. 1—are correctness/reproducibility issues rather than circularity, so they do not raise the score beyond 6. The framework's cross-corpus generalization rows (e.g., training on UCLASS, testing on FluencyBank) provide some independent content, which is why the score is not 8 or 10.

Axiom & Free-Parameter Ledger

9 free parameters · 6 axioms · 0 invented entities

The paper's predictive content reduces to roughly five hand-tuned thresholds plus a speaking-rate scaling constant, all fitted to the same corpora whose performance is then reported: alpha = 1.2 ('empirically tuned', Algorithm 1), theta_sim = 0.92, theta_f0 = 15 Hz, theta_hnr = 10 dB, theta_dtw = 0.3, plus a 350 ms block-silence cutoff, a 100 ms separation rule, and a 30-frame DTW window. Table 5 makes the fitting explicit by 'training' thresholds per corpus. No new entities are postulated. Background assumptions include the validity of MFCC spectral correlation as a prolongation marker (inherited from [18,19]), reliable forced alignment on stuttered speech (in tension with the paper's own intro, which notes ASR degrades on disfluent speech), WSOLA preserving annotation timing, and the existence of both official UCLASS splits and an unnamed Mandarin corpus.

free parameters (9)
  • alpha (rate-normalization constant) = ~1.2
    Sets prolongation duration threshold Tmin = alpha / SR; stated as 'empirically tuned' in Section 3.2.1, Algorithm 1.
  • theta_sim (MFCC frame-correlation threshold) = 0.92
    Prolongation gate in Algorithms 1 and 2; chosen via sensitivity analysis (Appendix B.2).
  • theta_f0 (F0 stability threshold) = 15 Hz
    Additional prolongation gate in Algorithm 2 (Appendix A.1); tuned on the same corpora.
  • theta_hnr (harmonic-to-noise ratio threshold) = 10 dB
    Voicing gate in Algorithm 2 (Appendix A.1); tuned.
  • theta_dtw (DTW distance threshold) = 0.3
    Sound-repetition detection cutoff in Algorithm 3 (Appendix A.2); tuned.
  • silent-block duration cutoff = >350 ms
    Block detection rule in Section 3.2.3; hand-chosen.
  • minimum same-type event separation = 100 ms
    Post-processing constraint in Section 3.3; hand-chosen.
  • DTW window size W = 30 frames
    Repetition window in Algorithm 3 (Appendix A.2); chosen.
  • theta_R (autocorrelation repetition threshold, Eq. 1) = unspecified
    Equation 1 in Section 3.2.2 claims to detect repetitions via weighted autocorrelation but gives no value for theta_R, leaving the detection rule incomplete.
axioms (6)
  • domain assumption MFCC frame-to-frame spectral correlation is a valid prolongation marker
    Inherited from Howell and Sackin [18] and Esmaili et al. [19]; used as the primary gate in Section 3.2.1.
  • domain assumption Montreal Forced Aligner produces reliable alignments on stuttered speech
    Word-repetition detection depends on MFA alignment (Section 3.2.2); the paper's own introduction cites evidence [10,11] that ASR degrades on disfluent speech.
  • domain assumption WSOLA time-stretching preserves dysfluency timings and labels
    Speaking-rate robustness experiment (Section 4.3.3) stretches audio by 0.5x to 2.0x and assumes annotations remain valid.
  • ad hoc to paper UCLASS provides official train/test splits
    Section 4.1 asserts 'We use official train/test splits'; the UCLASS archive as cited [1] defines no such splits and the paper specifies none.
  • domain assumption Syllable nuclei detection estimates speaking rate accurately
    The normalization Tmin = alpha/SR (Sections 3.1.2, 3.2.1) requires SR; no algorithm or error analysis for SR estimation is given.
  • ad hoc to paper Mandarin stuttering data for the zero-shot claim exists and is labeled comparably
    Section 5 reports 0.71 F1 on 'Mandarin stuttering data' without naming the corpus, split, or annotation scheme.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Revisiting Rule-Based Stuttering Detection: A Comprehensive Analysis of Interpretable Models for Clinical Applications." pith.science (2026). https://pith.science/paper/XZHKB6C2

@misc{pith2026250816681,
  author       = {Pith},
  title        = {Pith review of: Revisiting Rule-Based Stuttering Detection: A Comprehensive Analysis of Interpretable Models for Clinical Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XZHKB6C2}},
  note         = {Machine review of arXiv:2508.16681}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Stuttering affects approximately 1% of the global population, impacting communication and quality of life. While recent advances in deep learning have pushed the boundaries of automatic speech dysfluency detection, rule-based approaches remain crucial for clinical applications where interpretability and transparency are paramount. This paper presents a comprehensive analysis of rule-based stuttering detection systems, synthesizing insights from multiple corpora including UCLASS, FluencyBank, and SEP-28k. We propose an enhanced rule-based framework that incorporates speaking-rate normalization, multi-level acoustic feature analysis, and hierarchical decision structures. Our approach achieves competitive performance while maintaining complete interpretability-critical for clinical adoption. We demonstrate that rule-based systems excel particularly in prolongation detection (97-99% accuracy) and provide stable performance across varying speaking rates. Furthermore, we show how these interpretable models can be integrated with modern machine learning pipelines as proposal generators or constraint modules, bridging the gap between traditional speech pathology practices and contemporary AI systems. Our analysis reveals that while neural approaches may achieve marginally higher accuracy in unconstrained settings, rule-based methods offer unique advantages in clinical contexts where decision auditability, patient-specific tuning, and real-time feedback are essential.

Figures

Figures reproduced from arXiv: 2508.16681 by Eric Zhang.

Figure 1
Figure 1. Figure 1: Sensitivity of detection performance to threshold parameters. Optimal values shown with [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 32 canonical work pages · 3 internal anchors

  1. [1]

    The uclass archive of stuttered speech

    Peter Howell, Steve Davis, and Jon Bartrip. The uclass archive of stuttered speech. Journal of Speech, Language, and Hearing Research, 52(2):556–569, 2009

    work page 2009

  2. [2]

    A systematic review of anxiety levels in people who stutter

    Ashley Craig, Kylie Hancock, Yvonne Tran, Mark Craig, and Kimberley Peters. A systematic review of anxiety levels in people who stutter. Journal of Fluency Disorders, 34(4):203–221, 2009

    work page 2009

  3. [3]

    The relationship between mental health disorders and treatment outcomes among adults who stutter

    Lisa Iverach, Sue O’Brian, Mark Jones, Susan Block, Michelle Lincoln, Elisabeth Harrison, Sally Hewat, Ross G Menzies, Ann Packman, and Mark Onslow. The relationship between mental health disorders and treatment outcomes among adults who stutter. Journal of Fluency Disorders, 34(1):29–43, 2009

    work page 2009

  4. [4]

    Localization of stuttering based on causal brain lesions

    Catherine Theys, Juho Jaakkola, Luc F De Nil, Trevor Knittle, Priyanka Shah-Basak, Laura Battaglini, Vitoria Piai, Deryk Beal, Xuehai Kang, Sami Vitti-Diedrich, et al. Localization of stuttering based on causal brain lesions. Brain, 147(6):2203–2216, 2024

    work page 2024

  5. [5]

    Acquired stuttering in parkinson’s disease

    Emily Gooch, Tracy R Melzer, Kyla L Horne, Sophie Grenfell, Lynette Livingston, Toni Pitcher, John C Dalrymple-Alford, Tim J Anderson, Megan J McAuliffe, and Catherine Theys. Acquired stuttering in parkinson’s disease. Movement Disorders Clinical Practice, 10(7):1065–1074, 2023

    work page 2023

  6. [6]

    Neurogenic stuttering: Etiology, symptomatology, and treatment

    Catherine Theys and Luc F De Nil. Neurogenic stuttering: Etiology, symptomatology, and treatment. In Speech disorders: Causes, treatment and social effects, pages 1–36. Nova Science Publishers, 2011

    work page 2011

  7. [7]

    Neurogenic stuttering: a review of the literature

    Carla Cruz, Helena Amorim, Gonçalo Beca, et al. Neurogenic stuttering: a review of the literature. Revista de Neurología, 66(2):59–64, 2018

    work page 2018

  8. [8]

    A crucial role for the cortico-striato-cortical loop in the pathogenesis of stroke-related neurogenic stuttering

    Catherine Theys, Luc F De Nil, Vincent Thijs, Astrid van Wieringen, and Stefan Sunaert. A crucial role for the cortico-striato-cortical loop in the pathogenesis of stroke-related neurogenic stuttering. Human Brain Mapping, 34(9):2103–2112, 2013. 11

    work page 2013

  9. [9]

    A comprehensive review of stuttering therapy apps: Landscape analysis and quality assessment

    Lisa Koops and V olker Dellwo. A comprehensive review of stuttering therapy apps: Landscape analysis and quality assessment. Journal of Fluency Disorders, 71:105879, 2022

    work page 2022

  10. [10]

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

    Colin Lea, Vikramjit Mitra, Aparna Joshi, Sachin Kajarekar, and Jeffrey P Bigham. Sep-28k: A dataset for stuttering event detection from podcasts with people who stutter. arXiv preprint arXiv:2102.12394, 2021

    Pith/arXiv arXiv 2021

  11. [11]

    Machine learning for stuttering identi- fication: Review, challenges & future directions

    John Harvill, Payam Janbakhshi, and Mari Ostendorf. Machine learning for stuttering identi- fication: Review, challenges & future directions. Computer Speech & Language, 75:101343, 2022

    work page 2022

  12. [12]

    Fluency bank: a new resource for fluency research and practice

    Nan Bernstein Ratner and Brian MacWhinney. Fluency bank: a new resource for fluency research and practice. Language, Speech, and Hearing Services in Schools , 49(2):329–344, 2018

    work page 2018

  13. [13]

    Classification of stuttering–the compare challenge and beyond

    Sebastian P Bayerl, Florian Hönig, Elmar Nöth, and Korbinian Riedhammer. Classification of stuttering–the compare challenge and beyond. Computer Speech & Language, 81:101520, 2023

    work page 2023

  14. [14]

    Ssdm: Scalable speech dysfluency modeling

    Jiachen Lian, Xuanru Zhou, Zoe Ezzes, Jet V onk, Brittany Morin, David Paul Baquirin, Zachary Miller, Maria Luisa Gorno Tempini, and Gopala Anumanchipalli. Ssdm: Scalable speech dysfluency modeling. In Advances in Neural Information Processing Systems, volume 37, 2024

    work page 2024

  15. [15]

    SSDM 2.0: Time-Accurate Speech Rich Transcription with Non-Fluencies

    Jiachen Lian, Xuanru Zhou, Zoe Ezzes, Jet V onk, Brittany Morin, David Baquirin, Zachary Mille, Maria Luisa Gorno Tempini, and Gopala Krishna Anumanchipalli. Ssdm 2.0: Time- accurate speech rich transcription with non-fluencies. arXiv preprint arXiv:2412.00265, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  16. [16]

    Yolo-stutter: End-to-end region-wise speech dysfluency detection

    Xuanru Zhou, Anshul Kashyap, Steve Li, Ayati Sharma, Brittany Morin, David Baquirin, Jet V onk, Zoe Ezzes, Zachary Miller, Maria Tempini, Jiachen Lian, and Gopala Anumanchipalli. Yolo-stutter: End-to-end region-wise speech dysfluency detection. In Interspeech 2024, pages 937–941, 2024. doi: 10.21437/Interspeech.2024-1855

    work page doi:10.21437/interspeech.2024-1855 2024

  17. [17]

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

    Xuanru Zhou, Cheol Jun Cho, Ayati Sharma, Brittany Morin, David Baquirin, Jet V onk, Zoe Ezzes, Zachary Miller, Boon Lead Tee, Maria Luisa Gorno-Tempini, et al. Stutter-solver: End-to-end multi-lingual dysfluency detection. In 2024 IEEE Spoken Language Technology Workshop (SLT), pages 1039–1046. IEEE, 2024

    2024

  18. [18]

    Automatic recognition of repetitions and prolongations in stuttered speech

    Peter Howell and Steve Sackin. Automatic recognition of repetitions and prolongations in stuttered speech. Proceedings of the first World Congress on fluency disorders , 2:372–374, 1995

    work page 1995

  19. [19]

    An automatic prolongation detection approach in continuous speech with robustness against speaking rate variations

    Iman Esmaili, Nader Jafarnia Dabanloo, and Mansour Vali. An automatic prolongation detection approach in continuous speech with robustness against speaking rate variations. Journal of Medical Signals and Sensors, 7(1):1–11, 2017

    work page 2017

  20. [20]

    Seamless dysfluent speech text alignment for disordered speech analysis

    Zongli Ye, Jiachen Lian, Xuanru Zhou, Jinming Zhang, Haodong Li, Shuhe Li, Chenxu Guo, Anaisha Das, Peter Park, Zoe Ezzes, Jet V onk, Brittany Morin, Rian Bogley, Lisa Wauters, Zachary Miller, Maria Gorno-Tempini, and Gopala Anumanchipalli. Seamless dysfluent speech text alignment for disordered speech analysis. Interspeech, 2025

    2025

  21. [21]

    Lcs-ctc: Leveraging soft alignments to enhance phonetic transcription robustness

    Zongli Ye, Jiachen Lian, Akshaj Gupta, Xuanru Zhou, Krish Patel, Haodong Li, Hwi Joo Park, Chenxu Guo, Shuhe Li, Sam Wang, et al. Lcs-ctc: Leveraging soft alignments to enhance phonetic transcription robustness. arXiv preprint arXiv:2508.03937, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Universal phone recognition with a multilingual allophone system

    Xinjian Li, Siddharth Dalmia, Juncheng Li, Matthew Lee, Patrick Littell, Jiali Yao, Antonios Anastasopoulos, David R Mortensen, Graham Neubig, Alan W Black, et al. Universal phone recognition with a multilingual allophone system. In ICASSP, pages 8249–8253. IEEE, 2020

    work page 2020

  23. [23]

    Pattern search in dysfluent speech

    Juraj Pálfy and Jiˇrí Pospíchal. Pattern search in dysfluent speech. In 2012 IEEE International Workshop on Machine Learning for Signal Processing, pages 1–6, 2012. doi: 10.1109/MLSP. 2012.6349744. 12

    work page arXiv 2012

  24. [24]

    Lever- aging allophony in self-supervised speech models for atypical pronunciation assessment

    Kwanghee Choi, Eunjung Yeo, Kalvin Chang, Shinji Watanabe, and David Mortensen. Lever- aging allophony in self-supervised speech models for atypical pronunciation assessment. In NAACL, 2025

    work page 2025

  25. [25]

    Data-driven mispronunciation pattern discovery for robust speech recognition

    Anna Seo Gyeong Choi, Jonghyeon Park, and Myungwoo Oh. Data-driven mispronunciation pattern discovery for robust speech recognition. In ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , page 1–5. IEEE, April

    work page 2025

  26. [26]

    Towards hierarchical spoken language disfluency modeling

    Jiachen Lian and Gopala Anumanchipalli. Towards hierarchical spoken language disfluency modeling. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics, pages 539–551, 2024

    work page 2024

  27. [27]

    Unconstrained dysfluency modeling for dysfluent speech transcription and detection

    Jiachen Lian, Carly Feng, Naasir Farooqi, Steve Li, Anshul Kashyap, Cheol Jun Cho, Peter Wu, Robbie Netzorg, Tingle Li, and Gopala Krishna Anumanchipalli. Unconstrained dysfluency modeling for dysfluent speech transcription and detection. In 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 1–8, 2023. doi: 10.1109/ASRU57964. ...

    work page arXiv 2023

  28. [28]

    Dysfluent wfst: A framework for zero-shot speech dysfluency transcription and detection, 2025

    Chenxu Guo, Jiachen Lian, Xuanru Zhou, Jinming Zhang, Shuhe Li, Zongli Ye, Hwi Joo Park, Anaisha Das, Zoe Ezzes, Jet V onk, Brittany Morin, Rian Bogley, Lisa Wauters, Zachary Miller, Maria Gorno-Tempini, and Gopala Anumanchipalli. Dysfluent wfst: A framework for zero-shot speech dysfluency transcription and detection, 2025. URL https://arxiv.org/abs/2505. 16351

    work page 2025

  29. [29]

    Time and Tokens: Benchmarking End-to-End Speech Dysfluency Detection

    Xuanru Zhou, Jiachen Lian, Cheol Jun Cho, Jingwen Liu, Zongli Ye, Jinming Zhang, Brittany Morin, David Baquirin, Jet V onk, Zoe Ezzes, Zachary Miller, Maria Luisa Gorno Tempini, and Gopala Anumanchipalli. Time and tokens: Benchmarking end-to-end speech dysfluency detection, 2024. URL https://arxiv.org/abs/2409.13582

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  30. [30]

    Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit (version 0.92), 2019

    Junichi Yamagishi, Christophe Veaux, and Kirsten MacDonald. Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit (version 0.92), 2019. [sound], University of Edinburgh, The Centre for Speech Technology Research (CSTR)

    work page 2019

  31. [31]

    Analysis and Evaluation of Synthetic Data Generation in Speech Dysfluency Detection

    Jinming Zhang, Xuanru Zhou, Jiachen Lian, Shuhe Li, William Li, Zoe Ezzes, Rian Bogley, Lisa Wauters, Zachary Miller, Jet V onk, Brittany Morin, Maria Gorno-Tempini, and Gopala Anumanchipalli. Analysis and evaluation of synthetic data generation in speech dysfluency detection, 2025. URL https://arxiv.org/abs/2505.22029

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  32. [32]

    Montreal forced aligner: Trainable text-speech alignment using kaldi

    Michael McAuliffe, Michaela Socolof, Sarah Mihuc, Michael Wagner, and Morgan Sonderegger. Montreal forced aligner: Trainable text-speech alignment using kaldi. In Interspeech, volume 2017, pages 498–502, 2017

    2017

  33. [33]

    K-function: Joint pronunciation transcription and feedback for evaluating kids language function

    Shuhe Li, Chenxu Guo, Jiachen Lian, Cheol Jun Cho, Wenshuo Zhao, Xuanru Zhou, Dingkun Zhou, Sam Wang, Grace Wang, Jingze Yang, et al. K-function: Joint pronunciation transcription and feedback for evaluating kids language function. arXiv preprint arXiv:2507.03043, 2025

    work page arXiv 2025

  34. [34]

    Signature verification using a Siamese time delay neural network

    Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a Siamese time delay neural network. In Advances in Neural Information Processing Systems (NeurIPS), volume 6, pages 737–744, 1994

    work page 1994

  35. [35]

    Yikai Li, C. L. Philip Chen, and Tong Zhang. A survey on siamese network: Methodologies, applications, and opportunities. IEEE Transactions on Artificial Intelligence, 3(6):994–1014,

    work page

  36. [36]

    Conditional variational autoencoder with adver- sarial learning for end-to-end text-to-speech

    Jaehyeon Kim, Jungil Kong, and Juhee Son. Conditional variational autoencoder with adver- sarial learning for end-to-end text-to-speech. International Conference on Machine learning, 2021. 13

    work page 2021

  37. [2022]

    doi: 10.1109/TAI.2022.3207112

    work page arXiv 2022

  38. [2025]

    URL http://dx.doi.org/10.1109/ ICASSP49660.2025.10888676

    doi: 10.1109/icassp49660.2025.10888676. URL http://dx.doi.org/10.1109/ ICASSP49660.2025.10888676

    arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.