A Comparative Study of Controllability, Explainability, and Performance in Dysfluency Detection Models

Eric Zhang; Li Wei; Michael Wang; Sarah Chen

REVIEW 4 major objections 5 minor 14 references

Dysfluency detection models need more than accuracy: a three-axis comparison finds UDM best balanced for clinical use.

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 comparative study of dysfluency detection models finds UDM best balances accuracy and clinical interpretability, while SSDM is not reproducible from its published description.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Useful comparative map of dysfluency models, but the UDM recommendation rests on three unblinded raters and a small F1 gap. the 4 major comments →

arxiv 2509.00058 v1 pith:EL2B5C2K submitted 2025-08-25 cs.AI

A Comparative Study of Controllability, Explainability, and Performance in Dysfluency Detection Models

Eric Zhang , Li Wei , Sarah Chen , Michael Wang This is my paper

classification cs.AI

keywords dysfluency detectionstuttered speechclinical AIexplainabilitycontrollabilitymodel comparisonphoneme alignmentUClass benchmark

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 tries to show that accurate dysfluency detection alone is not enough for clinical use: models also need controllability and explainability. It introduces a three-axis benchmark, UClass, and applies it to four representative models. Its central finding is that UDM, a modular model with explicit phoneme alignment, achieves the best balance, with the highest detection scores and far higher clinician ratings for transparency and adjustability than YOLO-Stutter or FluentNet. The paper also reports that SSDM, a theoretically promising structured model, could not be reproduced, which limits the current landscape. If the ratings are trustworthy, the practical consequence is that clinical teams should favor UDM-like architectures and evaluation frameworks that measure usability alongside accuracy.

Core claim

The paper's central claim is that UDM is the most clinically applicable dysfluency detection model among the four compared: it scores highest on performance (F1 0.89, precision 0.88, recall 0.90, balanced accuracy 0.88) and is rated far above YOLO-Stutter and FluentNet by speech-language pathologists for controllability (4.0/5) and explainability (4.2/5). The authors attribute this to UDM's modular design, which exposes phoneme-level alignments and threshold parameters, making its decisions inspectable and adjustable. In contrast, YOLO-Stutter offers real-time efficiency and FluentNet offers simple stable classification, but both lack clinically meaningful intermediate outputs. SSDM could no

What carries the argument

The carrying mechanism is the UClass benchmark: a unified evaluation protocol that scores models on performance, controllability, and explainability, with the latter two rated by expert clinicians. Within that framework, UDM's modular architecture is the key object: explicit phoneme alignment via CTC-attention hybrids, a temporal pattern analyzer, and an unconstrained classifier produce intermediate representations clinicians can inspect, while adjustable thresholds and modular retraining enable adaptation. This architecture is what the paper argues earns UDM its clinical advantage.

Load-bearing premise

The recommendation that UDM is the best choice for clinical use rests on the assumption that the scores given by three clinician raters reliably measure real-world controllability and explainability; the paper does not report how the raters were selected, whether they were blinded to model identity, or whether the rating scale was validated.

What would settle it

A blinded usability study in which a larger, diverse panel of speech-language pathologists uses each model on identical clinical tasks, without knowing which model is which, would settle the claim. If UDM no longer scores higher on controllability and explainability than YOLO-Stutter or FluentNet, the paper's central conclusion fails.

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

If this is right

Clinical deployments of dysfluency detection will likely need to include interpretable intermediate outputs, because the clinicians in this study gave much higher usability scores to UDM's phoneme-aligned design.
YOLO-Stutter remains competitive for real-time interactive applications, where its speed offsets its opacity.
FluentNet's simplicity makes it a plausible screening or triage tool, but its binary classification cannot support detailed diagnosis of specific dysfluency types.
SSDM's theoretical promise cannot be evaluated until its code and training details are made reproducible.
Benchmark frameworks for clinical AI should follow UClass in measuring controllability and explainability alongside accuracy.

Where Pith is reading between the lines

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

A testable extension the paper does not run: equip YOLO-Stutter or FluentNet with a phoneme-alignment 'explanation head' and see whether clinician-rated explainability rises to UDM's level; if it does, modular transparency rather than the whole UDM architecture is the active ingredient.
UDM's advantage may shrink in clinical settings where phoneme transcriptions are unavailable or costly to obtain, since its alignment module depends on them.
The high reported inter-rater agreement (k > 0.8) among only three raters does not exclude systematic bias; a larger blinded panel could change the ranking.
If the trade-off generalizes, future hybrid designs could combine a fast detector with a lightweight, structured explanation module rather than choosing between speed and interpretability.

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.

Desk Editor's Note

Useful comparative map of dysfluency models, but the UDM recommendation rests on three unblinded raters and a small F1 gap.

read the letter

Short version: this is a readable, useful comparison of four dysfluency detectors, and the UClass framework is a reasonable attempt to broaden evaluation beyond F1. The central claim—that UDM is the clinically best choice—is driven mostly by clinician ratings from just three SLPs, and the paper gives no evidence those ratings are reliable or unbiased. Section 6.4 concedes potential bias, and the performance differences are modest (UDM 0.89 vs FluentNet 0.86, overlapping error bars).

What's new: the head-to-head comparison across controllability and explainability, dimensions usually ignored in dysfluency detection evals. The paper also does a good job articulating deployment trade-offs (efficiency vs transparency, training burden) and is honest about SSDM's reproducibility problems, at least at the level of stating them. The authors clearly know the literature and the clinical context.

The soft spots are where the reader's conditional verdict lands. The clinical assessment is load-bearing: three raters, no detail on recruitment, blinding, whether they saw live systems or paper descriptions, or whether the 1–5 scale was validated. Inter-rater reliability k>0.8 only shows agreement, not correctness; if the raters shared a prior that modular architectures are more clinically usable, the ratings would look exactly like this. The performance table reports only averages across datasets, no per-dataset numbers or significance tests, so we can't judge whether the F1 gaps are meaningful. The SSDM reproducibility claim is asserted without supporting evidence—no attempt logs, no specific inconsistencies—which makes it hard to evaluate and easy to read as convenient. And no code or data for UClass is released, limiting the benchmark's adoption. Minor: the name UClass collides with the UCLASS Corpus (ref [2]), which could confuse readers.

For whom: a clinical AI reading group would get a good discussion out of this, especially on how hard it is to evaluate interpretability. As a rigorous comparative benchmark, it needs work. I'd send it to peer review because the question matters and the framework is worth refining, but I'd expect major revisions: detail the clinical protocol, report per-dataset results with significance tests, and substantiate or soften the SSDM claim.

Recommendation: treat it as a discussion piece, not a source for the UDM conclusion yet.

Referee Report

4 major / 5 minor

Summary. The paper presents a comparative evaluation of four dysfluency-detection models (YOLO-Stutter, FluentNet, UDM, SSDM) along three axes: detection performance, controllability, and explainability. The authors introduce a 'UClass' benchmark that combines quantitative metrics with ratings from three certified speech-language pathologists on 1–5 scales. They report that UDM has the best overall performance (F1 0.89±0.03) and far higher controllability/explainability scores (4.0/4.2) than YOLO-Stutter (2.1/2.3) or FluentNet (2.4/2.6), while SSDM could not be reproduced. The paper concludes that UDM is the most clinically applicable model and that the UClass framework provides a more holistic evaluation standard for clinical adoption.

Significance. If the evidence were robust, this would be a useful contribution to the clinical-AI evaluation literature: it addresses the gap between benchmark accuracy and the interpretability/controllability requirements of speech-language pathology practice, and it makes the pragmatic trade-offs among object-detection, CNN, and modular alignment-based designs explicit. The paper also deserves credit for comparing external, non-proprietary models and for discussing deployment considerations such as real-time factor, memory, and training time. However, the central quantitative and clinical claims rest on incomplete evidence: the performance table gives only averaged metrics without inferential statistics, and the clinician ratings—which drive the main conclusion—are reported without methodological detail or validation. The contribution is therefore promising but not yet established.

major comments (4)

[§5.4.2, Table 2; §6.4; §7] The central conclusion that UDM is the best clinical choice depends on the clinician-rated controllability/explainability gaps, not on performance differences (which overlap within one standard deviation). The manuscript provides no protocol details for these ratings: how the three SLPs were recruited, whether they were blinded to model identity, whether they interacted with real systems or only paper descriptions, whether the 1–5 scale had been validated, or what individual ratings were. The statement that inter-rater reliability was high (k > 0.8) only shows agreement, and shared preconceptions about modular/alignment architectures would also produce high agreement. Section 6.4 itself concedes 'limited number of expert raters, potentially introducing bias.' Because the main clinical recommendation rests on this unvalidated rating procedure, this is a load-bearing gap.
[§5.4.1, Table 1] Table 1 reports only aggregate mean±std F1, precision, recall, and balanced accuracy across multiple datasets. There are no per-dataset results, sample sizes, significance tests, or confidence intervals. With UDM at 0.89±0.03 and FluentNet at 0.86±0.04, the differences are within one standard deviation and are not shown to be statistically distinguishable. The paper's claim that UDM achieves 'the highest overall performance across all metrics' and its summary in §7 that UDM is 'most balanced profile across all evaluation dimensions' are not supported by the reported numbers. Per-dataset results and appropriate statistical comparisons are needed.
[§3.4 and §5.4.1] The claim that SSDM is 'not reproducible' is stated as a finding in the abstract and conclusions, but the manuscript does not document what was actually attempted: which versions of the code and data were used, what implementation steps failed, what discrepancies were observed, or whether the authors contacted the original developers. Given that SSDM is one of only four models in the comparison, its exclusion should be supported by a concrete reproducibility protocol (e.g., exact commits, expected versus obtained metrics, and failure logs). Without this, the reproducibility claim is not verifiable.
[§5.4.3, Table 3] The computational-efficiency results are used in the Discussion to distinguish YOLO-Stutter as real-time and UDM as resource-heavy, but Table 3 provides no measurement methodology. It is unclear how the real-time factor was computed, which hardware/software settings were used beyond 'A100 GPUs and PyTorch 1.12' (Sec. 5.2), how memory was measured, or whether these numbers are from the authors' runs or the original papers. Because deployment guidance is part of the paper's contribution, the efficiency comparison needs a clearly described measurement protocol or should be presented as approximate.

minor comments (5)

[§5.3 and §5.4.2] The inter-rater reliability statistic is reported as 'k > 0.8' but the type of kappa (Cohen's, Fleiss's, weighted?) is not specified, and no per-dimension reliability values are given. Please clarify.
[§5.1] The corpus is called 'UCLASS Corpus' in the text and 'UClass benchmark' in the table captions, while reference [2] is the original UCLASS archive. The relationship between the UClass evaluation framework and the UCLASS corpus should be disambiguated.
[§7] The conclusion says UDM provides 'strong evidence for its clinical applicability'; given the limitations acknowledged in §6.4 and the issues above, a more cautious phrasing such as 'suggests potential' would be more appropriate.
[References] Some reference entries contain formatting errors (e.g., 'V onk' instead of 'Vonk', 'Zachary Mille' instead of 'Miller') and reference [2] lacks venue details (journal name and volume/pages are present, but the year and title formatting should be checked).
[General] No code or data availability statement is included. Because the paper proposes a benchmark (UClass) and reports reproducibility difficulties for SSDM, a public release of the evaluation pipeline and per-dataset results would substantially strengthen the contribution.

Circularity Check

0 steps flagged

No significant circularity; the comparison is an external empirical evaluation with no fitted-input-called-prediction chain.

full rationale

This paper performs a comparative evaluation of four existing models using standard performance metrics on external datasets (LibriStutter, UCLASS Corpus, FluencyBank, Clinical Validation Set) and clinician ratings. There is no mathematical derivation in which an output is constructed from its own input. None of the compared models is developed by the authors, so the performance numbers are not self-citations. The UClass framework is introduced in this paper, but it is an evaluation rubric, not a source of fitted parameters used to predict the same ratings; the controllability/explainability scores are directly reported expert judgments rather than predictions derived from those same judgments. The only notable weakness—Section 6.4's concession that 'Clinical assessment was performed by a limited number of expert raters, potentially introducing bias'—concerns measurement validity and generalizability, not circularity. No equation or fitted parameter is renamed as a prediction, and no load-bearing claim reduces by definition to its inputs. Accordingly, no circular step is identified.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

The paper introduces no fitted parameters and no new physical or conceptual entities. Its conclusions rest on the validity of the UClass framework, the reliability of the three clinician raters, and the fairness of the evaluation protocols. These are domain assumptions that are stated but not independently validated.

axioms (3)

domain assumption The UClass evaluation dimensions (controllability, explainability) are valid and meaningful proxies for clinical utility.
Invoked throughout Sections 4 and 5; the paper's main conclusion depends on these subjective ratings being meaningful.
domain assumption The three clinician raters provide unbiased, reliable assessments (inter-rater kappa > 0.8).
Stated in Section 5.3, but no details on rater selection, blinding, or scale validation are given.
domain assumption The models are evaluated fairly with identical preprocessing and protocols.
Stated in Section 5.2; if the implementations are not faithful to the original papers, the comparison is unfair.

reviewed 2026-08-05 · how reviews work

0 comments

Cite this review

Pith. "Pith review of A Comparative Study of Controllability, Explainability, and Performance in Dysfluency Detection Models." pith.science (2026). https://pith.science/paper/EL2B5C2K

@misc{pith2026250900058,
  author       = {Pith},
  title        = {Pith review of: A Comparative Study of Controllability, Explainability, and Performance in Dysfluency Detection Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EL2B5C2K}},
  note         = {Machine review of arXiv:2509.00058}
}

Share X Bluesky LinkedIn Reddit HN

read the original abstract

Recent advances in dysfluency detection have introduced a variety of modeling paradigms, ranging from lightweight object-detection inspired networks (YOLOStutter) to modular interpretable frameworks (UDM). While performance on benchmark datasets continues to improve, clinical adoption requires more than accuracy: models must be controllable and explainable. In this paper, we present a systematic comparative analysis of four representative approaches--YOLO-Stutter, FluentNet, UDM, and SSDM--along three dimensions: performance, controllability, and explainability. Through comprehensive evaluation on multiple datasets and expert clinician assessment, we find that YOLO-Stutter and FluentNet provide efficiency and simplicity, but with limited transparency; UDM achieves the best balance of accuracy and clinical interpretability; and SSDM, while promising, could not be fully reproduced in our experiments. Our analysis highlights the trade-offs among competing approaches and identifies future directions for clinically viable dysfluency modeling. We also provide detailed implementation insights and practical deployment considerations for each approach.

discussion (0)

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages · 2 internal anchors

[1]

FluentNet: End-to-End Detection of Speech Disfluency with Deep Learning

Tedd Kourkounakis, Amirhossein Hajavi, and Ali Etemad. Fluentnet: end-to-end detection of speech disfluency with deep learning. arXiv preprint arXiv:2009.11394, 2020

work page internal anchor Pith review Pith/arXiv arXiv 2009
[2]

The university college london archive of stuttered speech (uclass)

Peter Howell, Stephen Davis, and Jon Bartrip. The university college london archive of stuttered speech (uclass). Journal of speech, language, and hearing research, 52(2):556–569, 2009

work page 2009
[3]

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

work page 2024
[4]

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

work page 2024
[5]

Time and tokens: Benchmarking end-to-end speech dysfluency detection, 2024

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

work page 2024
[6]

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. IEEE, 2023

work page 2023
[7]

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 (Volume 1: Long Papers), 2024

work page 2024
[8]

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. 7

work page 2024
[9]

Automatic detection of articulatory-based disfluencies in primary progressive aphasia

Jiachen Lian, Xuanru Zhou, Chenxu Guo, Zongli Ye, Zoe Ezzes, Jet V onk, Brittany Morin, David Baquirin, Zachary Mille, Maria Luisa Gorno Tempini, and Gopala Krishna Anumanchipalli. Automatic detection of articulatory-based disfluencies in primary progressive aphasia. IEEE JSTSP, 2025

work page 2025
[10]

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

work page 2025
[11]

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

work page internal anchor Pith review Pith/arXiv arXiv 2025
[12]

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

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. Interspeech, 2025

work page 2025
[13]

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. Interspeech, 2025

work page 2025
[14]

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. Journal of fluency disorders, 56:69–80, 2018. 8

work page 2018

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

[1] [1]

FluentNet: End-to-End Detection of Speech Disfluency with Deep Learning

Tedd Kourkounakis, Amirhossein Hajavi, and Ali Etemad. Fluentnet: end-to-end detection of speech disfluency with deep learning. arXiv preprint arXiv:2009.11394, 2020

work page internal anchor Pith review Pith/arXiv arXiv 2009

[2] [2]

The university college london archive of stuttered speech (uclass)

Peter Howell, Stephen Davis, and Jon Bartrip. The university college london archive of stuttered speech (uclass). Journal of speech, language, and hearing research, 52(2):556–569, 2009

work page 2009

[3] [3]

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

work page 2024

[4] [4]

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

work page 2024

[5] [5]

Time and tokens: Benchmarking end-to-end speech dysfluency detection, 2024

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

work page 2024

[6] [6]

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. IEEE, 2023

work page 2023

[7] [7]

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 (Volume 1: Long Papers), 2024

work page 2024

[8] [8]

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. 7

work page 2024

[9] [9]

Automatic detection of articulatory-based disfluencies in primary progressive aphasia

Jiachen Lian, Xuanru Zhou, Chenxu Guo, Zongli Ye, Zoe Ezzes, Jet V onk, Brittany Morin, David Baquirin, Zachary Mille, Maria Luisa Gorno Tempini, and Gopala Krishna Anumanchipalli. Automatic detection of articulatory-based disfluencies in primary progressive aphasia. IEEE JSTSP, 2025

work page 2025

[10] [10]

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

work page 2025

[11] [11]

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

work page internal anchor Pith review Pith/arXiv arXiv 2025

[12] [12]

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

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. Interspeech, 2025

work page 2025

[13] [13]

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. Interspeech, 2025

work page 2025

[14] [14]

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. Journal of fluency disorders, 56:69–80, 2018. 8

work page 2018

A Comparative Study of Controllability, Explainability, and Performance in Dysfluency Detection Models

The reading

Core claim §

What carries the argument

Load-bearing premise §

What would settle it §

If this is right

Where Pith is reading between the lines

discussion (0)

Core claim

Load-bearing premise

What would settle it