Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Dysfluent WFST: A Framework for Zero-Shot Speech Dysfluency Transcription and Detection

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A parameter-free decoder that aligns speech emissions to a reference-text weighted finite-state transducer can simultaneously transcribe phonemes of dysfluent speech and label repetitions, deletions, and insertions, with no training on…

desk verdict Clever reference-constrained WFST decoding, but the PER gains come from knowing the target text and the output alphabet cannot represent substitutions, so the SOTA claims are premature. read the letter →

arxiv 2505.16351 v2 pith:MJOEY6KG submitted 2025-05-22 eess.AS cs.AI

Chenxu Guo , Jiachen Lian , Xuanru Zhou , Jinming Zhang , Shuhe Li , Zongli Ye , Hwi Joo Park , Anaisha Das
show 8 more authors
Zoe Ezzes Jet Vonk Brittany Morin Rian Bogley Lisa Wauters Zachary Miller Maria Gorno-Tempini Gopala Anumanchipalli
This is my paper · ORCID
classification eess.AScs.AI
keywords speechdysfluencydetectionzero-shotdecodingweightedfinite-statetransducerCTCphonemerecognitionWavLMprimaryprogressiveaphasiaphoneticerrorrate
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

The paper claims that the bottleneck in dysfluent speech processing is phoneme transcription, and that a parameter-free decoder can improve it without additional training: given a reference text and a speech encoder's phone posteriors, it builds a weighted finite-state transducer whose shortest path both transcribes phonemes and labels repetitions, deletions, and insertions. The core idea is to replace greedy or unconstrained beam search with decoding constrained by pronunciation priors, so phonemes in the reference can be revisited, skipped, or inserted with costs that grow with index distance. On simulated repetition, deletion, and insertion data and on real clinical speech from patients with nonfluent primary progressive aphasia, the method reports lower phonetic error rates than greedy decoding with the same WavLM and wav2vec2 encoders, and near-perfect repetition detection. If right, it shows that explicit decoding constraints, rather than more complex architectures, are the key to usable dysfluency transcription.

What carries the argument

The carrying object is a weighted finite-state transducer (WFST) built from the reference phoneme sequence. States are reference phoneme indices; horizontal arcs emit each expected phoneme with weight $\alpha = 1 - 10^{-\beta}$; return arcs implement repetition, skip arcs implement deletion, and lowered arcs implement insertion, with non-forward arcs charged as $\mathrm{err} = \mathrm{err}_0 \cdot \frac{1}{\sqrt{2\pi}} e^{-x^2/2}$, where $x = |i-j|$ is the $L^1$ distance between phoneme indices. Composing this transducer with the CTC topology and the log-posterior emission graph, and taking the shortest path, yields a phoneme sequence whose state visits label each unit as normal, repetition, insertion, or deletion. The parameter $\beta$ is the only tunable knob and sets the base error rate for non-forward arcs.

What would settle it

Decode a set of natural, script-free dysfluent utterances whose reference text must be guessed from the audio; if the WFST decoder's phoneme error rate is no better than greedy search when the reference is imperfect, the claim that reference constraint drives the gain is not supported. A second check: feed a substitution dysfluency that replaces a reference phoneme with a phoneme absent from the reference; since the graph contains no arc for that phone except via insertion paths, the decoder's label for it can be tested explicitly.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that dysfluency detection becomes easy once phoneme transcription is accurate, and transcription accuracy can be improved by constraining the decoder with a reference-text WFST rather than by training on dysfluent speech. The decoder starts from an encoder emission tensor, composes the CTC topology with a reference FST extended by return arcs for repetition, skip arcs for deletion, and arcs for insertion, then intersects the result with the emission graph and takes the shortest path. Arcs that leave the monotonic forward path are charged an error probability that decays as a Gaussian of the L1 distance between phoneme indices in the reference. With WavLM-CTC as the encoder, the paper reports PER of 10.71%, 4.46%, and 2.80% on simulated repetitions, deletions, and insertions respectively, against 19.37%, 21.33%, and 19.80% for greedy search, and 100% counting accuracy for repetition detection; on the real clinical data the weighted phonetic error rate drops from 49.95% (greedy) to 32.20% (WFST).

Load-bearing premise

The method assumes the reference text is available and phonetically correct, and that every dysfluency is a local repeat, skip, or insertion relative to that reference whose cost depends only on phoneme index distance.

Editorial extensions

If this is right

  • A speech-language pathologist who already has the expected utterance text can get a timestamped phoneme transcription with repetition, deletion, and insertion labels from any CTC-trained encoder, with no dysfluent training data.
  • Because the decoder is parameter-free, swapping in better upstream encoders should directly improve both transcription and detection without retraining the decoding stage.
  • The reported 100% repetition-count accuracy indicates that the return-arc structure reliably exposes repeated phonemes, making repetition the most interpretable output of the system.
  • The same shortest-path construction should carry over to other structured edit phenomena, such as sound prolongations or articulatory substitutions, if their arcs can be encoded as reference-based transitions.

Reading between the lines

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

  • The gain depends on the reference being right, so the method's practical ceiling is set by how the reference is obtained; tightening the loop by using the decoder's own output to revise the reference could be tested as a second pass.
  • The Gaussian index-distance prior is a stand-in for phonetic, articulatory, or lexical similarity; replacing it with a learned or feature-based distance matrix is a natural experiment and may improve the weak deletion and insertion detection.
  • The emission-noise sensitivity the paper reports suggests that upstream encoder reliability, not decoder complexity, is the binding constraint; a denoising or uncertainty-aware front end would likely transfer directly to the WFST decoder.
  • Because repetition detection is nearly perfect while deletion and insertion detection is weaker, a clinician-facing tool could start by trusting repetition labels and use them as anchors for correcting the rest of the transcription.
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

5 major / 6 minor

Summary. The paper introduces Dysfluent-WFST, a zero-shot decoder that takes an audio signal and a reference text, composes a CTC emission graph with a reference-derived WFST, and returns both a phoneme sequence and dysfluency labels via shortest-path decoding. The method requires no additional training and is evaluated on simulated repetition/deletion/insertion datasets and on a small nfvPPA corpus. The authors report large reductions in PER/WPER relative to greedy search and claim state-of-the-art dysfluency detection, particularly for repetitions. The central idea is to model dysfluency as local return, skip, and backward arcs in a reference-anchored weighted finite-state transducer.

Significance. Dysfluent speech transcription is clinically relevant, and a lightweight, interpretable, reference-guided decoder would be a useful tool if the claims were supported. The paper's strengths include an open-source implementation, a clear algorithmic description, and an explicit design intuition: dysfluencies are structured deviations from a reference phoneme sequence. However, the evaluation as presented does not support the stated claims. The decoder is given the reference text while the greedy baselines are unconstrained, the output alphabet is closed under the reference phoneme set, and the simulated detection results are partly guaranteed by construction. The paper also does not compare against SSDM, which it itself identifies as the current state of the art. These issues are load-bearing for the abstract and introduction claims of state-of-the-art PER and dysfluency detection.

major comments (5)
  1. [Section 2.1, D = (T ◦ S) ∩ ξx] Because the WFST T is built from the reference phoneme sequence R, composing with T restricts the output alphabet of the decoding graph to phonemes that appear in R (plus epsilon and transition markers). The decoder therefore cannot emit a substitution phoneme absent from the reference, even when the acoustic evidence strongly supports it. This structural limitation invalidates the general 'more accurate phoneme transcription' claim: the method is not transcribing freely, it is constrained to the reference vocabulary. It also explains the near-perfect PER on simulated insertion data in Table 1 (e.g., WavLMCTC-WFST on simu-ins at 2.80%), because the simulated dysfluencies are generated from the same reference phoneme set.
  2. [Table 1 and Section 3.3] The central comparison is confounded: the WFST decoder receives the reference text (as described in Section 2.1), while the greedy search baselines are unconstrained phoneme recognizers. The PER gap in Table 1 therefore conflates the availability of the ground-truth reference with the contribution of the proposed WFST topology. A fair comparison requires a reference-conditioned baseline, for example a forced-alignment decoder or a CTC decoder with a phoneme-level language model that does not use the reference. Without such a baseline, the claim that WFST decoding itself improves phoneme transcription is not supported.
  3. [Section 3.3 and Table 2] The detection evaluation on simulated data is partly circular. The simulated dysfluencies are created by asking an LLM to insert repetitions, deletions, and insertions into reference phoneme sequences, and Algorithm 1 labels repetition whenever a state index repeats, deletion when it skips, and insertion when it goes backward. The 100% repetition accuracy is thus an artifact of the search-space construction, not evidence of acoustic discrimination. Please report event-level precision, recall, and F1 on real clinical data, and compare with SSDM [18], which Section 1 identifies as the current state-of-the-art pipeline.
  4. [Sections 1 and 3.3, Tables 1-2] The claim of state-of-the-art performance is unsupported because the current SOTA model SSDM [18] (and its extension [19]) is never evaluated or compared in any table. Table 2 only compares with YOLO-Stutter, and Table 1 compares only with greedy search. Adding the SSDM baseline is necessary for any SOTA claim, especially since the paper itself states in the introduction that SSDM is the current state of the art.
  5. [Section 3.4 and Table 3] The robustness experiment undermines the paper's positive framing. For W2V2-WFST, adding Gaussian noise at sigma=0.1 raises WPER from 9.43% to 74.06%, while the greedy baseline stays at 22.68%. The paper acknowledges this sensitivity in Section 4 but does not reconcile it with the headline claim that WFST decoding yields more accurate phoneme transcription. This issue needs to be addressed before the practical utility of the method can be assessed.
minor comments (6)
  1. [Abstract, Section 2.2, Section 3.2] The paper calls the method 'parameter-free' in the abstract and introduction, but Section 2.2 introduces the severity parameter beta and Section 3.2 requires a 39x39 phoneme similarity matrix S for WPER. At minimum these should be described as hyperparameters or manually specified priors, not as parameter-free.
  2. [Equation (1)] The notation '(i,j)X' in the WPER definition is malformed, and the summation over substitutions is not formally defined. Please rewrite the equation with a clear alignment operator and state how substitutions are identified relative to the reference.
  3. [Section 2.2] The dynamic weighting formula 'err = err0 1√ 2π e− x2 2' is missing a multiplication sign and does not specify the standard deviation of the Gaussian term; as written it is not a well-defined probability density.
  4. [Algorithm 1] The pseudocode contains typos ('lexcion', 'emision', 'IN P U T', 'E N C O D E') and the function 'detect_dysfluency' is better named for clarity, since it both transcribes and labels states.
  5. [Table 2] The metric 'Accuracy' is not defined. For dysfluency detection, please report per-class precision, recall, and F1, and specify whether evaluation is at the count level, event level, or timestamp level.
  6. [Figure 3] The x-axis label 'Beta' and the legend are unclear; it should be clarified whether PER and WPER share the same y-axis scale and which beta values were tested.

Circularity Check

2 steps flagged · score 5.0 of 10

Reference-constrained decoder's output space is closed under the reference phoneme set, and simulated dysfluency labels are generated in the same representation; the claimed zero-shot transcription/detection gains are partly by construction.

  1. self definitional [Section 2.1, Eq. D=(T ◦ S) ∩ ξx; Table 1]
    "The reference text is used to construct a reference FST, which, unlike traditional ASR relying solely on FSAs, accounts for dysfluencies like insertions, deletions, and repetitions by extending the FSA into a WFST. ... The decoding graph is then formed as: D = (T ◦ S) ∩ ξx"

    S is built from the reference phoneme sequence R, so composition restricts all output labels to phonemes appearing in R (plus epsilons). Substitutions that introduce a phoneme absent from R are unrepresentable. PER/WPER is then computed against the same R, so a decoder that simply stays on reference arcs is never charged for emitting a non-reference phone; the large PER gains over greedy search in Table 1 come substantially from feeding the reference into the decoder, not from better acoustic transcription. The paper explicitly states the input is the reference text, so this is a constrained-alignment task renamed as phoneme recognition; the abstract's 'transcribes phonemes' overstates the generality.

  2. self definitional [Section 3.1 (Simulated Dataset), Table 2, Algorithm 1]
    "We provided claude-3-5-sonnet-20241022 [29] with clean text and its corresponding CMU/IPA sequences, instructing it to insert dysfluencies at positions it deemed naturalistic. This process yielded both dysfluent IPA sequences (used for subsequent VITS [30] speech synthesis) and CMU sequences with dysfluency labels (matching our model's output format for evaluation purposes)."

    Dysfluency types are defined in Algorithm 1 purely by path behavior relative to reference indices: repetition when `start in state_history`, insertion when `start < min_time`, deletion when `start > prev_end + 1`. The simulated test data was created by asking an LLM to insert dysfluencies into the very same CMU/IPA sequences used to build the reference FST, and labels were emitted in the model's output format. Thus the simu-rep/simu-del/simu-ins tests check whether the decoder's path deviates in the same representational units that were used to construct the data; the 100% repetition detection and poor insertion detection (12-34%) are largely consequences of this shared representation, not independent acoustic evidence. Detection numbers for real nfvPPA speech are not reported.

full rationale

The paper's core mechanism is a reference-conditioned decoder: given the reference text, it builds a WFST whose arcs are labeled with phonemes from that reference, and then runs CTC-constrained shortest path on the emission lattice. This is a legitimate and interpretable forced-alignment approach, and the paper is explicit that the reference text is an input. The circularity arises in how the results are presented and evaluated. First, the decoding graph D = (T ◦ S) ∩ ξx makes the output alphabet closed under the reference phoneme set, so any phoneme not in the reference cannot be emitted. The claimed 'phoneme transcription' is therefore really an edit-constrained alignment of the reference to speech, yet the abstract and introduction present it as general phoneme transcription and report SOTA PER against the very reference that defines the output space. The comparison to unconstrained greedy search is not apples-to-apples, because the WFST decoder is handed the lexical content that greedy search must infer. Second, the dysfluency detection rules are defined entirely by reference-index path behavior, and the simulated benchmark is generated by inserting dysfluencies into the same CMU/IPA reference sequences, with labels in the model's own output format. On such a benchmark, high repetition detection (100%) is largely a consistency check of the shared representation rather than evidence that the decoder recognizes acoustic dysfluency independently. The real nfvPPA data provide some external grounding for PER, but no detection results are reported on that data, so the only detection evidence is the construction-matched simulation. Separately, the 'parameter-free' claim is weakened by the β severity parameter that is tuned via ablation per dataset (Section 2.2, Figure 3), though this is hyperparameter selection rather than a full by-construction reduction. No load-bearing self-citation was found: prior work by the same authors is cited as related systems, and the key external reference [27] is used only for dynamic weighting. Overall, the central claimed contributions are partially forced by the reference-constrained formulation and by the simulation protocol, but the decoder still performs real alignment work and the nfvPPA experiments show that it does not trivially achieve low error on real disordered speech, so the circularity is partial rather than total.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

One explicit free parameter, beta, sets the base dysfluency error rate and is tuned by ablation rather than fixed. The WPER similarity matrix is an external choice that drives the headline metric. The decoder relies on reference-text availability, a local edit model for dysfluency, and the reliability of pretrained CTC emissions for disordered speech; none of these is derived in the paper. No new physical or conceptual entity is introduced.

free parameters (2)
  • beta = not stated; swept from 0 to 10 in ablation (Figure 3)
    Sets the base error rate err0 = 10^-beta for return and skip arcs. The paper does not specify how beta is chosen for the main results, and Figure 3 shows it affects performance.
  • phoneme similarity matrix S for WPER = 39x39, source not described in text
    Defines WPER (Equation 1) and therefore the main metric in Tables 1 and 3. Its origin is not specified, and results could change under a different similarity matrix.
assumptions (5)
  • domain assumption The reference text is available and is the intended fluent target.
    Section 2.1: 'Our framework takes the reference text and corresponding speech as input.' The entire WFST graph is built from this text.
  • domain assumption Dysfluent speech can be described by local repetitions, deletions, and insertions of reference phonemes with monotonic index order.
    Figure 1 and Algorithm 1 classify all output via return, skip, and insertion arcs; substitutions outside the reference are not representable.
  • domain assumption Pretrained CTC encoders (wav2vec2, WavLM) provide reliable phonetic posteriors for disordered speech without fine-tuning.
    Section 2.1 uses outputs of [25,26] as emissions, and Section 3.3 shows performance is highly sensitive to emission noise.
  • domain assumption The 39x39 similarity matrix used in WPER is a valid phoneme distance.
    Equation 1 and Section 3.2 introduce WPER with an external similarity matrix, but no derivation or validation of the matrix is given.
  • standard math WFST composition, intersection, and shortest path are correct as implemented in k2.
    Algorithm 1 relies on k2 operations compose, intersect, and shortest_path; this is standard WFST machinery.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dysfluent WFST: A Framework for Zero-Shot Speech Dysfluency Transcription and Detection." pith.science (2026). https://pith.science/paper/MJOEY6KG

@misc{pith2026250516351,
  author       = {Pith},
  title        = {Pith review of: Dysfluent WFST: A Framework for Zero-Shot Speech Dysfluency Transcription and Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MJOEY6KG}},
  note         = {Machine review of arXiv:2505.16351}
}
read the original abstract

Automatic detection of speech dysfluency aids speech-language pathologists in efficient transcription of disordered speech, enhancing diagnostics and treatment planning. Traditional methods, often limited to classification, provide insufficient clinical insight, and text-independent models misclassify dysfluency, especially in context-dependent cases. This work introduces Dysfluent-WFST, a zero-shot decoder that simultaneously transcribes phonemes and detects dysfluency. Unlike previous models, Dysfluent-WFST operates with upstream encoders like WavLM and requires no additional training. It achieves state-of-the-art performance in both phonetic error rate and dysfluency detection on simulated and real speech data. Our approach is lightweight, interpretable, and effective, demonstrating that explicit modeling of pronunciation behavior in decoding, rather than complex architectures, is key to improving dysfluency processing systems.

Figures

Figures reproduced from arXiv: 2505.16351 by the authors.

Figure 1
Figure 1. Illustrating Dysfluency Transitions: Repetition, Dele￾tion, and Insertion in WFST plored core dysfluency detection problems: transcription and detection. To our knowledge, Dysfluent-WFST is the most lightweight, interpretable, and effective speech dysfluency tran￾scriber. 2. Method A weighted finite-state acceptor (WFSA) is a 6-tuple A = (Σ, Q, Qs, Qa, π, ω), where Σ is the input alphabet, Q is the set of states, Qs… view at source ↗
Figure 2
Figure 2. Decoder Workflow Based on WFST: The framework takes as input the speech signal and the corresponding reference text, and outputs the phoneme transcription sequence along with dysfluency detection (e.g., repetition). In this example, the reference text is ”She’s not here,” while the spoken audio is ”She’s n-not (N AA N AA T) here.” When using greedy search, the model may incorrectly force-align the repetition part to… view at source ↗
Figure 3
Figure 3. Impact of varying β values on transcription perfor￾mance on different dataset 4. Conclusion and future work Our decoder achieves strong performance in dysfluency tran￾scription, with a low phonetic error rate (PER) despite not re￾quiring dysfluent speech in training. The architecture is inter￾pretable and efficient. While it excels at detecting repetitions, its accuracy for insertions and deletions is lower. The WFS… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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.

Reference graph

Works this paper leans on

40 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [27]

    Simple and effective zero-shot cross-lingual phoneme recognition,

    Q. Xu, A. Baevski, and M. Auli, “Simple and effective zero-shot cross-lingual phoneme recognition,” Interspeech, 2022

    work page 2022

  2. [18]

    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,” in 2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 937–944

    work page 2024

  3. [19]

    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,” in 2023 IEEE Automatic Speech Recognition and Under- standing Workshop (ASRU), 2023, pp. 1–8

    2023

  4. [1]

    Introduction Automatic detection of speech dysfluency is a key module that helps speech-language pathologists efficiently transcribe disor- dered speech, which then serves as strong evidence for diagnos- tic purposes or even for treatment feedback planning. For a long time, research in this area has been constrained to a sentence classification task: dete...

    arXiv 2025

  5. [2]

    She’s not here

    Method A weighted finite-state acceptor (WFSA) is a 6-tuple A = (Σ, Q, Qs, Qa, π, ω), where Σ is the input alphabet,Q is the set of states, Qs ⊆ Q and Qa ⊆ Q are the sets of start and accept- ing states, respectively. The transition functionπ : Q × Σ → Q defines state transitions, while the weight functionω : Q×Σ → R assigns a weight to each transition. A...

    work page

  6. [3]

    Dataset (a) PPA Dataset [28]: The nfvPPA dataset consists of speech samples from 35 individuals diagnosed with Nonfluent/Agram- matic Primary Progressive Aphasia

    Experiments 3.1. Dataset (a) PPA Dataset [28]: The nfvPPA dataset consists of speech samples from 35 individuals diagnosed with Nonfluent/Agram- matic Primary Progressive Aphasia. For evaluation purposes, we manually annotated the phonemes using Praat and Adobe 1https://github.com/k2-fsa/k2 Audition, totally 1h1min12s. (b) Simulated Dataset : To evaluate ...

    work page

  7. [4]

    The architecture is inter- pretable and efficient

    Conclusion and future work Our decoder achieves strong performance in dysfluency tran- scription, with a low phonetic error rate (PER) despite not re- quiring dysfluent speech in training. The architecture is inter- pretable and efficient. While it excels at detecting repetitions, its accuracy for insertions and deletions is lower. The WFST de- coding app...

    work page

  8. [5]

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

Show all 40 references
  1. [6]

    Automatic stuttering recog- nition using hidden markov models,

    E. N ¨oth, H. Niemann, T. Haderlein, M. Decher, U. Eysholdt, F. Rosanowski, and T. Wittenberg, “Automatic stuttering recog- nition using hidden markov models,” in Interspeech, 2000

    2000

  2. [7]

    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

  3. [8]

    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

  4. [9]

    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,” arXiv preprint arXiv:2204.03417, 2022

    2022

  5. [10]

    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

  6. [11]

    Audio features from the wav2vec 2.0 embeddings for the acm multimedia 2022 stut- tering challenge,

    C. Montaci ´e, M.-J. Caraty, and N. Lackovic, “Audio features from the wav2vec 2.0 embeddings for the acm multimedia 2022 stut- tering challenge,” in Proceedings of the 30th ACM International Conference on Multimedia, 2022, pp. 7195–7199

    2022

  7. [12]

    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

  8. [13]

    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 (WOCCI 2017), 2017, pp. 1–6

    2017

  9. [14]

    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

  10. [15]

    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, 2022, pp. 6482–6486

    2022

  11. [16]

    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. [17]

    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,” in Interspeech, 2024

    2024

  13. [20]

    Towards hierarchical spoken lan- guage disfluency modeling,

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

    2024

  14. [21]

    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,” in Interspeech 2024, 2024, pp. 937–941

    2024

  15. [22]

    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-Tempini et al. , “Stutter-solver: End-to-end multi-lingual dysfluency detection,” in 2024 IEEE Spoken Language Technology Workshop (SLT) . IEEE, 2024, pp. 1039–1046

    2024

  16. [23]

    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,” in Advances in Neural In- formation Processing Systems, vol. 37, 2024

    2024

  17. [24]

    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

  18. [25]

    Chatgpt,

    OpenAI, “Chatgpt,” 2022. [Online]. Available: https://openai. com/chatgpt/

    2022

  19. [26]

    Univer- sal phone recognition with a multilingual allophone system,

    X. Li, S. Dalmia, J. Li, M. Lee, P. Littell, J. Yao, A. Anastasopou- los, D. R. Mortensen, G. Neubig, A. W. Black et al. , “Univer- sal phone recognition with a multilingual allophone system,” in ICASSP. IEEE, 2020, pp. 8249–8253

    2020

  20. [28]

    Wavlm-ctc-hugginface,

    “Wavlm-ctc-hugginface,” https://huggingface.co/microsoft/ wavlm-large

  21. [29]

    Con- nectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,

    A. Graves, S. Fern ´andez, F. Gomez, and J. Schmidhuber, “Con- nectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,” in International Conference on Machine learning, 2006, pp. 369–376

    2006

  22. [30]

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

    A. Baevski, H. Zhou, A. rahman Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,”Advances in Neural Information Process- ing Systems,, 2020

    2020

  23. [31]

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

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, J. Wu, L. Zhou, S. Ren, Y . Qian, Y . Qian, J. Wu, M. Zeng, X. Yu, and F. Wei, “Wavlm: Large- scale self-supervised pre-training for full stack speech process- ing,” IEEE JSTSP, 2022

    2022

  24. [32]

    Weakly-supervised forced alignment of disfluent speech using phoneme-level modeling,

    T. Kouzelis, G. Paraskevopoulos, A. Katsamanis, and V . Kat- souros, “Weakly-supervised forced alignment of disfluent speech using phoneme-level modeling,” Interspeech, 2023

    2023

  25. [33]

    Classification of primary progressive aphasia and its variants,

    M. L. Gorno-Tempini, A. E. Hillis, S. Weintraub, A. Kertesz, M. Mendez, S. F. Cappa, J. M. Ogar, J. D. Rohrer, S. Black, B. F. Boeve et al., “Classification of primary progressive aphasia and its variants,” Neurology, vol. 76, no. 11, pp. 1006–1014, 2011

    2011

  26. [34]

    From llm to nmt: Ad- vancing low-resource machine translation with claude,

    M. Enis and M. Hopkins, “From llm to nmt: Ad- vancing low-resource machine translation with claude,” ArXiv, vol. abs/2404.13813, 2024. [Online]. Available: https://api.semanticscholar.org/CorpusID:269292906

    2024 arXiv

  27. [35]

    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,” Interna- tional Conference on Machine learning, 2021

    2021

  28. [36]

    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

  29. [37]

    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

  30. [38]

    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,” in Proc. INTER- SPEECH 2023, 2023, pp. 5132–5136

    2023

  31. [39]

    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

  32. [40]

    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,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (IC...

    2023

Pith tools

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