Pith. sign in

REVIEW 6 major objections 6 minor 3 cited by

LCS-CTC: Leveraging Soft Alignments to Enhance Phonetic Transcription Robustness

T0 review · 6 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read By using a longest-common-subsequence alignment mask to anchor high-confidence frames and prune CTC paths, LCS-CTC improves phoneme transcription and forced alignment on both fluent and non-fluent speech.

desk verdict A real but incremental CTC training trick that consistently beats vanilla CTC on the reported sets, but the forced-alignment claims rest on an unverified cost-model transfer and the WPER metric shares the same hand-built similarity as the method. read the letter →

arxiv 2508.03937 v2 pith:YMTDAW7V submitted 2025-08-05 eess.AS

Zongli Ye , Jiachen Lian , Akshaj Gupta , Xuanru Zhou , Haodong Li , Krish Patel , Hwi Joo Park , Dingkun Zhou
show 13 more authors
Chenxu Guo Shuhe Li Sam Wang Iris Zhou Cheol Jun Cho Zoe Ezzes Jet M.J. Vonk Brittany T. Morin Rian Bogley Lisa Wauters Zachary A. Miller Maria Luisa Gorno-Tempini Gopala Anumanchipalli
This is my paper · ORCID
classification eess.AS
keywords phonemetranscriptionCTClongestcommonsubsequencesoftalignmentforcednon-fluentspeechprimaryprogressiveaphasiarecognition
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

LCS-CTC is a phoneme-level speech recognizer that replaces the unconstrained path space of CTC with a partially constrained one. A lightweight network first predicts a frame-by-frame cost between each speech frame and each target phoneme; a modified Longest Common Subsequence routine turns that cost matrix into a binary alignment mask marking only high-confidence frame-phoneme matches. During training the mask both anchors the masked frames to phoneme labels via cross-entropy and prunes implausible CTC decoding paths. The paper reports that, across three encoders and three corpora—fluent LibriSpeech, clinical PPA speech, and synthetic dysfluent speech—LCS-CTC lowers phoneme error, weighted phoneme error, boundary loss, and articulatory reconstruction loss relative to vanilla CTC, and yields text-free forced alignment. The significance, if the result holds, is that one training objective can serve both accurate phoneme recognition and clinically interpretable alignment for fluent and non-fluent speech.

What carries the argument

The load-bearing object is the similarity-aware LCS alignment mask. Starting from a predicted cost matrix $C$, the algorithm collects phoneme-frame pairs $(i,j)$ whose cost is below a similarity-adjusted threshold $\tau_{i,j}=(1-s(p_i,p_k))\cdot\mathrm{tol}$ (with $\mathrm{tol}=1$), then runs a modified longest-common-subsequence dynamic program with horizontal propagation through consecutive valid matches and a traceback. The resulting binary mask is the single mechanism that both anchors frames through cross-entropy and restricts CTC emissions, so the whole improvement claim rests on whether this mask identifies trustworthy alignment zones.

What would settle it

Take the PPA or LLM-dys test set and compare LCS-CTC trained with predicted masks against the same model trained with oracle masks computed from ground-truth phonetic boundaries. If oracle masks do not improve over predicted masks, the claimed benefit of high-confidence mask selection is not what drives the gains. Alternatively, retrain the cost model without manual correction, or on one corpus only, and check whether the advantage over vanilla CTC on lvppa and nfvppa survives; if it vanishes, the method's robustness claim rests entirely on alignment-supervision transfer.

Watch

Extended reading notes

Core claim

The central claim is that CTC's weakness on unclear and non-fluent speech is best addressed not by full frame-level labels, which can inject noise, but by sparse, high-confidence alignment anchors. LCS-CTC predicts a frame-phoneme cost matrix and applies a similarity-aware LCS algorithm to identify only locally confident matches. These matches are turned into a mask that is used twice: masked positions get direct cross-entropy supervision, and the CTC emission probabilities are renormalized over the mask, restricting the set of paths the model can decode. The paper argues this reduces overfitting and peaky CTC behavior while preserving sequence-level flexibility, and it demonstrates consiste

Load-bearing premise

The central risk is that the cost model, trained on clean read speech with corrected alignments, still ranks frame-phoneme matches accurately on out-of-domain, pathological, and dysfluent audio; if its confidence rankings do not transfer, LCS-CTC's masks will anchor wrong frames and prune correct ones.

Editorial extensions

If this is right

  • If LCS-CTC's gains are real, CTC-based phoneme recognizers can be made more reliable on fluent and non-fluent speech without requiring manual frame-level labels at training time.
  • The same mask enables text-free forced alignment, so recognition and alignment quality improve together rather than being traded off.
  • Because the method helps on clean LibriSpeech as well as aphasic and dysfluent data, it suggests CTC's peaky-alignment problem, not dysfluency per se, is a major source of phoneme error.
  • The reduced boundary loss and articulatory reconstruction loss imply outputs are temporally and articulatorily more plausible, which matters for clinical analysis of disordered speech.
  • The approach is encoder-agnostic and reported to help across Wav2Vec2.0, HuBERT, and WavLM backbones.

Reading between the lines

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

  • A natural extension the paper does not test: applying the same mask-constrained objective to word- or subword-level CTC, where peaky alignments also hurt, may transfer the robustness gains beyond phonemes.
  • The current pipeline depends on a separately trained cost model supervised by alignments corrected on a fraction of clean read speech; the paper leaves open whether jointly learning the cost model with the recognizer, or training it self-supervised on unlabeled audio, would preserve the gains.
  • If the cost model's confidence rankings transfer, the similarity-aware threshold could be tuned per speaker or per disorder subtype, which the fixed global tolerance currently ignores.
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

6 major / 6 minor

Summary. The paper proposes LCS-CTC, a two-stage phoneme recognition framework. Stage 1 learns a frame-phoneme cost matrix on VCTK using MFA-derived and manually corrected alignments, then applies a similarity-aware Longest Common Subsequence (LCS) algorithm to produce a binary alignment mask. Stage 2 uses this mask to constrain CTC emissions and to add frame-level cross-entropy anchors, with total loss lambda*CE + (1-lambda)*masked CTC. Experiments on LibriSpeech, PPA, and the synthetic LLM-dys corpus report lower PER/WPER and improved boundary and articulatory reconstruction losses relative to vanilla CTC and a CE-CTC baseline across Wav2Vec2, HuBERT, and WavLM encoders. The paper also claims text-free forced alignment as a byproduct of the method.

Significance. If the results hold, the contribution is useful: a single training objective that improves both phoneme recognition and temporal alignment for fluent and non-fluent speech, with a conceptually interpretable alignment mechanism. The paper is strengthened by evaluating three self-supervised encoders, several corpora, and by open-sourcing code and checkpoints. The largest reported gain, WavLM-L on PPA lvppa (PER 20.33 to 18.72), is clinically meaningful if reproducible. However, the current empirical support is thin: all results are single runs, several margins are under 0.5 PER/WPER points, the test-time alignment protocol is unspecified, and the cost-matrix model is trained on clean VCTK only. These issues make the central claim plausible but not yet established at journal standards.

major comments (6)
  1. [§III-D, Tables I–IV] All results are single runs with no error bars, confidence intervals, or significance tests. Many differences are small (e.g., Wav2Vec2-L test-clean PER 16.16 vs 16.09; WavLM-L test-clean WPER 10.75 vs 10.49). Because the headline claim is "consistently outperforms" across all subsets and encoders, run-to-run variance should be reported, with a correction for multiple comparisons. As written, the reader cannot distinguish systematic gains from noise.
  2. [§II-A.c, §IV] The cost-matrix model is trained on only 30% of VCTK with MFA and manual alignments, then used to generate LCS masks for training on VCTK and for evaluations on LibriSpeech, PPA, and LLM-dys. There is no direct validation of cost-model transfer to out-of-domain or pathological speech. If the model's frame-phoneme confidence is miscalibrated on PPA or dysfluent speech, the masks anchor incorrect frames and may prune correct CTC paths, which could reverse the reported alignment gains in Tables III–IV. Section IV itself concedes that clean VCTK training is insufficient to assess generalization. The authors should provide a direct transfer evaluation of the cost model or explicitly restrict the claims to in-domain conditions.
  3. [§IV, Tables III–IV] The claimed "text-free forced alignment" protocol is never described. Boundary Loss requires a precise definition of predicted boundaries, but it is unclear whether boundaries come from Viterbi decoding of CTC emissions, from applying the cost model to test audio with predicted phonemes, or from the LCS mask itself. The comparison to vanilla CTC depends entirely on this protocol. Please specify the inference-time alignment procedure and the boundary extraction method, and clarify how the BL and ARL metrics are computed.
  4. [§III-C.2, §II-A.c] WPER is defined with the same hand-built articulatory similarity function s that is used in the cost-matrix targets, the LCS acceptance threshold, and the masked CTC/CE losses. This does not make the vanilla-CTC comparison circular, since all methods are scored with the same metric, but the metric is not independent of the training signal and may favor LCS-CTC by construction. Please report at least one alternative error metric (e.g., unweighted phoneme confusion or a separately derived distance) or discuss the coupling explicitly.
  5. [Algorithm 1, §II-A.c] The pseudocode's indexing is ambiguous and appears inconsistent with the text. Line 4 sets k = arg min C[:,j] and then uses Label[k] as phn2, so the threshold for cell (i,j) depends on k, which is independent of i; this does not match the stated per-pair threshold tau_{i,j} = (1 - s(pi,pj)) * tol. The set valid_res also contains the transposed pair (j,i), which is only valid if n=m. Lines 14–17 implement a nonstandard horizontal propagation whose intended recurrence is not clearly specified. As written, the alignment procedure is not exactly reproducible; please correct the indexing and define the DP recurrence explicitly.
  6. [§III-D, §I] The Introduction criticizes existing CTC-regularization methods [24]–[26], but the only non-vanilla baseline is CE-CTC. No comparison is made to Bayes-risk CTC, label-prior CTC, or CR-CTC under the same training recipe. Consequently, the claim that LCS-CTC is preferable to prior remedies is unsupported. Please add at least one of these baselines to the experimental comparison.
minor comments (6)
  1. [§II-A.b] After defining s as a raw articulatory similarity, the text says "For convenience, we denote s(pi,pj) as the similarity values after softmax." This redefinition of the same symbol is confusing; use a different notation or state the transformation once.
  2. [Fig. 2 caption] The caption says the LCS aligner uses ground-truth phonemes and predicted emissions. Ground-truth phonemes are not available at test time. Clarify whether the figure depicts training only and how the claimed text-free alignment is obtained at inference.
  3. [§III-B] No random seed is reported. With batch size 1, training is likely seed-sensitive; report seeds and any evaluation checkpoint selection procedure.
  4. [§II-A.b] The Gaussian edge attenuation applied to the ground-truth aligned region is not specified. Since it shapes all cost-matrix targets, its width and normalization should be given.
  5. [§II-B.2, §III-E] The weighting factor lambda is set to 0.5 without an ablation, while the ablation only studies tol. Report a sensitivity analysis for lambda or justify the choice.
  6. [§III-A] LLM-dys is a synthetic dataset generated with a specific LLM and VITS. Its representativeness for natural non-fluent speech should be discussed, especially since conclusions about "non-fluent speech" are partly based on this synthetic corpus.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: LCS-CTC's gains over vanilla CTC are empirically independent; only minor shared-metric and self-citation caveats.

full rationale

The central claim—that LCS-CTC outperforms vanilla CTC—is not circular. The recognition model is trained on 50% VCTK using masks produced by a separately trained cost matrix model (30% VCTK, MFA + manual corrections), and at test time decoding is standard CTC; the fitted tolerance tol and the hand-built similarity s do not enter inference. The PER comparisons (Table I) use unweighted Levenshtein distance, an external benchmark not defined in terms of the method's components. The CE-CTC baseline provides a fair additional comparison. Two caveats prevent a fully clean bill but do not amount to circularity. First, WPER (Section III-C.2) is defined using the same phoneme similarity s that defines the cost-matrix targets (Section II-A.b) and the LCS threshold, so WPER is not a fully independent metric; the PER results, however, are independent. Second, the ARL metric (Section III-C.4) uses the co-authors' AAI model [41]; this is a published external tool and is not used in training, so it is a minor self-citation rather than load-bearing. The paper also leaves the text-free forced-alignment protocol unspecified (Section III.D) and concedes in Section IV that VCTK-only training 'is insufficient to fully assess its generalizability' to pathological speech; these are missing-support/correctness risks, not circular derivations. No equation is fitted to the test labels, and no prediction is a renamed parameter of the training objective.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central claim rests mainly on the cost-matrix aligner and the hand-built similarity function. tol is fitted to validation data; lambda is hand-set; the Gaussian edge-attenuation width is not reported. The MFA/manual-correction supervision and the assumption that VCTK-trained costs transfer to LibriSpeech, PPA, and LLM-dys are the heaviest external inputs. No genuinely new physical entities are introduced; the similarity function s is an invented construction without independent validation.

free parameters (3)
  • tol (LCS acceptance tolerance) = 1.0 (swept over 0.9-1.3)
    Section III-E. Controls the similarity-adjusted threshold for accepting frame-phoneme matches; chosen by alignment accuracy and coverage on validation data.
  • lambda (loss weighting) = 0.5
    Section II-B.2. Weights frame-level cross-entropy against masked CTC; set by hand without an ablation table.
  • Gaussian edge-attenuation width for cost targets = unreported
    Section II-A.b. The paper smooths ground-truth cost labels near phoneme boundaries but does not specify the temporal sigma; this shapes all cost-model supervision.
assumptions (5)
  • domain assumption MFA plus manual corrections on a 30% VCTK split provide accurate frame-phoneme supervision for the cost matrix model.
    Section II-A.b. If the alignment labels are wrong, the masks inherit errors before the recognizer is trained.
  • ad hoc to paper The hand-constructed articulatory similarity function s is a valid measure of phoneme confusability.
    Section II-A.b. It defines cost targets, the LCS threshold, and the WPER metric and is never validated against behavioral or acoustic confusions.
  • domain assumption The cost matrix model trained on VCTK generalizes to LibriSpeech, PPA, and synthetic LLM-dys audio without adaptation.
    Section III-A and III-D. The masks used in training are computed by an aligner that never sees these corpora; cross-domain transfer is load-bearing.
  • ad hoc to paper Algorithm 1's modified LCS recurrence computes the intended partial alignment.
    Section II-A.c. The pseudocode as printed stores out-of-range pairs and uses an unusual horizontal propagation, so the published recurrence must be checked against the released code.
  • domain assumption VITS-synthesized dysfluencies from language-model prompts are a valid proxy for real dysfluent speech.
    Section III-A. LLM-dys is used as a benchmark for non-fluent recognition, but synthetic insertions may not match pathological acoustics.
invented entities (1)
  • Phonetic similarity function s(pi,pj) over CMU phoneme pairs
    purpose: Used to define cost-matrix targets, the LCS acceptance threshold, and WPER evaluation weights.
    The paper assigns similarity from eight articulatory feature categories and softmax normalizes the values; it does not validate the scale against human phoneme confusion or acoustic measures, and it appears in both the training pipeline and the evaluation, which can inflate reported WPER gains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LCS-CTC: Leveraging Soft Alignments to Enhance Phonetic Transcription Robustness." pith.science (2026). https://pith.science/paper/YMTDAW7V

@misc{pith2026250803937,
  author       = {Pith},
  title        = {Pith review of: LCS-CTC: Leveraging Soft Alignments to Enhance Phonetic Transcription Robustness},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YMTDAW7V}},
  note         = {Machine review of arXiv:2508.03937}
}
read the original abstract

Phonetic speech transcription is crucial for fine-grained linguistic analysis and downstream speech applications. While Connectionist Temporal Classification (CTC) is a widely used approach for such tasks due to its efficiency, it often falls short in recognition performance, especially under unclear and nonfluent speech. In this work, we propose LCS-CTC, a two-stage framework for phoneme-level speech recognition that combines a similarity-aware local alignment algorithm with a constrained CTC training objective. By predicting fine-grained frame-phoneme cost matrices and applying a modified Longest Common Subsequence (LCS) algorithm, our method identifies high-confidence alignment zones which are used to constrain the CTC decoding path space, thereby reducing overfitting and improving generalization ability, which enables both robust recognition and text-free forced alignment. Experiments on both LibriSpeech and PPA demonstrate that LCS-CTC consistently outperforms vanilla CTC baselines, suggesting its potential to unify phoneme modeling across fluent and non-fluent speech.

Figures

Figures reproduced from arXiv: 2508.03937 by the authors.

Figure 1
Figure 1. (a) shows the structure of cost matrix learning model. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of LCS-CTC framework: (a) Vanilla CTC allows all valid alignment paths that collapse to the target sequence, which can result in peaky distributions. (b) Our LCS-CTC introduces alignment constraints from the LCS-based aligner (black nodes), which eliminate implausible paths (red crosses) by fixing ”high-confidence” frame-phoneme matches. (c) The overall training pipeline of LCS-CTC. The LCS aligner uses gro… view at source ↗
Figure 3
Figure 3. Visualization of token emission probabilities for vanilla CTC and our proposed LCS-CTC on three randomly selected [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Effect of tol on alignment accuracy and constrained region ratio. We perform an ablation study on the tolerance parameter tol in the LCS alignment algorithm, which controls the threshold for accepting frame-phoneme matches based on predicted cost and phoneme similarity…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

  1. AV-EMO-Reasoning: Benchmarking Emotional Reasoning Capabilities in Omni-modal LLMS with Audio-visual Cues

    cs.MM 2025-10 conditional novelty 6.0 of 10

    Current omni-modal LLMs underperform on audio-visual emotional reasoning, and automatic scores diverge from human perceptual judgments; AV-EMO-Reasoning provides a benchmark to measure this.

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

    cs.AI 2025-08 conditional novelty 4.0 of 10

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

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

    cs.AI 2025-08 reject novelty 4.0 of 10

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

Reference graph

Works this paper leans on

51 extracted references · 41 canonical work pages · cited by 3 Pith papers

  1. [24]

    Bayes risk ctc: Controllable ctc alignment in sequence-to-sequence tasks,

    J. Tian, B. Yan, J. Yu, C. Weng, D. Yu, and S. Watanabe, “Bayes risk ctc: Controllable ctc alignment in sequence-to-sequence tasks,” ICLR, 2023

    work page 2023

  2. [26]

    Cr-ctc: Consistency regularization on ctc for improved speech recognition,

    Z. Yao, W. Kang, X. Yang, F. Kuang, L. Guo, H. Zhu, Z. Jin, Z. Li, L. Lin, and D. Povey, “Cr-ctc: Consistency regularization on ctc for improved speech recognition,” ICLR, 2025

    work page 2025

  3. [1]

    Robust speech recognition via large-scale weak super- vision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak super- vision,” in International Conference on Machine Learning . PMLR, 2023, pp. 28 492–28 518

    2023

  4. [2]

    Google usm: Scaling automatic speech recognition beyond 100 languages,

    Y . Zhang, W. Han, J. Qin, Y . Wang, A. Bapna, Z. Chen, N. Chen, B. Li, V . Axelrod, G. Wang et al. , “Google usm: Scaling automatic speech recognition beyond 100 languages,” arXiv preprint arXiv:2303.01037 , 2023

    arXiv 2023

  5. [3]

    Scaling speech technology to 1,000+ languages,

    V . Pratap, A. Tjandra, B. Shi, P. Tomasello, A. Babu, S. Kundu, A. Elkahky, Z. Ni, A. Vyas, M. Fazel-Zarandi et al. , “Scaling speech technology to 1,000+ languages,” Journal of Machine Learning Re- search, vol. 25, no. 97, pp. 1–52, 2024

    2024

  6. [4]

    Less is more: Accurate speech recognition & translation without web- scale data,

    K. C. Puvvada, P. ˙Zelasko, H. Huang, O. Hrinchuk, N. R. Koluguri, K. Dhawan, S. Majumdar, E. Rastorgueva, Z. Chen, V . Lavrukhin et al., “Less is more: Accurate speech recognition & translation without web- scale data,” arXiv preprint arXiv:2406.19674 , 2024

    arXiv 2024

  7. [5]

    Parakeet-rnnt-1.1b: A fastconformer transducer model for english asr,

    NVIDIA NeMo and Suno.ai, “Parakeet-rnnt-1.1b: A fastconformer transducer model for english asr,” https://huggingface.co/nvidia/ parakeet-rnnt-1.1b, 2025

    work page 2025

  8. [6]

    Automatic pronunciation assessment - a review,

    Y . El Kheir, A. Ali, and S. A. Chowdhury, “Automatic pronunciation assessment - a review,” in Findings of the Association for Computational Linguistics: EMNLP 2023 , pp. 8304–8324

    work page 2023

Show all 51 references
  1. [7]

    Automatic pronunci- ation error detection: an acoustic-phonetic approach,

    K. Truong, A. Neri, C. Cucchiarini, and H. Strik, “Automatic pronunci- ation error detection: an acoustic-phonetic approach,” in InSTIL/ICALL 2004 Symposium on Computer Assisted Learning , 2004, p. paper 032

    2004

  2. [8]

    Language-independent methods for computer-assisted pronunciation training,

    A. Lee et al. , “Language-independent methods for computer-assisted pronunciation training,” Ph.D. dissertation, Massachusetts Institute of Technology, 2016

    2016

  3. [9]

    Connected speech fluency in poststroke and progressive aphasia: A scoping review of quantitative approaches and features,

    C. Cordella, L. Di Filippo, V . B. Kolachalama, and S. Kiran, “Connected speech fluency in poststroke and progressive aphasia: A scoping review of quantitative approaches and features,” American Journal of Speech- Language Pathology, vol. 33, no. 4, pp. 2083–2120, 2024

  4. [10]

    Automatically measuring speech fluency in people with apha- sia: First achievements using read-speech data,

    L. Fontan, T. Prince, A. Nowakowska, H. Sahraoui, and S. Martinez- Ferreiro, “Automatically measuring speech fluency in people with apha- sia: First achievements using read-speech data,” Aphasiology, vol. 38, no. 5, pp. 939–956, 2024

    2024

  5. [11]

    The fluency dimension in aphasia,

    J. K. Gordon, “The fluency dimension in aphasia,” Aphasiology, vol. 12, no. 7-8, pp. 673–688, 1998

    1998

  6. [12]

    How do clinicians judge fluency in aphasia?

    J. K. Gordon and S. Clough, “How do clinicians judge fluency in aphasia?” Journal of Speech, Language, and Hearing Research , vol. 65, no. 4, pp. 1521–1542, 2022

    2022

  7. [13]

    Evaluating fluency in aphasia: Fluency scales, trichotomous judgements, or machine learning,

    J. Metu, V . Kotha, and A. E. Hillis, “Evaluating fluency in aphasia: Fluency scales, trichotomous judgements, or machine learning,” Apha- siology, pp. 1–13, 2023

    2023

  8. [14]

    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 Information Processing Systems, vol. 37, 2024

    2024

  9. [15]

    Automatic detection of articulatory-based disfluencies in primary pro- gressive aphasia,

    J. Lian, X. Zhou, C. Guo, Z. Ye, Z. Ezzes, J. V onk, B. Morin, D. Baquirin, Z. Mille, M. L. G. Tempini, and G. K. Anumanchipalli, “Automatic detection of articulatory-based disfluencies in primary pro- gressive aphasia,” IEEE JSTSP, 2025

    2025

  10. [16]

    Yolo-stutter: End-to-end region-wise speech dysfluency 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. Anumanchipalli, “Yolo-stutter: End-to-end region-wise speech dysfluency detection,” in Interspeech 2024, 2024, pp. 937–941

    2024

  11. [17]

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

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

    2024

  12. [18]

    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

  13. [19]

    Leveraging allophony in self-supervised speech models for atypical pronunciation assessment,

    K. Choi, E. Yeo, K. Chang, S. Watanabe, and D. Mortensen, “Leveraging allophony in self-supervised speech models for atypical pronunciation assessment,” in NAACL, 2025

    2025

  14. [20]

    Connection- ist temporal classification: labelling unsegmented sequence data with recurrent neural networks,

    A. Graves, S. Fern ´andez, F. Gomez, and J. Schmidhuber, “Connection- ist temporal classification: labelling unsegmented sequence data with recurrent neural networks,” in Proceedings of the 23rd international conference on Machine learning , 2006, pp. 369–376

    2006

  15. [21]

    Attention-based models for speech recognition,

    J. K. Chorowski, D. Bahdanau, D. Serdyuk, K. Cho, and Y . Bengio, “Attention-based models for speech recognition,” Advances in neural information processing systems , vol. 28, 2015

    2015

  16. [22]

    Why does ctc result in peaky behavior?

    A. Zeyer, R. Schl ¨uter, and H. Ney, “Why does ctc result in peaky behavior?” arXiv preprint arXiv:2105.14849 , 2021

    2021 arXiv

  17. [23]

    Careless whisper: Speech-to-text hallucination harms,

    A. Koenecke, A. S. G. Choi, K. X. Mei, H. Schellmann, and M. Sloane, “Careless whisper: Speech-to-text hallucination harms,” in Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Trans- parency, 2024, pp. 1672–1681

    2024

  18. [25]

    Less peaky and more accu- rate ctc forced alignment by label priors,

    R. Huang, X. Zhang, Z. Ni, L. Sun, M. Hira, J. Hwang, V . Manohar, V . Pratap, M. Wiesner, S. Watanabe et al., “Less peaky and more accu- rate ctc forced alignment by label priors,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (I...

    2024

  19. [27]

    Unconstrained dysfluency modeling for dysfluent speech transcription and detection,

    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 detection,” in2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) . IE...

    2023

  20. [28]

    Towards hierarchical spoken language disfluency modeling,

    J. Lian and G. 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

    2024

  21. [29]

    Seamless dysfluent speech text alignment for disordered speech analysis,

    Z. Ye, J. Lian, X. Zhou, J. Zhang, H. Li, S. Li, C. Guo, A. Das, P. Park, Z. Ezzes, J. V onk, B. Morin, R. Bogley, L. Wauters, Z. Miller, M. Gorno-Tempini, and G. Anumanchipalli, “Seamless dysfluent speech text alignment for disordered speech analysis,” Interspeech, 2025

    2025

  22. [30]

    Algorithms for the longest common subsequence problem,

    D. S. Hirschberg, “Algorithms for the longest common subsequence problem,” Journal of the ACM (JACM) , vol. 24, no. 4, pp. 664–675, 1977

    1977

  23. [31]

    Librispeech: An asr corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Librispeech: An asr corpus based on public domain audio books,” in 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015, pp. 5206–5210

    2015

  24. [32]

    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

  25. [33]

    Analysis and evaluation of synthetic data gener- ation in speech dysfluency detection,

    J. Zhang, X. Zhou, J. Lian, S. Li, W. Li, Z. Ezzes, R. Bogley, L. Wauters, Z. Miller, J. V onk, B. Morin, M. Gorno-Tempini, and G. Anumanchipalli, “Analysis and evaluation of synthetic data gener- ation in speech dysfluency detection,” Interspeech, 2025

    2025

  26. [34]

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

    J. Yamagishi, C. Veaux, K. MacDonald et al., “Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit (version 0.92),” University of Edinburgh. The Centre for Speech Technology Research (CSTR), 2019

    2019

  27. [35]

    Montreal forced aligner: Trainable text-speech alignment using kaldi,

    M. McAuliffe, M. Socolof, S. Mihuc, M. Wagner, and M. Sonderegger, “Montreal forced aligner: Trainable text-speech alignment using kaldi,” in Interspeech 2017, 2017, pp. 498–502

    2017

  28. [36]

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

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

    2020

  29. [37]

    Conformer: Convolution-augmented transformer for speech recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu et al., “Conformer: Convolution-augmented transformer for speech recognition,” Interspeech, 2020

    2020

  30. [38]

    Model card addendum: Claude 3.5 haiku and upgraded claude 3.5 sonnet,

    Anthropic, “Model card addendum: Claude 3.5 haiku and upgraded claude 3.5 sonnet,” 2024. [Online]. Available: https://www.anthropic. com

    2024

  31. [39]

    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,” in International Conference on Machine Learning . PMLR, 2021, pp. 5530–5540

    2021

  32. [40]

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

    C. Guo, J. Lian, X. Zhou, J. Zhang, S. Li, Z. Ye, H. J. Park, A. Das, Z. Ezzes, J. V onk, B. Morin, R. Bogley, L. Wauters, Z. Miller, M. Gorno- Tempini, and G. Anumanchipalli, “Dysfluent wfst: A framework for zero-shot speech dysfluency transcription and detection,” Interspeech, 2025

    2025

  33. [41]

    Coding speech through vocal tract kinematics,

    C. J. Cho, P. Wu, T. S. Prabhune, D. Agarwal, and G. K. Anumanchipalli, “Coding speech through vocal tract kinematics,” IEEE Journal of Selected Topics in Signal Processing , vol. 18, no. 8, p. 1427–1440, Dec. 2024. [Online]. Available: http://dx.doi.org/10.1109/ JSTSP.2024.3497655

    2024

  34. [42]

    Hubert: Self-supervised speech representation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “Hubert: Self-supervised speech representation learning by masked prediction of hidden units,” IEEE TASLP, 2021

    2021

  35. [43]

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

    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 processing,” IEEE JSTSP, 2022

    2022

  36. [44]

    Phone-to-audio alignment with- out text: A semi-supervised approach,

    J. Zhu, C. Zhang, and D. Jurgens, “Phone-to-audio alignment with- out text: A semi-supervised approach,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 8167–8171

    2022

  37. [45]

    Neufa: Neural network based end-to-end forced alignment with bidirectional attention mechanism,

    J. Li, Y . Meng, Z. Wu, H. Meng, Q. Tian, Y . Wang, and Y . Wang, “Neufa: Neural network based end-to-end forced alignment with bidirectional attention mechanism,” in ICASSP 2022-2022 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2...

    2022

  38. [46]

    Joint speech and text machine translation for up to 100 languages,

    “Joint speech and text machine translation for up to 100 languages,” Nature, vol. 637, no. 8046, pp. 587–593, 2025

    2025

  39. [47]

    Towards accurate phonetic error detection through phoneme similarity modeling,

    X. Zhou, J. Lian, C. J. Cho, T. Prabhune, S. Li, W. Li, R. Ortiz, Z. Ezzes, J. V onk, B. Morin, R. Bogley, L. Wauters, Z. Miller, M. Gorno-Tempini, and G. Anumanchipalli, “Towards accurate phonetic error detection through phoneme similarity modeling,” Interspeech, 2025

    2025

  40. [48]

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

    S. Li, C. Guo, J. Lian, C. J. Cho, W. Zhao, X. Zhou, D. Zhou, S. Wang, G. Wang, J. Yang et al., “K-function: Joint pronunciation transcription and feedback for evaluating kids language function,” arXiv preprint arXiv:2507.03043, 2025

    2025

  41. [49]

    Coding speech through vocal tract kinematics,

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

    2025

  42. [50]

    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,” in Proc. Interspeech 2022 , 2022, pp. 4686–4690

    2022

  43. [51]

    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 6, 2026 · model on record in the stance chip above.