Pith. sign in

REVIEW 4 major objections 6 minor 38 references

Leveraging LLM for Stuttering Speech: A Unified Architecture Bridging Recognition and Event Detection

T0 review · 4 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 unified LLM system pairs transcription of stuttered Mandarin speech with stuttering-event detection, reporting CER 5.45% and SED F1 73.63% on AS-70.

desk verdict A plausible LLM+CTC multi-task system for stuttered Mandarin ASR/SED, but the headline numbers hinge on a missing CTC-only baseline and an unclear scoring normalization. read the letter →

arxiv 2505.22005 v1 pith:GWZZHCJS submitted 2025-05-28 cs.SD eess.AS

Shangkun Huang , Jing Deng , Jintao Kang , Rong Zheng This is my paper
classification cs.SDeess.AS
keywords Mandarinstutteredspeechautomaticrecognitionstutteringeventdetectionmulti-tasklearninglargelanguagemodelCTChypothesisembeddingscontrastivefocalloss
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

This paper tries to establish that a single LLM-driven system can handle both halves of stuttered-speech processing: transcribing audio into fluent text and detecting where stuttering events occur. The authors design a multi-task architecture that feeds a small instruction-tuned LLM two extra signals alongside the audio—a text hypothesis produced by a CTC decoder and a sentence-level stuttering embedding from a detection branch—plus a prompt asking for a fluent transcription. On the AS-70 Mandarin stuttering dataset, they report a character error rate of 5.45% and an average stuttering-event-detection F1 of 73.63%, relative improvements of 37.71% and 46.58% over the earlier systems they compare against. If the attribution holds, the same model could serve as a technical backbone for speech-rehabilitation tools that both listen and annotate disfluencies.

What carries the argument

The load-bearing mechanism is the multimodal input template assembled for the LLM: audio embeddings from a shared speech encoder, a hypothesis embedding from a connectionist temporal classification (CTC) decoder, a sentence-level stuttering embedding from the detection branch, a prompt embedding, and, during training, the target transcription. The CTC hypothesis supplies a global text prior that steers generation away from stutter-induced repetitions; the stuttering embedding tells the LLM which acoustic content is disfluent. The LLM is adapted with LoRA, and the total loss couples the LLM cross-entropy loss with CTC loss, multi-label supervised contrastive loss, and focal loss, the last two addressing the long-tailed event distribution.

What would settle it

Run the AS-70 test set with the shared encoder and CTC decoder alone, without the LLM or SED branch, under the same label preprocessing that yields the 5.45% CER; then repeat with stuttering markers and disfluent characters kept in the labels. If the CTC-only CER is close to 5.45%, or if the preprocessing difference explains most of the gap to the baseline numbers, the central attribution is refuted.

Watch

Extended reading notes

Core claim

The central claim is that stuttering-induced recognition failures are best addressed by treating recognition and stuttering-event detection as one task rather than two. The paper argues that feeding the LLM a CTC-generated text hypothesis gives it a global-text prior that suppresses the repetitive hallucination patterns stuttered audio tends to provoke, while feeding it a sentence-level stuttering embedding tells it which acoustic content is disfluent, letting it generate a fluent transcription. With contrastive learning and focal loss on the detection branch to handle the long-tailed event distribution, the combined system reaches a CER of 5.45% and an average SED F1-score of 73.63% on AS-70, which the authors report as a 37.71% relative CER reduction and a 46.58% relative F1 improvement over previous work.

Load-bearing premise

The load-bearing premise is that the gains come from the LLM and the added signal streams rather than from the CTC branch alone or from label preprocessing; the paper does not provide the CTC-only error rate or a matched preprocessing baseline, so this premise is unverified.

Editorial extensions

If this is right

  • Adding the CTC hypothesis embedding to the LLM suppresses stutter-induced repetitive hallucinations, making recognition more robust than speech-embedding-only LLM input.
  • Adding the stuttering embedding further lowers CER across all severity groups and scenarios reported, including severe stuttering and conversation speech.
  • Multi-label supervised contrastive learning together with focal loss raises detection F1 especially for the rare Block (/b) and Word/Phrase Repetition ([]) event types.
  • The joint ASR-SED architecture offers an end-to-end path for stuttered-speech processing tasks such as speech rehabilitation.
  • The same architecture is positioned to extend to multilingual or low-resource speech-disorder scenarios.

Reading between the lines

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

  • A decisive test of the causal story would compare the frozen shared encoder plus CTC decoder alone (no LLM, no SED branch) under the same label preprocessing; without that number, the LLM's marginal contribution over the CTC hypothesis is not established.
  • Because the paper strips stuttering markers and disfluent characters from the labels before computing CER, the metric measures transcription of intended fluent content; a standardized preprocessing protocol would be needed to compare fairly against systems that keep stutter annotations.
  • The dynamic-embedding pattern could transfer to other disfluent or atypical speech (aphasia, heavy accent) and to streaming therapy feedback, though the sentence-level stuttering embedding would have to be adapted to short windows.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes an LLM-driven multi-task framework for Mandarin stuttered speech that jointly performs automatic speech recognition (ASR) and stuttering event detection (SED). A shared encoder feeds both a CTC decoder and an SED branch; the CTC decoder's 1-best hypothesis and the SED branch's sentence-level stutter embedding are tokenized and fed, together with speech, prompt, and text embeddings, into Qwen2.5-3B-Instruct via LoRA. The training loss combines LLM cross-entropy, CTC, focal loss, and multi-label supervised contrastive loss. On the AS-70 dataset the authors report a CER of 5.45% (a 37.71% relative reduction) and an average SED F1 of 73.63% (a 46.58% relative improvement), and they present ablations showing the contribution of the CTC hypothesis embedding and the stutter embedding to the LLM inputs.

Significance. If the reported results hold, the framework would be a useful contribution to stuttered-speech processing: it integrates CTC hypotheses into an LLM to suppress repetitive hallucinations, incorporates sentence-level stutter embeddings from an SED branch, and combines contrastive learning with focal loss to address long-tailed stuttering events. The evaluation uses an external public dataset (AS-70) and compares against published baselines, which is a strength. However, the central attribution claim is not yet supported because no CTC-only baseline is reported, the label preprocessing used for CER may not match the baselines, and no uncertainty estimates or test-set statistics are given. The architecture is plausible, but the evidence provided is insufficient to establish that the LLM and stutter embeddings, rather than the CTC hypothesis itself, produce the headline gains.

major comments (4)
  1. [Section 2.2, Section 2.4, Table 1] The paper never reports the CER of the CTC decoder's own 1-best output. Since the CTC hypothesis is tokenized and fed to the LLM in all variants except the first row of Table 1, the drop from 10.13% to 5.63% could be caused by the CTC text alone, with the LLM largely copying or lightly editing it. To support the claim that the LLM and stutter embeddings 'suppress stuttering-induced repetitive hallucinations' and are responsible for the 37.71% relative reduction, the authors must report the greedy (and optionally beam-searched) CTC decoder's CER on the same normalized references, and ideally a variant with the LLM but without the CTC hypothesis.
  2. [Section 3.3.1, Table 1] The manuscript states that labels were preprocessed by removing stuttering event markers and disfluent characters, but it does not state whether the published baseline numbers (Whisper-large-v2 8.75, Hubert-large 7.25, Conformer 6.24) were computed on the same normalized references. If those baselines were not re-evaluated with the same preprocessing, the reported 37.71% relative reduction is not a like-for-like comparison. The authors should specify the exact normalization, state whether the baseline CERs are taken from the original AS-70 paper or recomputed, and report the CER both with and without normalization.
  3. [Section 3.1, Section 3.3] No train/validation/test split description, number of runs, or error bars are provided. With a single training run and hand-tuned hyperparameters (α, β, μ, δ, τ, LoRA rank, etc.), the reported differences such as 5.63% vs. 5.45% CER and the SED F1 improvements cannot be assessed for statistical significance. The authors should describe the AS-70 split, report results over multiple seeds or runs, and provide uncertainty measures, or at least clearly state how many runs were performed.
  4. [Section 3.3.2, Table 2] The abstract and Section 3.3.3 describe a '46.58% relative improvement' in SED F1, but this figure compares the proposed model (73.63) against StutterNet (50.23), not against the strongest prior baseline Wav2vec2.0 (68.21). Against Wav2vec2.0 the relative gain is only about 7.9%. The claim of state-of-the-art performance may still be true, but the headline relative improvement should be computed against the strongest prior baseline, and the choice of baseline for the relative improvement should be explicitly justified.
minor comments (6)
  1. [Section 3.1] The training configuration says 'only the CTC decoder parameters were updated, while the encoder parameters remained frozen,' but it is unclear whether the entire SenseVoice-Small encoder and CTC decoder are frozen or only the CTC decoder; this should be clarified because it affects the interpretation of Table 1.
  2. [Section 2.3 and Section 2.4, Eq. (8)] The symbol H_SED used in Eq. (8) is described as 'the intermediate output of the SED classifier,' but Section 2.3 defines h1 and h2 without explicitly naming H_SED; please align the notation with the classifier architecture.
  3. [Section 2.4, Eq. (9)] The 'Template' function is not formally defined; the reader is left to infer the ordering and concatenation of the five embeddings from the prose and Figure 1. A precise definition of the template, including token boundaries, would improve reproducibility.
  4. [Figure 1 and Section 2.4] The figure labels (a) and (b) are mentioned in the caption, but the main text does not explicitly reference panel (b) when discussing the SED branch; adding a reference would help the reader.
  5. [Throughout] The abstract uses past tense ('This paper proposed', 'We proposed') where present tense is conventional, and there are minor typographical issues such as a space in 'SenseV oiceSmall' and inconsistent punctuation around equations.
  6. [Section 3.3.1] The phrase 'disfluent characters' is not defined; it is unclear whether it refers to stuttering event markers, repeated phonemes, fillers, or some other annotation. Please define the normalization explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the reported CER and SED F1 are empirical results against external AS-70 baselines, not consequences of the model's definitions; the missing standalone CTC baseline is an attribution gap, not a definitional reduction.

full rationale

The paper's derivation is not circular. The final CER (5.45%) and SED F1 (73.63%) are obtained by evaluating the trained system on the AS-70 test set, not by algebraic identity with any input. The CTC hypothesis embedding (Eq. 3) and stutter embedding (Eq. 8) are conditioned inputs to the LLM, but the LLM output is not defined as equal to them; it is a separate autoregressive prediction optimized by cross-entropy (Eq. 10). The ablations in Table 1 provide internal evidence for the contribution of these inputs, and the comparison tables use external baselines (Whisper-large-v2, Hubert-large, Conformer for ASR; StutterNet, ConvLSTM, Wav2vec2.0 for SED), so the headline improvements are not relative to a self-defined reference. The hand-set loss weights (alpha=[0.3,0.3,0.2,0.1,0.1], delta=0.3, beta=0.3, mu=0.1) are tuned hyperparameters, not fitted parameters renamed as predictions. The label preprocessing in Section 3.3.1 (removing stuttering markers and disfluent characters) is an evaluation-protocol choice; it could affect comparability with published baselines, but it is not a hidden equivalence between input and output. The only self-citations ([3] and [4]) are background prior work and are not load-bearing; no uniqueness theorem or methodological premise is imported from them. The strongest concern—that the standalone CTC 1-best CER is never reported, so the 37.71% relative reduction cannot be attributed to the LLM rather than to the CTC hypothesis—is a valid evidence gap and a correctness risk, but it is not circularity: the paper nowhere defines the final CER in terms of the CTC hypothesis or reduces Eq. (10) to Eq. (2).

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim relies on the AS-70 dataset and pretrained models; no new physical or formal entities are introduced. The main risk is the unverified evaluation protocol and hand-tuned loss weights.

free parameters (6)
  • Focal loss alpha vector = [0.3, 0.3, 0.2, 0.1, 0.1]
    Balancing factors for five stuttering event classes, chosen by hand for the AS-70 class distribution (Section 3.1).
  • CTC loss weight beta = 0.3
    Weight for CTC loss in the total loss; no sensitivity analysis is shown (Section 3.1).
  • SED loss weight mu = 0.1
    Weight for SED loss; no sensitivity analysis is shown (Section 3.1).
  • Contrastive loss weight delta = 0.3
    Weight for supervised contrastive loss; no sensitivity analysis is shown (Section 3.1).
  • LoRA parameters = rank=8, alpha=16, dropout=0.1
    LoRA configuration chosen without ablation; affects LLM adaptation (Section 3.1).
  • Contrastive temperature tau = not reported
    Temperature in the supervised contrastive loss (Eq. 7); value omitted, preventing exact reproduction.
assumptions (4)
  • domain assumption AS-70 dataset labels and splits are correct and match the evaluation protocol of prior baselines
    All results depend on AS-70 annotation quality and on the official test set; the paper does not describe the split (Section 3.2).
  • domain assumption Removing stuttering markers and disfluent characters from reference labels is the same preprocessing used for the compared baselines
    Section 3.3.1 states this preprocessing; if baselines did not do it, the CER comparison is unfair.
  • domain assumption Pretrained Qwen2.5-3B and SenseVoice models provide useful priors for Mandarin and stuttered speech
    The method inherits capabilities from pretrained encoders and an LLM without training them from scratch (Section 3.1).
  • domain assumption Stuttering event detection can be performed from a sentence-level pooled embedding
    Section 2.3 averages the encoder output over time; if stutter events are local, the pooling assumption may weaken detection.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Leveraging LLM for Stuttering Speech: A Unified Architecture Bridging Recognition and Event Detection." pith.science (2026). https://pith.science/paper/GWZZHCJS

@misc{pith2026250522005,
  author       = {Pith},
  title        = {Pith review of: Leveraging LLM for Stuttering Speech: A Unified Architecture Bridging Recognition and Event Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GWZZHCJS}},
  note         = {Machine review of arXiv:2505.22005}
}
read the original abstract

The performance bottleneck of Automatic Speech Recognition (ASR) in stuttering speech scenarios has limited its applicability in domains such as speech rehabilitation. This paper proposed an LLM-driven ASR-SED multi-task learning framework that jointly optimized the ASR and Stuttering Event Detection (SED) tasks. We proposed a dynamic interaction mechanism where the ASR branch leveraged CTC-generated soft prompts to assist LLM context modeling, while the SED branch output stutter embeddings to enhance LLM comprehension of stuttered speech. We incorporated contrastive learning to strengthen the discriminative power of stuttering acoustic features and applied Focal Loss to mitigate the long-tailed distribution in stuttering event categories. Evaluations on the AS-70 Mandarin stuttering dataset demonstrated that our framework reduced the ASR character error rate (CER) to 5.45% (-37.71% relative reduction) and achieved an average SED F1-score of 73.63% (+46.58% relative improvement).

Figures

Figures reproduced from arXiv: 2505.22005 by the authors.

Figure 1
Figure 1. (a) LLM-driven ASR-SED multi-task learning framework; (b) SED branch architecture 2. Proposed Method 2.1. Framework Overview As shown in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 22 canonical work pages

  1. [1]

    Leveraging LLM for Stuttering Speech: A Unified Architecture Bridging Recognition and Event Detection

    Introduction Stuttering is a multifaceted neurodevelopmental disorder affect- ing over 80 million people worldwide[1]. Common manifesta- tions include word, phrase, or sound repetitions, silent pauses (blocking), elongation of specific syllables or words (prolonga- tion), and interjections. Stuttering impedes daily communica- tion for people who stutter, ...

    work page Pith review arXiv 2024

  2. [2]

    Framework Overview As shown in Figure 1 (a), the proposed LLM-driven ASR- SED multi-task learning framework comprised two parallel branches: the ASR branch and the SED branch

    Proposed Method 2.1. Framework Overview As shown in Figure 1 (a), the proposed LLM-driven ASR- SED multi-task learning framework comprised two parallel branches: the ASR branch and the SED branch. The frame- work included six core components: the LLM, the LLM tok- enizer, the shared encoder, the Connectionist Temporal Clas- sification (CTC)[26] decoder, t...

    work page

  3. [3]

    <|im_start|>user<S><D><C ><P><|im_end|><|im_start|>assistant< T><|im_end|>|

    Experiments 3.1. Experimental Configuration For the shared encoder, we employed the SenseV oice-Small[29] model1, which included a SAN-M-based encoder, task-specific embeddings, and a CTC loss function. This non-autoregressive architecture provided efficient low-latency multilingual speech processing. We selected Qwen2.5-3B-Instruct 2, a model pre- optimi...

    work page

  4. [4]

    Conclusion This paper presents an LLM-driven ASR-SED multi-task learn- ing framework that achieves dual breakthroughs in recognition and detection performance in Mandarin stuttered speech sce- narios through dynamic interaction mechanisms and hybrid op- timization strategies. Experiments demonstrate that the frame- work, by integrating CTC hypothesis embe...

    work page

  5. [5]

    Acknowledgements This work is supported by National Engineering Research Cen- ter of Multi-dimensional Identification and Trusted Authentica- tion Technology (No.IDNERC202403)

    work page

  6. [6]

    “the world is designed for fluent people

    S. Wu, ““the world is designed for fluent people”: Benefits and challenges of videoconferencing technologies for people who stutter,” in Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, 2023, pp. 1–17

    work page 2023

  7. [7]

    Microaggression and the adult stuttering experience,

    G. A. Coalson, A. Crawford, S. B. Treleaven, C. T. Byrd, L. Davis, L. Dang, J. Edgerly, and A. Turk, “Microaggression and the adult stuttering experience,” Journal of Communication Disor- ders, vol. 95, p. 106180, 2022

    work page 2022

  8. [8]

    Enhanced asr for stuttering speech: Combining adversarial and signal-based data augmentation,

    S. Huang, D. Zhang, J. Deng, and R. Zheng, “Enhanced asr for stuttering speech: Combining adversarial and signal-based data augmentation,” in 2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 393–400

    work page 2024

Show all 38 references
  1. [9]

    The fos- afer system for the chime-8 mmcsg challenge,

    S. Huang, D. Zhang, Y . Wang, J. Deng, and R. Zheng, “The fos- afer system for the chime-8 mmcsg challenge,” in CHiME Work- shop on Speech Processing in Everyday Environments, 2024

    2024

  2. [10]

    As-70: A mandarin stuttered speech dataset for automatic speech recognition and stuttering event detection,

    R. Gong, H. Xue, L. Wang, X. Xu, Q. Li, L. Xie, H. Bu, S. Wu, J. Zhou, Y . Qinet al., “As-70: A mandarin stuttered speech dataset for automatic speech recognition and stuttering event detection,” arXiv preprint arXiv:2406.07256, 2024

    2024 arXiv

  3. [11]

    Findings of the 2024 mandarin stuttering event detection and automatic speech recognition chal- lenge,

    H. Xue, R. Gong, M. Shao, X. Xu, L. Wang, L. Xie, H. Bu, J. Zhou, Y . Qin, J. Du et al. , “Findings of the 2024 mandarin stuttering event detection and automatic speech recognition chal- lenge,” in 2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 385–392

    2024

  4. [12]

    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

  5. [13]

    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

  6. [14]

    From user percep- tions to technical improvement: Enabling people who stutter to better use speech recognition,

    C. Lea, Z. Huang, J. Narain, L. Tooley, D. Yee, D. T. Tran, P. Georgiou, J. P. Bigham, and L. Findlater, “From user percep- tions to technical improvement: Enabling people who stutter to better use speech recognition,” in Proceedings of the 2023 CHI Conference on Human Factors...

    2023

  7. [15]

    Anal- ysis and tuning of a voice assistant system for dysfluent speech,

    V . Mitra, Z. Huang, C. Lea, L. Tooley, S. Wu, D. Botten, A. Palekar, S. Thelapurath, P. Georgiou, S. Kajarekaret al., “Anal- ysis and tuning of a voice assistant system for dysfluent speech,” arXiv preprint arXiv:2106.11759, 2021

    2021 arXiv

  8. [16]

    Stutter-tts: Con- trolled synthesis and improved recognition of stuttered speech,

    X. Zhang, I. Vall ´es-P´erez, A. Stolcke, C. Yu, J. Droppo, O. Shon- ibare, R. Barra-Chicote, and V . Ravichandran, “Stutter-tts: Con- trolled synthesis and improved recognition of stuttered speech,” arXiv preprint arXiv:2211.09731, 2022

    2022 arXiv

  9. [17]

    Stutter- net: Stuttering detection using time delay neural network,

    S. A. Sheikh, M. Sahidullah, F. Hirsch, and S. Ouni, “Stutter- net: Stuttering detection using time delay neural network,” in 2021 29th European Signal Processing Conference (EUSIPCO) . IEEE, 2021, pp. 426–430

    2021

  10. [18]

    Stuttering detec- tion using atrous convolutional neural networks,

    A.-K. Al-Banna, E. Edirisinghe, and H. Fang, “Stuttering detec- tion using atrous convolutional neural networks,” in2022 13th In- ternational Conference on Information and Communication Sys- tems (ICICS). IEEE, 2022, pp. 252–256

    2022

  11. [19]

    Fluentnet: end- to-end detection of speech disfluency with deep learning,

    T. Kourkounakis, A. Hajavi, and A. Etemad, “Fluentnet: end- to-end detection of speech disfluency with deep learning,” arXiv preprint arXiv:2009.11394, 2020

    2009 arXiv

  12. [20]

    An end-to-end stut- tering detection method based on conformer and bilstm,

    X. Liu, C. Xu, Y . Yang, L. Wang, and N. Yan, “An end-to-end stut- tering detection method based on conformer and bilstm,” arXiv preprint arXiv:2411.09479, 2024

    2024 arXiv

  13. [21]

    Momentum con- trast for unsupervised visual representation learning,

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum con- trast for unsupervised visual representation learning,” inProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 9729–9738

    2020

  14. [22]

    Cola: Weakly- supervised temporal action localization with snippet contrastive learning,

    C. Zhang, M. Cao, D. Yang, J. Chen, and Y . Zou, “Cola: Weakly- supervised temporal action localization with snippet contrastive learning,” in Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, 2021, pp. 16 010–16 019

    2021

  15. [23]

    Mmger: Multi-modal and multi-granularity generative error correction with llm for joint accent and speech recognition,

    B. Mu, Y . Li, Q. Shao, K. Wei, X. Wan, N. Zheng, H. Zhou, and L. Xie, “Mmger: Multi-modal and multi-granularity generative error correction with llm for joint accent and speech recognition,” arXiv preprint arXiv:2405.03152, 2024

    2024 arXiv

  16. [24]

    Speech recognition meets large language model: Benchmarking, models, and exploration,

    Z. Ma, G. Yang, Y . Yang, Z. Gao, J. Wang, Z. Du, F. Yu, Q. Chen, S. Zheng, S. Zhang et al., “Speech recognition meets large language model: Benchmarking, models, and exploration,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 23, 2025, pp. 24 ...

    2025

  17. [25]

    A transcrip- tion prompt-based efficient audio large language model for robust speech recognition,

    Y . Li, X. Wang, S. Cao, Y . Zhang, L. Ma, and L. Xie, “A transcrip- tion prompt-based efficient audio large language model for robust speech recognition,” arXiv preprint arXiv:2408.09491, 2024

    2024 arXiv

  18. [26]

    Lauragpt: Listen, attend, understand, and re- generate audio with gpt,

    Z. Du, J. Wang, Q. Chen, Y . Chu, Z. Gao, Z. Li, K. Hu, X. Zhou, J. Xu, Z. Ma et al., “Lauragpt: Listen, attend, understand, and re- generate audio with gpt,”arXiv preprint arXiv:2310.04673, 2023

    2023 arXiv

  19. [27]

    Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities,

    D. Zhang, S. Li, X. Zhang, J. Zhan, P. Wang, Y . Zhou, and X. Qiu, “Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities,” arXiv preprint arXiv:2305.11000, 2023

    2023 arXiv

  20. [28]

    Qwen-audio: Advancing universal audio understand- ing via unified large-scale audio-language models,

    Y . Chu, J. Xu, X. Zhou, Q. Yang, S. Zhang, Z. Yan, C. Zhou, and J. Zhou, “Qwen-audio: Advancing universal audio understand- ing via unified large-scale audio-language models,”arXiv preprint arXiv:2311.07919, 2023

    2023 arXiv

  21. [29]

    Salmonn: Towards generic hearing abilities for large language models,

    C. Tang, W. Yu, G. Sun, X. Chen, T. Tan, W. Li, L. Lu, Z. Ma, and C. Zhang, “Salmonn: Towards generic hearing abilities for large language models,” arXiv preprint arXiv:2310.13289, 2023

    2023 arXiv

  22. [30]

    Large language models are efficient learners of noise- robust speech recognition,

    Y . Hu, C. Chen, C.-H. H. Yang, R. Li, C. Zhang, P.-Y . Chen, and E. Chng, “Large language models are efficient learners of noise- robust speech recognition,” arXiv preprint arXiv:2401.10446 , 2024

    2024 arXiv

  23. [31]

    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 Proceedings of the 23rd international conference on Machine learning, 2006, pp. 369–376

    2006

  24. [32]

    Focal loss for dense object detection,

    T.-Y . Ross and G. Doll´ar, “Focal loss for dense object detection,” in proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2980–2988

    2017

  25. [33]

    Lora: Low-rank adaptation of large lan- guage models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large lan- guage models,” arXiv preprint arXiv:2106.09685, 2021

    2021 arXiv

  26. [34]

    Funaudiollm: V oice understanding and gen- eration foundation models for natural interaction between humans and llms,

    K. An, Q. Chen, C. Deng, Z. Du, C. Gao, Z. Gao, Y . Gu, T. He, H. Hu, K. Hu et al., “Funaudiollm: V oice understanding and gen- eration foundation models for natural interaction between humans and llms,” arXiv preprint arXiv:2407.04051, 2024

    2024 arXiv

  27. [35]

    Advanc- ing stuttering detection via data augmentation, class-balanced loss and multi-contextual deep learning,

    S. A. Sheikh, M. Sahidullah, F. Hirsch, and S. Ouni, “Advanc- ing stuttering detection via data augmentation, class-balanced loss and multi-contextual deep learning,” IEEE Journal of Biomedical and Health Informatics, vol. 27, no. 5, pp. 2553–2564, 2023

    2023

  28. [36]

    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,” arXiv preprint arXiv:2005.08100, 2020

    2005 arXiv

  29. [37]

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

    C. Lea, V . Mitra, A. Joshi, S. Kajarekar, and J. P. Bigham, “Sep-28k: A dataset for stuttering event detection from podcasts with people who stutter,” in ICASSP 2021-2021 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 6798–6802

    2021

  30. [38]

    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

Pith tools

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