Pith. sign in

REVIEW 4 major objections 5 minor 97 references

The Role of Disfluencies in Speech Translation

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Current speech translation systems delete disfluencies, and false starts and self-repairs cause most translation-quality loss, as shown by a new eight-language benchmark.

desk verdict Solid benchmark paper that gives speech translation a new disfluency axis, but the hand-picked 80-utterance sample undercuts the generality of the emotion claims, so treat those as suggestive rather than definitive. read the letter →

arxiv 2608.02138 v1 pith:ZCEVE67X submitted 2026-08-03 cs.CL

Maike Züfle , Maria Teleki , Fabian Retkowski , Vilém Zouhar , Oliver Grabner , Alexander Waibel , James Caverlee , Jan Niehues This is my paper
classification cs.CL
keywords speechdisfluenciesbenchmarkcleanedfalsefilledlanguagespauses
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

People often speak with "um's," repeated words, and false starts. Most translation systems are trained on tidy text and remove these disfluencies. This paper argues that cleaning them away is not free: hesitation and self-correction carry emotional and conversational signals. In a small human study, listeners rated written or audio versions of the same utterances and reported stronger negative emotions when disfluencies were stripped out; the effect was large in a few examples, though the overall statistical effects are modest.

To make the problem measurable, the authors built Uh-Mazing, a benchmark of 80 disfluent English telephone conversations from Switchboard, translated by human annotators into eight languages. Each translation exists in a disfluent and a fluent version, and disfluencies are marked. Then several model types, cascaded ASR plus translation, end-to-end systems, speech LLMs, and commercial chatbots, were asked to translate the same speech. Across systems, meaning was preserved better than style. When a disfluency was lost, it was almost always deleted rather than mistranslated. False starts and self-repairs, not fillers like "uh" or discourse markers, were the main source of the quality drop.

The paper also tests cheap fixes at decoding time. Beam search generally helped, and giving the model two disfluent examples before translating, in-context learning, improved the disfluency-sensitive chrF score without hurting meaning. These interventions do not close the gap but point to a practical path while training data remain scarce. The benchmark and code are planned for release through LDC.

Extended reading notes

Core claim

The paper's central assertion, from the abstract, is that "disfluencies carry meaning that gets lost when speech is cleaned up," and, from Sections 5.3 and 5.4, that "false starts and self-repairs, not filled pauses or discourse markers, drive most of the translation-quality loss," while "when models fail to preserve a disfluency, they typically omit it outright rather than mistranslating it" (deletion spans number 1,098 versus 446 disfluency mistranslations). If correct, disfluency preservation is a measurable, currently missing dimension of speech translation quality, concentrated in edited speech rather than fillers.

Load-bearing premise

Load-bearing premise: the 80 manually selected Switchboard utterances form a sufficiently representative sample of disfluent conversational speech for the multilingual benchmark and the emotion study. Section 2.1 says the utterances were "manually chosen to contain multiple disfluencies and to span a range of lengths (17-212 words)", and Section 3.1 translates those same 80 into eight languages. If this curated sample is not representative of spontaneous speech, all cross-lingual quality comparisons, the false-start ablation, and the emotional-shift results inherit that bias. This is distinct from the claim that current systems delete disfluencies.

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 / 5 minor

Summary. The paper argues that speech translation systems, including SpeechLLMs, strip disfluencies at a cost: disfluencies carry affective and interpretive meaning that is lost when utterances are cleaned. The authors introduce UH-MAZING, a human-translated, disfluency-annotated benchmark built from 80 manually selected Switchboard utterances, translated into eight target languages. Through human emotion annotation, a human error taxonomy over 3,154 spans, and a disfluency-type ablation, they claim that false starts and self-repairs—not filled pauses or discourse markers—drive most translation-quality loss, and that models typically delete disfluencies rather than mistranslate them. They also propose inference-time interventions (beam search, in-context learning) that improve disfluency preservation without retraining. The central empirical claims are triangulated but rest on a small, hand-picked sample and on metric choices that need closer scrutiny.

Significance. If the claims hold, the paper identifies a measurable and currently missing dimension of speech translation quality: preserving disfluencies as meaningful signals rather than noise. The benchmark, the fine-grained error taxonomy, and the cross-lingual setup are valuable contributions, and the paper is generally careful: it uses human annotation, reports effect sizes, checks leave-one-out stability, and will release code and data. The main significance is moderated by the fact that the entire benchmark and the emotion study rest on 80 manually selected utterances, which limits the generality of the cross-lingual and disfluency-type conclusions, and by a potential confound in the central Figure 5 ablation that uses COMET-Kiwi to measure the effect of false starts.

major comments (4)
  1. [§2.1, §3.1] The entire benchmark and emotion study rest on 80 manually chosen Switchboard utterances ('manually chosen to contain multiple disfluencies and to span a range of lengths'). This is not a random or stratified sample, and the selection criterion may preferentially include utterances with salient false starts and emotionally charged disfluencies. Because the same 80 utterances are used for all eight-language translations, the false-start ablation (Figure 5), and the deletion-vs-mistranslation counts (Figure 3), the cross-lingual and disfluency-type generalizations are underdetermined. A random-sample validation set—or at least a demonstration that the 80 utterances match Switchboard's disfluency-type and emotion distributions—is needed to support the paper's central claims.
  2. [§5.4, Figure 5] The disfluency-type ablation uses COMET-Kiwi, yet §5.1 and Table 1 show that COMET-Kiwi is the metric least sensitive to disfluency removal and is trained predominantly on fluent data. EDITED spans contain lexical content (false starts and self-repairs), so a reference-free metric can penalize the source content itself rather than the translation system's failure to preserve disfluency. This confound directly affects the claim that false starts, not filled pauses, drive translation-quality loss. Please recompute Figure 5 with chrF and, crucially, with the human reference translations in place of model outputs; if the same drop appears for human translations, the effect is not a deficiency of current systems.
  3. [§2.2] The emotion claim is based on small aggregate effects (Cramér's V = 0.07–0.08, ε² = 0.01), while the abstract emphasizes a 71 percentage-point extreme case. With open-ended emotion labels mapped through two pipelines, the extreme-case framing can be misleading. Please report the full distribution of per-utterance negative-label shifts, the proportion of utterances with shifts above a meaningful threshold, and the effect sizes after excluding the most extreme utterances. The reported leave-one-out analysis is a useful robustness check but does not address the influence of the hand-picked sample itself.
  4. [§4.2, §5.1] Human evaluation is central to the metric validation and to the error taxonomy (3,154 spans), but only the first 20 items were triple-annotated; the remaining items received a single annotation. Given that the aggregate taxonomy counts and the correlation numbers in Table 1 rely on these human judgments, please provide per-item reliability on a larger subset, or confidence intervals for the aggregate counts and for the segment- and system-level correlations. This is especially important because the main deletion-vs-mistranslation comparison (1,098 vs. 446 spans) depends on the consistency of span annotations across items.
minor comments (5)
  1. [§5.2] The sentence 'Cascaded systems lead by a wide margin over all audio-input approaches' is not supported by Table 2 for GEMINI Audio on several languages (e.g., AR 37.15 vs. TOWER Canary 22.87; HI 57.60 vs. 52.24). Please clarify whether the comparison is to all evaluated systems or only to open-source audio-input models.
  2. [Table 3] The header 'utt. neg-share A/D/F' and values like '3623/60/94%' are ambiguous. Show the three percentages separately and define whether the final number is the Audio-to-Fluent shift.
  3. [§3.2] The expert↔Prolific agreement is measured by embedding cosine after removing underscore-marked disfluencies. This validates general translation similarity but not the placement or type of preserved disfluencies; a token-level disfluency agreement metric would be more informative for the benchmark's stated purpose.
  4. [References] The claim that COMET-Kiwi is fluency-biased (used in §5.1 and §5.4) cites an author self-publication (Züfle et al., 2026). An independent reference or a small internal experiment demonstrating the bias would strengthen the argument.
  5. [Abstract, §1] The abstract and contributions say the benchmark and code are 'released', but §1 conditions release on LDC approval. Please state the current availability status and timeline.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims rest on independent human evaluation and benchmark measurements; self-citations are contextual, not load-bearing.

full rationale

The paper's central claims are empirical and self-contained. The emotion study (Section 2) compares human ratings across Audio, Disfluent, and Fluent conditions; the shifts in perceived emotion are measured, not derived from the definition of disfluency. The UH-MAZING benchmark is constructed with reference translations that intentionally preserve disfluencies, but the finding that systems delete rather than mistranslate disfluencies comes from a separate human error-span annotation (Figure 3), not from the benchmark's construction alone. The false-start ablation (Section 5.4) reinserts annotated disfluency types into fluent text and measures COMET-Kiwi changes; this is a controlled comparison, not a tautology. Self-citations (e.g., Teleki et al. 2025b for E/Z-scores, Zouhar and Kocmi 2026 for Pearmut, Züfle et al. 2026 for COMET-Kiwi's training data) provide tools or contextual explanations; the paper also reports its own metric-correlation results (Table 1), so those citations are not the sole support for any central claim. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported, and no ansatz is smuggled in via citation. The 80-utterance sample is a limitation for generality, but it is not a circularity.

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

The paper is an empirical benchmark study; it introduces no fitted physical parameters, no new particles or mediators, and no mathematical entities. It rests on domain assumptions about annotation reliability, sample representativeness, and metric validity, listed above.

assumptions (4)
  • domain assumption Switchboard disfluency annotations (INTJ, PRN, EDITED) are accurate and can be transferred to translations by crowd annotators.
    Used in Section 2.1 to select utterances and in Section 3.1 to build UH-MAZING; the benchmark quality assumes the Round 2 highlighting is correct.
  • ad hoc to paper 80 manually selected utterances represent conversational disfluent speech well enough for cross-lingual generalization.
    Section 2.1 says they were manually chosen for high disfluency density; no random sample or population justification is given.
  • domain assumption Free-text emotion labels mapped through NRC EmoLex and an embedding pipeline preserve the intended emotion categories.
    Section 2.1 describes a three-pass cascade and cosine-similarity valence assignment; these mappings are not validated by an external gold standard.
  • domain assumption Automatic metrics (chrF, COMET-Kiwi) and LLM judge rankings generalize beyond the four systems selected for human evaluation.
    Section 5.1 validates metrics against human judgments on four systems; these correlations are then used to rank all systems in Table 2 and Figure 5.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Role of Disfluencies in Speech Translation." pith.science (2026). https://pith.science/paper/ZCEVE67X

@misc{pith2026260802138,
  author       = {Pith},
  title        = {Pith review of: The Role of Disfluencies in Speech Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZCEVE67X}},
  note         = {Machine review of arXiv:2608.02138}
}
read the original abstract

Current speech translation systems, including SpeechLLMs, are trained on cleaned text and tend to strip disfluencies like filled pauses and false starts rather than translate them. We show this comes at a cost: disfluencies carry meaning that gets lost when speech is cleaned up. To study this systematically, we introduce Uh-Mazing, a benchmark of human-translated, disfluency-annotated Switchboard speech covering English into eight target languages. Across these languages and several architectures, we find that false starts and self-repairs, not filled pauses or discourse markers, drive most of the translation-quality loss, and that models which fail to preserve a disfluency tend to omit it rather than mistranslate it. We show inference-time decoding can mitigate this without retraining, and release the benchmark and code.

Figures

Figures reproduced from arXiv: 2608.02138 by the authors.

Figure 1
Figure 1. Removing audio signal and disfluencies increases perceived intensity for seven of the eight emotions. Disgust reverses. Six of the eight increase at both steps: Audio→Disfluent and Disfluent→Fluent. A tick inside the bar marks these emotions. grouped by valence into negative (anger, fear, sad￾ness, disgust), positive (joy, trust), and variable (surprise, anticipation). Because NRC EmoLex indexes surface forms only, … view at source ↗
Figure 2
Figure 2. Across languages, disfluencies most often follow a coordinating conjunction and precede the sentence subject. Position of disfluencies by the de￾pendency relation of neighboring content words. with disfluencies frequently arising right after con￾nectives such as and or but. The strength of this asymmetry still differs by language. Context Shift on Translation. We next ask whether the syntactic context of a disfluenc… view at source ↗
Figure 3
Figure 3. Three failure modes (⋆), incorrect language, missing content, and deletion, dominate consistently, regardless of whether error spans are broken down by language (left) or model (right). Human annotators marked the spans, assigning each a category from our taxonomy and a severity (Major/Minor). while CHATGPT improves under most conditions but stays roughly flat with an automatic tran￾script. SpeechLLMs diverge in opp… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: All systems preserve meaning better than style in translation. Style Preservation vs. Meaning Preservation scores for one representative system per architecture category, across eight languages. ceeds style preservation for nearly every system and language, with almost…
Figure 5
Figure 5. Figure 5: EDITED-type disfluencies drive the drop in translation performance. We show the change in COMET-Kiwi relative to the Fluent source baseline, by which disfluency type(s) the source text contains, pooled across four translation models (Llama, Tower, Gemini, ChatGPT) and …
Figure 6
Figure 6. Figure 6: Decoding methods shape overall quality (COMET), while ICL improves disfluency preserva￾tion (chrF); the two interventions complement each other, as do the metrics. Decoding strategies per model (rows), in COMET (left) and chrF (right), aver￾aged over eight EN→X pairs. …
Figure 7
Figure 7. Figure 7: Instructions to the annotators to identify the meaning of disfluencies using the Pearmut platform ( [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Mean intensity ratings per NRC emotion across the three conditions (Audio, Disfluent, Fluent), with [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Mean expert↔Prolific cosine similarity per target language under three OpenAI embed￾ders (text-embedding-3-small, -3-large, and -ada-002). The Latin- vs. non-Latin-script gap shrinks and eventually inverts as the embedder changes, indicating that part of the apparent n…
Figure 10
Figure 10. Figure 10: Instructions given to annotators for the creation of U [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Top: Translation prompts for cascaded models, SpeechLLMs, and commercial models. The high￾lighted clause is added only in the disfluency-aware variant. Bottom: LLM-as-a-judge prompt for disfluency￾aware translation evaluation. The highlighted line is omitted in the re…
Figure 12
Figure 12. Figure 12: Annotation guidelines for marking and categorising error spans and general judgments of translation [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

97 extracted references · 13 canonical work pages

  1. [1]

    Fillers , year = 2025, publisher =

    Rice, Alexander , title =. Fillers , year = 2025, publisher =. doi:10.5281/zenodo.15697595 , url =

    work page doi:10.5281/zenodo.15697595 2025

  2. [2]

    Speech Translation and Metrics in 2026: Findings of the IWSLT Campaign

    Adelani, David Ifeoluwa and Agostinelli, Victor and Anastasopoulos, Antonios and Bentivogli, Luisa and Bojar, Ond. Speech Translation and Metrics in 2026: Findings of the IWSLT Campaign. Proceedings of the 23rd International Conference on Spoken Language Translation ( IWSLT 2026). 2026. doi:10.18653/v1/2026.iwslt-1.39

    work page doi:10.18653/v1/2026.iwslt-1.39 2026

  3. [3]

    Direct Speech Translation for Automatic Subtitling

    Papi, Sara and Gaido, Marco and Karakanta, Alina and Cettolo, Mauro and Negri, Matteo and Turchi, Marco. Direct Speech Translation for Automatic Subtitling. Transactions of the Association for Computational Linguistics. 2023. doi:10.1162/tacl_a_00607

    work page doi:10.1162/tacl_a_00607 2023

  4. [4]

    How ``Real'' is Your Real-Time Simultaneous Speech-to-Text Translation System?

    Papi, Sara and Pol \'a k, Peter and Mach \'a c ek, Dominik and Bojar, Ond r ej. How ``Real'' is Your Real-Time Simultaneous Speech-to-Text Translation System?. Transactions of the Association for Computational Linguistics. 2025. doi:10.1162/tacl_a_00740

    doi:10.1162/tacl_a_00740 2025

  5. [5]

    Improved speech-to-text translation with the Fisher and Callhome S panish- E nglish speech translation corpus

    Post, Matt and Kumar, Gaurav and Lopez, Adam and Karakos, Damianos and Callison-Burch, Chris and Khudanpur, Sanjeev. Improved speech-to-text translation with the Fisher and Callhome S panish- E nglish speech translation corpus. Proceedings of the 10th International Workshop on Spoken Language Translation: Papers. 2013

    2013

  6. [6]

    2026 , eprint=

    Why We Need Speech to Evaluate Speech Translation , author=. 2026 , eprint=

    2026

  7. [7]

    2025 , eprint=

    Enhancing Naturalness in LLM-Generated Utterances through Disfluency Insertion , author=. 2025 , eprint=

    2025

  8. [8]

    Mind the Pause: Disfluency-Aware Objective Tuning for Multilingual Speech Correction with LLM s

    Kumar, Deepak and Gain, Baban and Ekbal, Asif. Mind the Pause: Disfluency-Aware Objective Tuning for Multilingual Speech Correction with LLM s. Proceedings of the 64th Annual Meeting of the A ssociation for C omputational L inguistics (Volume 1: Long Papers). 2026. doi:10.18653/v1/2026.acl-long.2137

    work page doi:10.18653/v1/2026.acl-long.2137 2026

Show all 97 references
  1. [9]

    2026 , eprint=

    Transcription Policy as a Latent Variable: Activating Controllable Verbatim ASR with Word-Level Timing , author=. 2026 , eprint=

    2026

  2. [10]

    Augmenting Automatic Speech Recognition Models With Disfluency Detection , year=

    Amann, Robin and Li, Zhaolin and Bruno, Barbara and Niehues, Jan , booktitle=. Augmenting Automatic Speech Recognition Models With Disfluency Detection , year=

  3. [11]

    End-to-End Speech Recognition and Disfluency Removal

    Jamshid Lou, Paria and Johnson, Mark. End-to-End Speech Recognition and Disfluency Removal. Findings of the Association for Computational Linguistics: EMNLP 2020. 2020. doi:10.18653/v1/2020.findings-emnlp.186

    2020 doi

  4. [12]

    The Thirty-Third

    Qianqian Dong and Feng Wang and Zhen Yang and Wei Chen and Shuang Xu and Bo Xu , title =. The Thirty-Third. 2019 , url =. doi:10.1609/AAAI.V33I01.33016351 , timestamp =

    2019 doi

  5. [13]

    Towards Fluent Translations From Disfluent Speech , year=

    Salesky, Elizabeth and Burger, Susanne and Niehues, Jan and Waibel, Alex , booktitle=. Towards Fluent Translations From Disfluent Speech , year=

  6. [14]

    Fluent Translations from Disfluent Speech in End-to-End Speech Translation

    Salesky, Elizabeth and Sperber, Matthias and Waibel, Alexander. Fluent Translations from Disfluent Speech in End-to-End Speech Translation. Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Techno...

    2019 doi

  7. [15]

    2020 , eprint=

    Language Models are Few-Shot Learners , author=. 2020 , eprint=

    2020

  8. [16]

    Gemini 2.5:

    Gheorghe. Gemini 2.5:. 2025 , eprint=

    2025

  9. [17]

    , booktitle=

    Miller, George A. , booktitle=. 1994 , url=

    1994

  10. [18]

    Porter , year=

    Martin F. Porter , year=. Snowball:

  11. [19]

    , editor=

    Qi, Peng and Zhang, Yuhao and Zhang, Yuhui and Bolton, Jason and Manning, Christopher D. , editor=. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations , month=. 2020 , address=. doi:10.18653/v1/2020.acl-demos.14 , pages=

    2020 doi

  12. [20]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month=

    Large Language Models are not Fair Evaluators , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month=. 2024 , address=. doi:10.18653/v1/2024.acl-long.511 , pages=

    2024 doi

  13. [21]

    Pearmut:

    Vilém Zouhar and Tom Kocmi , year=. Pearmut:. 2601.02933 , archiveprefix=

    arXiv

  14. [22]

    Error Span Annotation:

    Kocmi, Tom and Zouhar, Vil. Error Span Annotation:. Proceedings of the Ninth Conference on Machine Translation , month=. 2024 , address=. doi:10.18653/v1/2024.wmt-1.131 , pages=

    2024 doi

  15. [23]

    Experts, Errors, and Context:

    Freitag, Markus and Foster, George and Grangier, David and Ratnakar, Viresh and Tan, Qijun and Macherey, Wolfgang , editor=. Experts, Errors, and Context:. Transactions of the Association for Computational Linguistics , volume=. 2021 , address=. doi:10.1162/tacl_a_00437 , pages=

    2021 doi

  16. [24]

    and Sulik, Justin and Hauser, David and Pepin, Kathleen and Jones, Spencer , title=

    Gordon, Andrew and Rothschild, David and Affonso, Filipe M. and Sulik, Justin and Hauser, David and Pepin, Kathleen and Jones, Spencer , title=. 2026 , month=. doi:10.31234/osf.io/pvdjr , url=

    2026 doi

  17. [25]

    Popovi. chr. Proceedings of the Tenth Workshop on Statistical Machine Translation , month=. 2015 , address=. doi:10.18653/v1/W15-3049 , pages=

    2015 doi

  18. [26]

    Proceedings of the IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) , year=

    Reproducing Whisper-Style Training Using An Open-Source Toolkit And Publicly Available Data , author=. Proceedings of the IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) , year=

  19. [27]

    de Souza, Jos

    Rei, Ricardo and C. de Souza, Jos. Proceedings of the Seventh Conference on Machine Translation (WMT) , month=. 2022 , address=. doi:10.18653/v1/2022.wmt-1.52 , pages=

    2022 doi

  20. [28]

    Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics , month=

    Papineni, Kishore and Roukos, Salim and Ward, Todd and Zhu, Wei-Jing , editor=. Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics , month=. 2002 , address=. doi:10.3115/1073083.1073135 , pages=

    2002

  21. [29]

    Consortium Meetween , year=

  22. [30]

    2025 , eprint=

    Qwen2.5-Omni Technical Report , author=. 2025 , eprint=

    2025

  23. [31]

    Comparing

    Maria Teleki and Xiangjue Dong and Soohwan Kim and James Caverlee , year=. Comparing. Interspeech 2024 , pages=

    2024

  24. [32]

    2024 , eprint=

    Phi-4 Technical Report , author=. 2024 , eprint=

    2024

  25. [33]

    2024 , eprint=

    The Llama 3 Herd of Models , author=. 2024 , eprint=

    2024

  26. [34]

    Guerreiro and José Pombal and João Alves and Pedro Teixeirinha and Amin Farajian and André F

    Ricardo Rei and Nuno M. Guerreiro and José Pombal and João Alves and Pedro Teixeirinha and Amin Farajian and André F. T. Martins , year=. Tower+:. 2506.17080 , archiveprefix=

    arXiv

  27. [35]

    2022 , copyright=

    Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman, Greg and McLeavey, Christine and Sutskever, Ilya , title=. 2022 , copyright=. doi:10.48550/ARXIV.2212.04356 , url=

    2022 doi arXiv

  28. [36]

    Canary-1B-v2 &

    Monica Sekoyan and Nithin Rao Koluguri and Nune Tadevosyan and Piotr Zelasko and Travis Bartley and Nikolay Karpov and Jagadeesh Balam and Boris Ginsburg , year=. Canary-1B-v2 &. 2509.14128 , archiveprefix=

  29. [37]

    Preliminaries to a Theory of Speech Disfluencies , school=

    Shriberg, Elizabeth , year=. Preliminaries to a Theory of Speech Disfluencies , school=

  30. [38]

    On The Landscape of Spoken Language Models:

    Siddhant Arora and Kai-Wei Chang and Chung-Ming Chien and Yifan Peng and Haibin Wu and Yossi Adi and Emmanuel Dupoux and Hung-Yi Lee and Karen Livescu and Shinji Watanabe , year=. On The Landscape of Spoken Language Models:. 2504.08528 , archiveprefix=

    arXiv

  31. [39]

    The timing bottleneck:

    Liesenfeld, Andreas and Lopez, Alianda and Dingemanse, Mark , editor=. The timing bottleneck:. Proceedings of the 24th Annual Meeting of the Special Interest Group on Discourse and Dialogue , month=. 2023 , address=. doi:10.18653/v1/2023.sigdial-1.45 , pages=

    2023 doi

  32. [40]

    Tools for Analyzing Talk:

    MacWhinney, Brian , organization=. Tools for Analyzing Talk:. 2025 , month=. doi:10.21415/3mhn-0z89 , note=

    2025 doi

  33. [41]

    MacWhinney, Brian , edition=. The. 2000 , publisher=

    2000

  34. [42]

    and Branson, M

    Ardila, R. and Branson, M. and Davis, K. and Henretty, M. and Kohler, M. and Meyer, J. and Morais, R. and Saunders, L. and Tyers, F. M. and Weber, G. , title=. Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020) , pages=. 2020 , url=

    2020

  35. [43]

    Mozilla Common Voice , url=

  36. [44]

    1999 , note=

    Treebank-3 , author=. 1999 , note=

    1999

  37. [45]

    1993 , doi=

    Godfrey, John and Holliman, Edward , title=. 1993 , doi=

    1993

  38. [46]

    Summarizing Speech:

    Retkowski, Fabian and Z. Summarizing Speech:. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , month=. 2025 , address=. doi:10.18653/v1/2025.emnlp-main.1388 , pages=

    2025 doi

  39. [47]

    1992 , organization=

    Godfrey, John J and Holliman, Edward C and McDaniel, Jane , booktitle=. 1992 , organization=

    1992

  40. [48]

    2026 , url=

    Meta , title=. 2026 , url=

    2026

  41. [49]

    2025 , url=

    TranscribeGlass , author=. 2025 , url=

    2025

  42. [50]

    2025 , url=

    Xander Captioning Glasses , author=. 2025 , url=

    2025

  43. [51]

    Hu, Junjie and Ruder, Sebastian and Siddhant, Aditya and Neubig, Graham and Firat, Orhan and Johnson, Melvin , booktitle=. Xtreme:. 2020 , organization=

    2020

  44. [52]

    Common voice:

    Ardila, Rosana and Branson, Megan and Davis, Kelly and Kohler, Michael and Meyer, Josh and Henretty, Michael and Morais, Reuben and Saunders, Lindsay and Tyers, Francis and Weber, Gregor , booktitle=. Common voice:. 2020 , url=

    2020

  45. [53]

    2025 , url=

    Mozilla Common Voice Spontaneous Speech 2.0 Collection , author=. 2025 , url=

    2025

  46. [54]

    Fisher and

    Post, Matt and Kumar, Gaurav and Lopez, Adam and Karakos, Damianos and Callison-Burch, Chris and Khudanpur, Sanjeev , year=. Fisher and

  47. [55]

    Conversational Speech Reveals Structural Robustness Failures in

    Maria Teleki and Sai Janjur and Haoran Liu and Oliver Grabner and Ketan Verma and Thomas Docog and Xiangjue Dong and Lingfeng Shi and Cong Wang and Stephanie Birkelbach and Jason Kim and Yin Zhang and Éva Székely and James Caverlee , year=. Conversational Speech Reveals Struct...

  48. [56]

    2024 , eprint=

    A Survey on LLM Inference-Time Self-Improvement , author=. 2024 , eprint=

    2024

  49. [57]

    Z-Scores:

    Maria Teleki and Sai Janjur and Haoran Liu and Oliver Grabner and Ketan Verma and Thomas Docog and Xiangjue Dong and Lingfeng Shi and Cong Wang and Stephanie Birkelbach and Jason Kim and Yin Zhang and James Caverlee , year=. Z-Scores:. ICASSP , url=

  50. [58]

    SpeechSpectrum:

    Anna Seo Gyeong Choi and Maria Teleki and Miguel del Rio and James Caverlee and Corey Miller and Allison Koenecke , year=. SpeechSpectrum:

  51. [59]

    Beyond Single Ground Truth:

    Anna Seo Gyeong Choi and Maria Teleki and James Caverlee and Miguel del Rio and Corey Miller and Hoon Choi , year=. Beyond Single Ground Truth:. 2605.07084 , archiveprefix=

    arXiv

  52. [60]

    Masculine Defaults via Gendered Discourse in Podcasts and Large Language Models

    Teleki, Maria and Dong, Xiangjue and Liu, Haoran and Caverlee, James. Masculine Defaults via Gendered Discourse in Podcasts and Large Language Models. Proceedings of the Third Workshop on Social Influence in Conversations (SICon 2025). 2025. doi:10.18653/v1/2025.sicon-1.7

    2025 doi

  53. [61]

    Advances in neural information processing systems , volume=

    Judging llm-as-a-judge with mt-bench and chatbot arena , author=. Advances in neural information processing systems , volume=. 2023 , url=

    2023

  54. [62]

    From generation to judgment:

    Li, Dawei and Jiang, Bohan and Huang, Liangjie and Beigi, Alimohammad and Zhao, Chengshuai and Tan, Zhen and Bhattacharjee, Amrita and Jiang, Yuxuan and Chen, Canyu and Wu, Tianhao and others , booktitle=. From generation to judgment:. 2025 , url=

    2025

  55. [63]

    Languages Still Left Behind:

    Taguchi, Chihiro and Mai, Seng and Kurabe, Keita and Sakai, Yusuke and Agyei, Georgina and Eslami, Soudabeh and Chiang, David , booktitle=. Languages Still Left Behind:. 2025 , url=

    2025

  56. [64]

    Generating Fluent Translations from Disfluent Text Without Access to Fluent References:

    Saini, Nikhil and Khatri, Jyotsana and Jyothi, Preethi and Bhattacharyya, Pushpak , editor=. Generating Fluent Translations from Disfluent Text Without Access to Fluent References:. Proceedings of the 17th International Conference on Spoken Language Translation , month=. 2020 ...

    2020 doi

  57. [65]

    Leveraging llms for written to spoken style data transformation to enhance spoken dialog state tracking , author=. Proc. Interspeech , volume=. 2025 , url=

    2025

  58. [66]

    I want a horror -- comedy -- movie:

    Maria Teleki and Lingfeng Shi and Chengkai Liu and James Caverlee , year=. I want a horror -- comedy -- movie:. INTERSPEECH , url=

  59. [67]

    2025 , eprint=

    A Survey of Context Engineering for Large Language Models , author=. 2025 , eprint=

    2025

  60. [68]

    doi:10.21437/Interspeech.2007-650 , issn =

    Darren Scott Appling and Nick Campbell , year =. doi:10.21437/Interspeech.2007-650 , issn =

    2007 doi

  61. [69]

    Rohit Kundu and Preethi Jyothi and Pushpak Bhattacharyya , year=. Survey:

  62. [70]

    Effect of machine translation in interlingual conversation:

    Hara, Kotaro and Iqbal, Shamsi T , booktitle=. Effect of machine translation in interlingual conversation:. 2015 , url=

    2015

  63. [71]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    An interdisciplinary approach to human-centered machine translation , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=. 2025 , url=

    2025

  64. [72]

    Proceedings of the 2022 ACM conference on fairness, accountability, and transparency , pages=

    Reliable and safe use of machine translation in medical settings , author=. Proceedings of the 2022 ACM conference on fairness, accountability, and transparency , pages=. 2022 , url=

    2022

  65. [73]

    Findings of the Association for Computational Linguistics: NAACL 2022 , pages=

    Opportunities for human-centered evaluation of machine translation systems , author=. Findings of the Association for Computational Linguistics: NAACL 2022 , pages=. 2022 , url=

    2022

  66. [74]

    Speech Rate Comparison When Talking to a System and Talking to a Human:

    Akira, Hayakawa and Vogel, Carl and Luz, Saturnino and Campbell, Nick , booktitle=. Speech Rate Comparison When Talking to a System and Talking to a Human:. 2017 , url=

    2017

  67. [75]

    Traveling with

    Kundan, Zheng and Hossain, Md Sazzad and Shahidul Islam, Mohammad and Alam, Mohammad Nurul and Parvez, M Omar , journal=. Traveling with. 2025 , url=

    2025

  68. [76]

    Patterns , volume=

    A decade of gender bias in machine translation , author=. Patterns , volume=. 2025 , url=

    2025

  69. [77]

    Google Translate is our best friend here

    “Google Translate is our best friend here” A vignette-based interview study on machine translation use for health communication , author=. Translation Spaces , volume=. 2025 , url=

    2025

  70. [78]

    Mohammad and Peter D

    Saif M. Mohammad and Peter D. Turney , title =. Comput. Intell. , volume =. 2013 , url =. doi:10.1111/J.1467-8640.2012.00460.X , timestamp =

    2013

  71. [79]

    Invited talk:

    Brennan, Susan E , booktitle=. Invited talk:. 2000 , url=

    2000

  72. [80]

    Psychonomic Bulletin & Review , volume=

    Linguistic features of spontaneous speech predict conversational recall , author=. Psychonomic Bulletin & Review , volume=. 2024 , url=

    2024

  73. [81]

    , author=

    The effect of disfluency on memory for what was said. , author=. Journal of Experimental Psychology: Learning, Memory, and Cognition , volume=. 2023 , url=

    2023

  74. [82]

    2607.08034 , archiveprefix=

    Dhruv Agarwal and Anya Shukla and Tanya Goyal and Aditya Vashistha , year=. 2607.08034 , archiveprefix=

    arXiv

  75. [83]

    Plurals:

    Ashkinaze, Joshua and Fry, Emily and Edara, Narendra and Gilbert, Eric and Budak, Ceren , booktitle=. Plurals:. 2025 , url=

    2025

  76. [84]

    Multi-Perspective

    Navya Mehrotra and Adam Visokay and Kristina Gligorić , year=. Multi-Perspective. 2603.21404 , archiveprefix=

  77. [85]

    Cultivating Pluralism In Algorithmic Monoculture:

    Lily Hong Zhang and Smitha Milli and Karen Jusko and Jonathan Smith and Brandon Amos and Wassim Bouaziz and Manon Revel and Jack Kussman and Yasha Sheynin and Lisa Titus and Bhaktipriya Radharapu and Jane Yu and Vidya Sarma and Kris Rose and Maximilian Nickel , year=. Cultivat...

  78. [86]

    2004 , publisher=

    Medical interpreting and cross-cultural communication , author=. 2004 , publisher=

    2004

  79. [87]

    Cognition , volume=

    Using uh and um in spontaneous speaking , author=. Cognition , volume=. 2002 , url=

    2002

  80. [88]

    2026 , url=

    Thebaud, Thomas and Wang, Yuzhe and Zhang, Hao and Ugandhar, Sathvik Manikantan Napa and Hallur, Ashish and Tinchev, Georgi and Ravichandran, Venkatesh and Moro-Velazquez, Laureano , journal=. 2026 , url=

    2026

  81. [89]

    Proceedings of the 22nd annual conference of the European association for machine translation , pages=

    On the differences between human translations , author=. Proceedings of the 22nd annual conference of the European association for machine translation , pages=. 2020 , url=

    2020

  82. [90]

    Romance studies , volume=

    Translation, intertextuality, interpretation , author=. Romance studies , volume=. 2009 , url=

    2009

  83. [91]

    Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) , month=

    Quantifying the Impact of Disfluency on Spoken Content Summarization , author=. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) , month=. 2024 , address=

    2024

  84. [92]

    2025 , eprint=

    VocalBench-DF: A Benchmark for Evaluating Speech LLM Robustness to Disfluency , author=. 2025 , eprint=

    2025

  85. [93]

    Full-duplex-bench-v3:

    Lin, Guan-Ting and Chen, Chen and Chen, Zhehuai and Lee, Hung-yi , journal=. Full-duplex-bench-v3:. 2026 , url=

    2026

  86. [94]

    Closing the Gap Between Text and Speech Understanding in

    Santiago Cuervo and Skyler Seto and Maureen de Seyssel and Richard He Bai and Zijin Gu and Tatiana Likhomanenko and Navdeep Jaitly and Zakaria Aldeneh , booktitle=. Closing the Gap Between Text and Speech Understanding in. 2026 , url=

    2026

  87. [95]

    2015 international conference on affective computing and intelligent interaction (ACII) , pages=

    Emotion recognition in spontaneous and acted dialogues , author=. 2015 international conference on affective computing and intelligent interaction (ACII) , pages=. 2015 , organization=

    2015

  88. [96]

    , author=

    Detection of real-life emotions in call centers. , author=. Interspeech , volume=. 2005 , url=

    2005

  89. [97]

    Theories of emotion , pages=

    A general psychoevolutionary theory of emotion , author=. Theories of emotion , pages=. 1980 , publisher=

    1980

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.