mlsanityData to 14 Oct 2026

The Role of Disfluencies in Speech Translation

arXiv 2608.02138 ↗First posted 3 Aug 2026cs.CLcited by 0

Not on the map. Why?

Abstract (arXiv)

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.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 25 of its 55 references that are in the map. First posted on arXiv 3 Aug 2026.

Builds on

25 of 55 references.

References · 25

  1. 2023Judging LLM-as-a-Judge with MT-Bench and Chatbot Arenacited by 8,626
  2. 2022Robust Speech Recognition via Large-Scale Weak Supervisioncited by 4,057
  3. 2020Stanza: A Python Natural Language Processing Toolkit for Many Human Languagescited by 723
  4. 2024From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judgecited by 494
  5. 2021Experts, Errors, and Context: A Large-Scale Study of Human Evaluation for Machine Translationcited by 338
  6. 2013Crowdsourcing a Word-Emotion Association Lexiconcited by 196
  7. 2024Error Span Annotation: A Balanced Approach for Human Evaluation of Machine Translationcited by 62
  8. 2025Canary-1B-v2 & Parakeet-TDT-0.6B-v3: Efficient and High-Performance Models for Multilingual ASR and ASTcited by 60
  9. 2019Fluent Translations from Disfluent Speech in End-to-End Speech Translationcited by 22
  10. 2024A Survey on LLM Inference-Time Self-Improvementcited by 17
  11. 2024How "Real" is Your Real-Time Simultaneous Speech-to-Text Translation System?cited by 17
  12. 2020End-to-End Speech Recognition and Disfluency Removalcited by 16
13 more
  1. 2026Pearmut: Human Evaluation of Translation Made Trivialcited by 11
  2. 2023The timing bottleneck: Why timing and overlap are mission-critical for conversational user interfaces, speech recognition and dialogue systemscited by 9
  3. 2025VocalBench-DF: A Benchmark for Evaluating Speech LLM Robustness to Disfluencycited by 8
  4. 2018Towards Fluent Translations from Disfluent Speechcited by 7
  5. 2025Languages Still Left Behind: Toward a Better Multilingual Machine Translation Benchmarkcited by 7
  6. 2024Augmenting Automatic Speech Recognition Models with Disfluency Detectioncited by 6
  7. 2025An Interdisciplinary Approach to Human-Centered Machine Translationcited by 6
  8. 2024Enhancing Naturalness in LLM-Generated Utterances through Disfluency Insertioncited by 3
  9. 2025Conversational Speech Reveals Structural Robustness Failures in SpeechLLM Backbonescited by 2
  10. 2026Beyond Single Ground Truth: Reference Monism as Epistemic Injustice in ASR Evaluationcited by 2
  11. 2025Z-Scores: A Metric for Linguistically Assessing Disfluency Removalcited by 1
  12. 2026Mind the Pause: Disfluency-Aware Objective Tuning for Multilingual Speech Correction with LLMscited by 1
  13. 2026Why We Need Speech to Evaluate Speech Translationcited by 1

Built on by

Not cited yet.