# The Role of Disfluencies in Speech Translation

arXiv: https://arxiv.org/abs/2608.02138 · HTML: https://mlsanity.com/paper/2608.02138/

## 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 (most cited first, 25 of 25)

- [Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena](https://mlsanity.com/paper/2306.05685/) (2023)
- [Robust Speech Recognition via Large-Scale Weak Supervision](https://mlsanity.com/paper/2212.04356/) (2022)
- [Stanza: A Python Natural Language Processing Toolkit for Many Human Languages](https://mlsanity.com/paper/2003.07082/) (2020)
- [From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge](https://mlsanity.com/paper/2411.16594/) (2024)
- [Experts, Errors, and Context: A Large-Scale Study of Human Evaluation for Machine Translation](https://mlsanity.com/paper/2104.14478/) (2021)
- [Crowdsourcing a Word-Emotion Association Lexicon](https://mlsanity.com/paper/1308.6297/) (2013)
- [Error Span Annotation: A Balanced Approach for Human Evaluation of Machine Translation](https://mlsanity.com/paper/2406.11580/) (2024)
- [Canary-1B-v2 & Parakeet-TDT-0.6B-v3: Efficient and High-Performance Models for Multilingual ASR and AST](https://mlsanity.com/paper/2509.14128/) (2025)
- [Fluent Translations from Disfluent Speech in End-to-End Speech Translation](https://mlsanity.com/paper/1906.00556/) (2019)
- [A Survey on LLM Inference-Time Self-Improvement](https://mlsanity.com/paper/2412.14352/) (2024)
- [How "Real" is Your Real-Time Simultaneous Speech-to-Text Translation System?](https://mlsanity.com/paper/2412.18495/) (2024)
- [End-to-End Speech Recognition and Disfluency Removal](https://mlsanity.com/paper/2009.10298/) (2020)
- [Pearmut: Human Evaluation of Translation Made Trivial](https://mlsanity.com/paper/2601.02933/) (2026)
- [The timing bottleneck: Why timing and overlap are mission-critical for conversational user interfaces, speech recognition and dialogue systems](https://mlsanity.com/paper/2307.15493/) (2023)
- [VocalBench-DF: A Benchmark for Evaluating Speech LLM Robustness to Disfluency](https://mlsanity.com/paper/2510.15406/) (2025)
- [Towards Fluent Translations from Disfluent Speech](https://mlsanity.com/paper/1811.03189/) (2018)
- [Languages Still Left Behind: Toward a Better Multilingual Machine Translation Benchmark](https://mlsanity.com/paper/2508.20511/) (2025)
- [Augmenting Automatic Speech Recognition Models with Disfluency Detection](https://mlsanity.com/paper/2409.10177/) (2024)
- [An Interdisciplinary Approach to Human-Centered Machine Translation](https://mlsanity.com/paper/2506.13468/) (2025)
- [Enhancing Naturalness in LLM-Generated Utterances through Disfluency Insertion](https://mlsanity.com/paper/2412.12710/) (2024)
- [Conversational Speech Reveals Structural Robustness Failures in SpeechLLM Backbones](https://mlsanity.com/paper/2509.20321/) (2025)
- [Beyond Single Ground Truth: Reference Monism as Epistemic Injustice in ASR Evaluation](https://mlsanity.com/paper/2605.07084/) (2026)
- [Z-Scores: A Metric for Linguistically Assessing Disfluency Removal](https://mlsanity.com/paper/2509.20319/) (2025)
- [Mind the Pause: Disfluency-Aware Objective Tuning for Multilingual Speech Correction with LLMs](https://mlsanity.com/paper/2605.12242/) (2026)
- [Why We Need Speech to Evaluate Speech Translation](https://mlsanity.com/paper/2605.28227/) (2026)

## 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.