The Role of Disfluencies in Speech Translation

2026-08-03

AI Summary

Current speech translation systems commonly filter out disfluencies—such as hesitations and repetitions—in spoken language, often leading to loss of semantic information. This work presents Uh-Mazing, the first multilingual human-translated benchmark with explicit disfluency annotations, built upon the Switchboard corpus, to systematically analyze the impact of different disfluency types on translation quality. The study reveals that self-corrections incur the greatest translation degradation, and models tend to omit disfluent content rather than mistranslate it. To address this issue, the authors propose a training-free, inference-time decoding strategy that effectively mitigates information loss. The benchmark dataset and implementation code are publicly released to support future research in disfluency-aware speech translation.

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.

Problem

• disfluencies
• speech translation
• false starts
• translation quality
• filled pauses

Innovation

• disfluencies
• speech translation
• Uh-Mazing benchmark
• false starts
• self-repairs