A new benchmark called Uh-Mazing has been developed to study the impact of disfluencies in speech translation. Current systems often remove these speech irregularities, leading to a loss of meaning. The benchmark, which includes human-translated, disfluency-annotated speech from the London Lesbian and Gay Switchboard, reveals that false starts and self-repairs significantly degrade translation quality. Researchers found that models tend to omit disfluencies rather than mistranslate them, and that inference-time decoding can help mitigate this issue without retraining. AI
IMPACT Highlights the importance of preserving speech disfluencies for more accurate machine translation.
RANK_REASON The cluster describes a new academic paper and benchmark released on arXiv.
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
