{"slug": "dial-s2st-learning-latency-control-for-simultaneous-speech-to-speech-translation", "title": "DIAL-S2ST: Learning Latency Control for Simultaneous Speech-to-Speech Translation", "summary": "Researchers introduced DIAL-S2ST, a post-training method that learns inference-time latency control for simultaneous speech-to-speech translation models, according to arXiv paper 2610.32214v1. The method self-distills from lookahead predictions to supervise earlier emission without paired target data or reward computation, and jointly trains a single LoRA adapter across scales with matched lookahead guidance. In French-to-English experiments with the Hibiki model, only 10 hours of source speech enabled the adaptation, yielding flexible latency control with a favorable quality-latency trade-off.", "body_md": "arXiv:2610.32214v1 Announce Type: new \nAbstract: We propose DIAL-S2ST, a post-training method for learning inference-time latency control in simultaneous speech-to-speech translation models. Our approach reuses pretrained translation capabilities through self-distillation from lookahead predictions, while limiting the influence of large text distillation losses to account for the imitation gap. This provides supervision for earlier emission without paired target data or reward computation. To make the learned timing adjustable at inference, we jointly train a single LoRA adapter across scales with matched lookahead guidance. Experiments on French-to-English translation with Hibiki demonstrate that only 10 hours of source speech enable this adaptation, providing flexible latency control with a favorable quality-latency trade-off.", "url": "https://wpnews.pro/news/dial-s2st-learning-latency-control-for-simultaneous-speech-to-speech-translation", "canonical_source": "https://www.machinebrief.com/news/dial-s2st-learning-latency-control-for-simultaneous-speech-t-emy6", "published_at": "2026-10-26 04:00:00+00:00", "updated_at": "2026-10-26 04:42:17.308338+00:00", "lang": "en", "topics": ["machine-learning", "natural-language-processing", "ai-research", "large-language-models"], "entities": ["DIAL-S2ST", "Hibiki", "arXiv", "LoRA"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/dial-s2st-learning-latency-control-for-simultaneous-speech-to-speech-translation", "markdown": "https://wpnews.pro/news/dial-s2st-learning-latency-control-for-simultaneous-speech-to-speech-translation.md", "text": "https://wpnews.pro/news/dial-s2st-learning-latency-control-for-simultaneous-speech-to-speech-translation.txt", "jsonld": "https://wpnews.pro/news/dial-s2st-learning-latency-control-for-simultaneous-speech-to-speech-translation.jsonld"}}