New AI predicts when you'll stop talking, cutting voice latency by 505ms

Researchers from Carnegie Mellon University and Brno University of Technology (Sathwik Udupa, Shinji Watanabe, Petr Schwarz, Jan Cernocky) have introduced Endpoint Anticipation, a novel approach to reducing latency in spoken dialogue systems. Published at Interspeech 2026, the paper shifts from reactive turn‑completion detection to proactive forecasting. Their speech‑based model predicts when a speaker will finish up to 2.56 seconds in advance, enabling speculative execution of large language model (LLM) and text‑to‑speech (TTS) pipelines on partial context.

When integrated with the Unmute framework, Endpoint Anticipation achieved a 505 ms average latency reduction at the cost of a 28.4% increase in speculative computation. This effectively masks sequential bottlenecks, allowing complex reasoning in real‑time speech‑to‑speech interaction. The model consistently outperforms competitive VAP (Voice Activity Projection) baselines across both conversational and task‑oriented datasets. The authors also introduced metrics to quantify the trade‑off between realized latency reduction and computational redundancy.