FastTurn: Unifying Acoustic and Streaming Semantic Cues for Low-Latency and Robust Turn Detection

A research team led by Chengyou Wang has introduced FastTurn, a novel framework designed to solve a critical bottleneck in real-time AI conversations. Current full-duplex systems either rely on basic voice activity detection (which lacks semantic understanding) or ASR-based modules (which introduce latency and degrade with noise). FastTurn unifies streaming Connectionist Temporal Classification (CTC) decoding with acoustic features, enabling early decisions from partial audio observations while preserving semantic context. This hybrid approach allows the system to detect conversational turns with both speed and intelligence.

The researchers also addressed a significant data limitation by releasing a test set based on real human dialogue that captures authentic interaction dynamics including overlapping speech, backchannels, pauses, and environmental noise. Experiments demonstrate that FastTurn achieves higher decision accuracy with lower interruption latency compared to existing baselines. The 5-page paper, submitted to arXiv, shows the system's robustness under challenging acoustic conditions, making it practical for deployment in real-world spoken dialogue systems where natural, fluid conversation is essential.