Optimizing endpointing thresholds using dialogue features in a spoken dialogue system

This paper describes a novel algorithm to dynamically set endpointing thresholds based on a rich set of dialogue features to detect the end of user utterances in a dialogue system. By analyzing the relationship between silences in user's speech to a spoken dialogue system and a wide range of automatically extracted features from discourse, semantics, prosody, timing and speaker characteristics, we found that all features correlate with pause duration and with whether a silence indicates the end of the turn, with semantics and timing being the most informative. Based on these features, the proposed method reduces latency by up to 24% over a fixed threshold baseline. Offline evaluation results were confirmed by implementing the proposed algorithm in the Let's Go system.