A Distributed Architecture for Real-time Dialogue and On-task Learning of Efficient Co-operative Turn-taking - Coverbal Synchrony in Human-Machine Interaction - page 319

Figure 8. Overlapped turns in our three evaluations.
with a length 0 to 1000 msecs between them; people tend to answer in
shorter sentences, not allowing for as many opportunities for mistakes.
7. Conclusions and Future Work
Our system learns to optimize STW and minimize speech overlaps
and awkward silences, using prosody analysis to predict interlocutor
behavior. It learns this on the fly, in a full-duplex “open-mic” (dynamic
interaction) setup, and can take turns very efficiently in dialogues
with copies of itself and with people, in relatively human-like ways.