Full‑duplex speech models were originally trained to converse with humans, but they are increasingly used to converse with each other—e.g., in self‑play data generation, agent societies, and model‑based evaluation. In this loop no human corrects timing errors: each model’s turn‑taking becomes the input for the other. We investigate what timing the loop converges to.
We let two PersonaPlex‑7B instances exchange audio tokens on a shared clock in unscripted conversation and apply the same floor‑transfer rule to them and to the Switchboard corpus. The timing of the two models is coupled: re‑pairing speakers across conversations destroys the coupling. Compared with the human median floor‑transfer latency of 137 ms, the models transfer the floor much later, with a median of 400‑560 ms.
A finer analysis shows that humans place about one‑tenth of their transfers in the last 120 ms of the partner’s turn, whereas the models allocate only 1 % of theirs to that interval. Introducing a delay in one direction of the channel shifts the model’s response one‑for‑one and leaves the run‑up to the transfer empty, consistent with a reactive wait after the perceived end rather than the turn‑end projection that human timing requires.
Review: This work reveals that full‑duplex models develop a stable yet delayed turn‑taking pattern in self‑play settings, highlighting the need for more human‑like projection mechanisms to improve the naturalness of model‑to‑model interaction.