Turn-taking latency seems larger than expected
Hi, thanks for releasing this impressive model.
I have a question regarding the reported turn-taking latency. The highlights section mentions a latency of around 450 ms, but when listening to the demo samples, the pause between the user’s speech and the model’s response seems noticeably longer than I expected.
Is the reported latency measured under a specific definition or evaluation setup? For example, does it exclude endpoint detection or some other processing time? I just wanted to make sure I’m interpreting the number correctly.
Thanks!
Hi Jihwan Park,
Thank you for your kind words about the model!
Regarding your question, the 450 ms turn-taking latency comes from the Full Duplex Bench (FDB) 1.0 evaluation (https://github.com/DanielLin94144/Full-Duplex-Bench).
The recordings were done in an interactive environment, which includes factors like network latency, VPN, and other real-world conditions that can impact the perceived pause.
Thanks for the clarification! That makes sense.