Barge-in
×2
the latency budget, paid twice per interruption
A listener talking over synthesised speech, expecting it to stop and answer the way a person would. Speech that keeps running reads as a machine immediately.
What barge-in demands of TTS
- Instant stop: kill the audio stream mid-word the moment the caller speaks. A persistent socket per call makes this one message.
- Fast restart: the answer is a fresh synthesis inside the same turn budget, so you pay the first-audio wait a second time.
- No billing penalty: the redone sentence must cost nothing, or every interruption becomes a small invoice for words nobody heard.
The interrupted turn, to scale
fig.
A normal turn vs an interrupted turn, same one-second budget
Fig
Normal turn, STT → LLM → TTS
550 ms and up
Interrupted turn, detect, stop, redo all three
the same stages, twice
▲ the restart begins here
the caller’s patience
Every fixed cost in the pipeline is paid twice per interruption, the stages you can buy down, synthesis first among them, decide whether interruptions feel handled or fumbled.
Every term on this page is measurable. Take a key and read the numbers off your own requests.
Get a key and measure it yourself