Realtime and turn-based voice
The turn-based engine runs on the carrier path: the caller's turn becomes text, the model answers in text, and the text is spoken — so code sits between the answer and the caller. The realtime engine is speech-to-speech, opted into per workspace, and nothing sits in that gap. You are trading a checkpoint for immediacy, and that is the decision.
The structural difference#
| Turn-based (carrier path) | Realtime (speech-to-speech) | |
|---|---|---|
| Shape of a turn | Audio in, text through the model, audio out | Audio in, audio out |
| Providers | A carrier for the PSTN leg | LiveKit SIP plus a Gemini Live model, over the same carrier leg |
| Enabled | The default path | A workspace opts in |
| A rule can be | Prevented — each sentence is checked before it is spoken | Detected — the rule lives in the instructions, and a breach is counted afterwards |
| Interruption | Bounded by the turn | Barge-in is native to the conversation |
| Cost driver | Text tokens, plus speech synthesis | Audio tokens, billed at several times the text rate in both directions |
Everything else on this page follows from row four. A checkpoint you can stand at is worth a great deal, and it is exactly what speech-to-speech removes.
Latency, with the real number#
The floor on the realtime path is the model's first token plus its end-of-turn detection: 3.3 seconds median on the best measured call. Not one to two seconds. Roughly 60% of response time is model-bound, which means the remaining tuning surface is smaller than it looks — the target used internally is 2.5 seconds and anything past 4 seconds counts as slow, but the floor is the floor.
Two consequences that matter to a caller. A response watchdog sits at about 5.5 seconds, because dead air is worse than a filler. And barge-in is ignored for the first couple of seconds, so the greeting is not cut in half by a cough — see barge-in without cutting the greeting.
What each one is good for#
- Choose turn-based
- When something said on the call would be expensive to get wrong — commercial terms, regulated language, anything you must be able to guarantee rather than review. The check runs before the caller hears it.
- Choose realtime
- When conversation quality is the product: interruptions, overlapping speech, a caller who changes their mind mid-sentence. It sounds like a conversation because it is structured as one.
- Choose turn-based on cost grounds
- Audio tokens are billed at a multiple of text tokens in both directions; a long realtime call is not a cheap call.
- Choose realtime for languages and accents
- Speech-to-speech keeps prosody and code-mixing that a transcribe-then-speak pipeline flattens.
Failure surfaces are not the same#
- A failed model session ends a realtime call in a way a text-model failure on the turn-based path does not; the call is attributed rather than recorded as a success.
- Prompt size is paid per call on both, but it lands hardest on the realtime path where the caller is listening to the delay: about 7,500 characters of instructions gave a first token in 1.2–1.8 seconds, about 9,600 gave 2.3–3.4.
- Worker capacity bounds concurrent realtime calls; a line at capacity says so through line health rather than answering into silence.
- A hard ceiling of 30 minutes applies to a call, on either engine.
- Rules become counts.
price_spokencounts lines the voice priced using the same expression the carrier path checks each sentence with, so the two paths at least agree on what a price is.
Questions#
Can I run both?
The realtime engine is something a workspace opts into, so the turn-based carrier path is what runs otherwise. Treat the choice as a property of the line rather than something to switch mid-call.
Will the realtime engine get under two seconds?
This documentation makes no prediction about future performance. What is measured today is a 3.3-second median on the best call, with about 60% of it model-bound — and the honest planning assumption is the measurement, not a hoped-for number.
Which is safer for a business that quotes prices?
Turn-based, unambiguously. The refusal to invent commercial terms is enforced there before a sentence is spoken; on the realtime path the same rule lives in the instructions and a breach is detected afterwards. See rules and instructions.