How Oya handles a meeting with five languages at once

The usual assumption in multilingual meeting tools is that there's a dominant language and everything else is an exception. One primary language for the meeting, one translation option for the minority speaker. That model breaks immediately in a genuinely multilingual team, where nobody shares a single dominant language and the meeting itself doesn't have a natural "base" tongue.

The per-participant architecture

MeetOye's translation is per-participant, not per-room. Each person selects their caption language before joining the call — independently of every other participant. A call with five participants can have five different caption languages simultaneously, and each person sees only their own. The system translates each speaker's audio into each active caption language independently, so there are no shared translation streams that require everyone to read the same second language.

What this looks like in practice

• Participant A speaks Urdu, reads Urdu captions — hears everyone else in Urdu.
• Participant B speaks English, reads English captions — the Urdu speaker's words appear in English.
• Participant C speaks Arabic, reads Arabic captions — both speakers above are translated into Arabic.
• Participant D speaks French, reads French captions — all other speakers translated into French.
• Participant E speaks Mandarin, reads Mandarin captions — all four other speakers translated into Mandarin.

What gets preserved

The original-language transcript is stored alongside all translations. If Participant A says something in Urdu that gets translated differently for different participants, the source record is always the Urdu utterance — not any of the translations. This matters for accuracy: if a translation is disputed, the original is available.

“Do not broadcast every translation to every participant. Only send each participant the caption language they selected.”

MEETOYE LIVE-TRANSLATION ARCHITECTURE

Performance at scale

Translation only runs for languages that participants have actively selected. A call with five participants and four different caption languages doesn't run twenty translation paths — it runs four. This keeps latency and compute cost proportional to actual usage, not worst-case theoretical load. Translated captions target a 2–4 second lag behind the speaker; original captions land in 1–2 seconds.