Live Captions by Subanana - Your whole audience follows, in their own language
by•
You speak once. Your audience scans a QR code, picks a language, and follows you live on their own phone — reading it, or hearing it. Put two languages on the big screen, or send a clean feed into OBS and vMix.
Replies
Is there any latency in multi-vendor routing compared to single-model applications, or does it occur instantly?
Congratulations @lakshya_singh & team @Subanana . 🚀
Subanana
@aymi_malik Good question, and the routing itself isn't the cost. We benchmark models per source language and per use case ahead of time, so a job goes straight to the one already picked for it — there's no vendor lookup while you wait. For uploaded files it's batch processing anyway, and for live sessions the model is fixed when the session starts, so nothing switches mid-event.
Where the real latency sits is the correction pass on live captions. From what I see running these myself: roughly half a second per sentence for the first pass, and around a second when the line gets enhanced using surrounding context and the event glossary. It stays inside the window where captions still feel live to the room.
That second one is the trade we chose deliberately. A caption that appears instantly and says the wrong name isn't much use at a real event.
Subanana
@aymi_malik Thanks! @aric_fung will be able to answer that better.