The secret of the 1-second caption — the latency we measured
How OpenAI Realtime Translate plus a Socket.IO relay holds average latency at 0.8 seconds, including what happens when the connection drops.
'Captions within a second' isn't marketing copy — it's a measured number. From the speaker's mouth to a caption line on the LED screen takes 0.8 seconds on average. We measure it minute by minute at the events we run every week, and log it in the operator manual.
Where the 0.8 seconds goes
- OpenAI Realtime API, speech → text → translation: 0.55s on average
- The broadcast server (Node + Socket.IO) fanning out to on-site devices: 0.15s on average
- Rendering on attendee phones and the LED laptop: 0.10s on average
When the connection drops
If the venue Wi-Fi gets shaky, the operator laptop holds the last 30 seconds of captions in a local cache. When the line comes back, a backoff reconnect restores everything automatically; Korean spoken during the outage is dropped (the caption queue restarts from zero). Across a year of live events, the line has been down for more than a minute exactly once.
What's next
We're experimenting with a Deepgram Nova-3 STT + Claude Haiku translation pipeline to bring the average down to 0.5 seconds. Pushing captions out faster is technically possible, but the threshold at which audiences perceive captions as 'natural' is about one second — so within that budget, we prioritize stability.