If live translated captions appear out of sync, first check which sound they follow: the live stage feed or the delayed stream that viewers hear. A separate caption link and a video player can arrive at different times even when both receive the correct content.
This guide is for organisers covering a venue and an online audience together. It is a documentation review and an original rehearsal worksheet, not a measurement of any translation service’s latency. Sources were checked on 17 October 2026.
Start with the viewer’s playback, not the production monitor
Write down the two paths before changing settings:
- Video path: speaker → production feed → streaming platform → viewer’s player.
- Translation path: selected audio input → transcription and translation → guest captions or translated audio.
The second path might start before or after the stream’s delay. Ask the operator to name the actual input. A microphone beside the stage and audio captured from the public player are different starting points.
Check a remote viewer’s actual player and caption page together. A clear result at the venue’s production desk does not establish what that viewer receives. For the separate task of checking which microphones reach translation, use the audio input routing worksheet.
What the published delay figures mean
YouTube defines stream latency as the time from camera capture to display for a viewer. Its guidance also warns that reducing latency can increase buffering.1 The figures below describe the named delivery route, not AI translation speed or a promise for your event.
| Documented route | Provider’s published figure | Planning use |
|---|---|---|
| Zoom meeting or webinar streamed to YouTube | Approximately 20 seconds of delay | Check the public player separately from Zoom. |
| YouTube low-latency setting | Less than 10 seconds for most viewers | Rehearse the chosen stream configuration. |
| YouTube ultra-low-latency setting | Less than 5 seconds for most viewers | Check buffering as well as responsiveness. |
Sources: Zoom’s YouTube livestream instructions2 and YouTube’s latency guidance.1 These are different contexts; do not add the rows together or substitute them for a measurement. A paused viewer, connection problem or changed setup needs another check.
Measure the caption offset in a clearly labelled rehearsal
Choose a short phrase with an identifiable last word. Have a consenting colleague capture both the viewer’s video and caption display in one recording, where permitted. Compare the end of the spoken phrase in that playback with the arrival of its corresponding translated caption. Translation may reorganise a sentence, so use a complete phrase rather than trying to align individual words.
Define the sign before recording results:
Caption offset = caption arrival time − corresponding speech time in the viewer’s playback.
- A negative result means the caption arrives first.
- A positive result means it arrives after the speech.
- This comparison says nothing by itself about whether the translation is correct.
For an invented arithmetic example, suppose the speaker finishes at elapsed time 0, the translated phrase appears at 3 seconds and the viewer hears the end of that phrase at 12 seconds. The caption offset is 3 − 12 = −9 seconds: it is nine seconds early relative to that viewer’s video. Those numbers are hypothetical, not product test results or recommended targets.
Record the method and any timing uncertainty. Repeat after a stream restart or input change, and test translated audio separately if you offer it. Do not assume that caption text and synthesised speech arrive together.
Copy this room-and-stream rehearsal sheet
Use one row per viewing route. Record the device, browser, network and player position above the table. Leave results blank until someone checks them.
| Viewing route | What feeds translation? | Caption offset / result | Owner and next action |
|---|---|---|---|
| Guest listening in the venue | ___ | ___ | ___ |
| Participant inside the meeting | ___ | ___ | ___ |
| Viewer on the public livestream | ___ | ___ | ___ |
| Viewer after pause or reconnection | ___ | ___ | ___ |
If one route is acceptable and another is not, identify that difference in the event instructions. A single guest link does not establish synchronisation with every player.
Choose a fix that the delivery system actually supports
Ask the production supplier whether translation can follow the appropriate playback feed, or whether the caption or media route has a documented delay control. Agree who owns that control and repeat the viewer check after using it. An independent translation service is not automatically an encoder, subtitle track or playback synchronisation system.
Where supported and licensed, an integrated event player may be the simpler route. Zoom Events, for example, documents translated-caption selection in its event-lobby livestream window, with stated licence requirements.3 This is a specific option, not proof that every Zoom-to-YouTube stream carries those features.
For a Hong Kong event using Cantonese on stage, tlive’s presenter and guest flow delivers captions and enabled translated voices through a separate guest page; the documented room limit is the source language plus at most two translation languages. Treat that page and the broadcast as separate paths to rehearse, without assuming automatic player synchronisation. See the hybrid-event use case for joining arrangements.4
If guests need a tightly controlled interpreted audio programme, ask an AV supplier and professional interpreters to design the delivery together. For legal, medical or negotiation content, arrange qualified human interpretation rather than relying on an unattended AI output. The timing worksheet does not assess professional suitability or translation accuracy.
Give the host an instruction they can use
Before opening the room, write down which guest routes passed, which remain unsupported, and who can stop or change each feed. Have the host announce the confirmed viewing and listening arrangement, including earphones where needed. Do not advertise a route that the rehearsal could not verify.
A usable handover might read: “Venue guests use ___. Online viewers use ___. If captions no longer match the current speaker, contact ___; the operator checks the viewer feed before changing settings.” Fill the blanks with the actual arrangement, not a generic product promise.