Live events stop at language.

In plain terms: the hard trade is between translation quality and delay — every moment spent getting a better translation is a moment your audience falls further behind the room. Every accuracy gain in this pipeline is bought with time. Wait for more audio and the recognition is better, the translation has more context, and the sentence is more likely to be right — and the listener is now far enough behind the speaker that they are watching a different event. Under about a second, the room stays together. Past a couple of seconds, the audience in the translated channel stops experiencing a live event and starts experiencing a delayed one, and they can hear the original in the room, which makes the gap worse rather than better.

So the pipeline is built against a latency budget, and every stage is allocated a slice of it. Recognition segments on natural boundaries where it can and forces a boundary where it must, because waiting indefinitely for a speaker to finish a sentence is how a live channel falls behind and never recovers. Translation is per language and independent — a slow language is a slow channel, never a slow broadcast. That isolation is the single most important structural decision in the system: without it, adding a fifth language degrades the other four, and the failure looks like the platform is bad rather than like one language pack is.

Then the languages themselves. The commercial pipelines are excellent at French and German and thin on Yorùbá, Igbo, and Hausa — which is precisely inverted from what a Nigerian broadcast needs. That gap is why interpreter-assist exists. A human interpreter working over a machine pass is doing correction rather than origination: a far lower cognitive load than booth interpreting, sustainable across a long service, and it puts a competent human exactly where the machine is weakest. The design admits what the models cannot do yet, rather than pretending.

The unglamorous half is the connection. Venues have the network they have. SRT exists because RTMP over a bad link degrades in ways an audience notices immediately, and SRT recovers packet loss at the cost of a little more latency — which then has to come out of a budget that was already tight. That trade is made per venue, not once in an architecture document.