Human speech
Keeping a live translation under a second, when every quality gain costs time.
Keeping live speech translation fast enough to follow, with every language on its own channel and a human placed exactly where the AI is weakest.
The problem
PolyCast
A room with four languages in it and one interpreter serving one of them. Everyone else is attending an event they can hear and cannot follow.
The technology to fix this existed, assembled for conference centres and priced for them. Nothing in the market was built for a Nigerian congregation, a Nairobi broadcast, or the languages either of them actually speak.
The engineering problem is a trade, not a feature: every accuracy gain in the pipeline is bought with time, and a live audience notices time.
The constraints
Fixed before anything was drawn.
Not requirements, which move. These are the facts of the problem that no amount of engineering was going to change, and every decision downstream had to survive them.
Under a second, or the room comes apart
Below roughly one second, the translated audience stays with the event. Past a couple of seconds they are watching something delayed — and they can hear the original in the room, which makes the lag worse than silence rather than better.
The venue's network is the network
Halls have the connection they have. Designing for the studio and deploying to the hall is how a broadcast fails in front of people.
The models are thin exactly where we need them
Commercial speech pipelines are strong in French and German, weak in Yorùbá, Igbo, and Hausa — precisely inverted from the requirement.
Adding a language must not cost the other languages
A shared pipeline means a fifth language degrades the other four, and the failure reads as 'the platform is bad'.
The architecture
What the constraints forced.
A latency budget, allocated per stage
The budget is the design input. Recognition, translation, and synthesis each get a slice, and every stage is built to hit its slice rather than to be as good as possible.
Forced segmentation
Recognition segments on natural boundaries where it can and forces one where it must. Waiting indefinitely for a speaker to finish a sentence is how a channel falls behind and never recovers.
Per-language isolation
Each target language is translated and voiced independently. A slow language is a slow channel, never a slow broadcast. This is the structural decision the system stands on.
Interpreter-assist over the machine pass
A human corrects rather than originates — far lower cognitive load than booth interpreting, sustainable across a long service, and positioned exactly where the models are weakest.
SRT where the link is bad
SRT recovers packet loss at the cost of latency, taken out of a budget that was already tight. The trade is made per venue, on the day.
The decisions
Each one cost something.
A decision presented without its cost is a decision nobody had to make. These are the ones we would defend, and what we gave up to take them.
Treat the latency budget as a hard constraint, not a goal
A pipeline optimised for transcript quality produces a channel nobody can watch live, which is the only thing it was for.
The trade-off
Individual sentences are sometimes less polished than a batch system would produce. In a live room that is the right side of the trade, and in a recording it is the wrong one — which is why the archive is per-channel rather than a re-render.
Build interpreter-assist instead of waiting for the models
The languages our audiences speak are the ones the models are worst at, and that gap is not closing on a schedule we can plan around.
The trade-off
It means a session that needs a strong Yorùbá channel needs a person, and we have to say so during scoping rather than after.
Isolate channels even though it costs more to run
Shared pipelines make every added language a tax on the existing ones.
The trade-off
Per-language cost is higher and scales linearly. The alternative scales worse in the way that matters — in front of an audience.
Where it stands
The platform is built and running: one live source, translated and voiced to independent per-language channels, with latency and channel health visible to the operator while the session runs.
The design admits what the models cannot do yet, and puts a human there, rather than shipping a Yorùbá channel that is fluent and wrong.
We build systems like this for other organizations.
PolyCast exists because we met the same operational failure enough times to engineer it properly. Yours is probably a different failure. That is the conversation we want, and you will have it with an engineer.