Keep your interpreter and add AI for the other languages

You do not have to choose, and you should not have to run two systems to avoid choosing. The arrangement that works is one audio feed off the desk, one QR code for the whole room, and a list of languages decided per service — your interpreter on the language he has always carried, the machine on the ones nobody in the building can cover.

Most churches arrive at this question from the wrong end. They go looking for a tool, find that every vendor's pitch is written as a replacement for the interpreter, and conclude that adopting anything means telling a faithful volunteer his seat is gone. That is a real fear and it is worth naming out loud. But the operational problem underneath it is smaller and more boring: two translation arrangements in one room usually means two sets of instructions from the front, two things that can fail, and two bills. If you are already past that and weighing actual options, the things worth putting side by side are latency, voice, glossary and handling rather than whether a tool replaces a person.

Can a human interpreter and AI translation run in the same service?

Yes, and the part that decides whether it is pleasant or miserable is how many things the congregation has to be told.

If the interpreter uses a transmitter and a box of receivers while the machine lives on people's phones, you have two announcements, two queues after the service, and a visitor who guessed wrong. If both live on the phone behind the same code, you have one sentence from the front: scan this, pick your language.

That second arrangement is what we have been building toward and are still refining. In practice a service is created with its language list, one of those languages is marked as carried by a human interpreter, and the interpreter opens his own link — phone or laptop, headset on — picks the microphone he is speaking into and presses start. The congregation scans the single code on the screen, and whoever chooses his language hears his voice. Every other language on the list is translated as before, under a second behind the speaker, in a natural voice that keeps the preacher's tone and rhythm. No second app for anyone, no second code, no second announcement.

I want to be straight about the status of this: the interpreter channel is the newest part of what we run, and we are still testing and refining it. The decision it supports — keep the interpreter, add the machine around him — is the settled part. The plumbing is young.

What the interpreter actually needs from the room

Three things, and none of them is a product feature.

A headset, closed enough that he hears the preacher and not himself. Interpreters who work off a loudspeaker drift into the room's delay and start stepping on the speaker's sentences. Thirty euros of closed headphones fixes more than any software setting.

A seat where he can talk without being heard by the people next to him. The back row of a side gallery, the sound booth, a side room with a monitor feed. In a small room the honest answer is often a chair beside the desk, or a corridor with the door open and the service on a phone.

And a clean source of the preacher's voice, which is the same requirement the machine has. The capture device wants to be within about a metre of the preacher or next to a loudspeaker that is fed from the desk — not a phone propped on the front pew hearing the room, and not a desk output that someone mutes between songs. If you take nothing else from this page, take the habit of looking at the input level before the service starts rather than after someone complains. We wrote the longer version of that in which audio signal to feed a translation system.

Which language goes to the human and which to the machine

Put the human where the relationship is. In most congregations the person interpreting into Farsi or Ukrainian is also the person who visits that family, prays with them and knows what the preacher means when he gets ironic. That is not a translation task you should want to automate, and the group in question will tell you so.

Put the machine where there is nobody. A church with three Polish students, two Armenian families and a Nigerian couple does not have a rota problem, it has an impossibility. Two interpreters is a rota; five is a fantasy. The machine speaks the fifth language as readily as the first and does not get tired at minute thirty-five of the sermon. If the machine half is the unfamiliar part, what AI translation actually does in a service covers what the listener hears and where it slips.

And treat the split as something that can change over a season rather than a settled policy. A language can start on the machine and move to a person the month a fluent member is ready to carry it; it can move back when that person's job changes or he goes home for the summer. A language list decided per service is what makes that a two-minute decision instead of a project. There is a fuller map of the choice in when a church should keep its interpreter.

Do we still pay for the interpreter's equipment?

This is the question that actually sits behind the whole conversation, and it has a dull, good answer: you stop buying receivers.

A transmitter and twenty receivers is the cost most churches are trying not to repeat, along with the charging, the lost earpieces and the volunteer who hands them out. We laid out the three arrangements side by side in translation equipment for a church: headsets, interpreters, AI. If the interpreter's voice reaches people on their own phones behind the same code as the machine's languages, that hardware stops growing. You will still want a stock of cheap wired earphones in a basket by the door, because the person who needs translation most is usually the one who came without any — but that is a basket, not an installation. We counted the real per-Sunday arithmetic of this in what a translated service costs per Sunday.

What to do this Sunday

Before the service, check two levels instead of one: the feed the machine is hearing, and the interpreter's own microphone. Both should show quiet passages still moving, not a flat line that is technically connected and carrying nothing.

During the worship set, decide deliberately whether translation is running at all. In Vienna we start it after the songs, because the half hour of singing does not need it and nobody wants to hear a sung line translated. Pausing cleanly matters more than pausing early — a pause that holds everyone's connection is a pause; one that drops the room and makes forty people rescan a code is an incident.

After the blessing, ask the interpreter one question: where did you have to guess? His answer is usually a name, a place or a term from your own church's vocabulary, and those are exactly the entries worth adding to an approved glossary so that both he and the machine say them the same way next week.

If you want to see what that looks like with your own languages and your own interpreter in the loop, we will set a service up with you and watch one Sunday together before you decide anything.