Church translation accuracy: when fluent is not faithful

When live translation fails in a service, it almost never fails the way a church expects. The audio stays clean, the captions stay readable, the voice keeps the preacher's pace, and the sentence that reaches the listener is perfectly fluent. What has changed is a single word inside it: a negation, or a condition. "If you do not forgive" arrives as "if you forgive". Nobody in the room who speaks only one of the two languages can hear that happen, and the sound desk has nothing to show for it, because nothing went wrong with the sound.

It is worth saying plainly, because it is true of every system on the market, ours included, and because churches discover it late. A congregation that tried AI translation and quietly dropped it usually did not drop it over robotic audio. It dropped it because somebody who spoke both languages finally sat down with a phone and heard the sermon say something the preacher had not said.

Why is this not an audio problem?

Because the audio chain is doing its job. In our own church in Vienna we run four languages on an ordinary Sunday, and when we listen right through a service rather than sampling it, the pattern holds: feed fine, text fine, voice fine, meaning in an instruction not fine. If your Sunday check is "is there sound, and does it sound like a person", you will pass a system that is inverting the application of the sermon.

It also means the first place a volunteer looks is the wrong place. A quiet feed, a muted desk output, a phone too far from the lectern: those are real faults and they announce themselves, because somebody's phone goes silent. A flipped condition announces nothing. No gap, no distortion, nothing to notice. The only instrument that detects it is a person who understands both languages.

Where does it show up first?

In exhortations, conditions and negations. Almost never in narrative.

That asymmetry is not bad luck, and it is worth understanding before you test anything. A story carries its meaning redundantly: if one clause lands a little loose, the next three repair it, and a listener reconstructs what was meant. An instruction carries its meaning in one word. Drop the "not", soften the "unless", turn "no longer" into "still", and the sentence remains grammatical, idiomatic and wrong.

So the damage is concentrated exactly where a sermon is supposed to do its work. The genealogy survives anything. The appeal at the end does not.

Why does one spot check give the wrong answer?

Because it is not a steady state, and this is the point we got wrong ourselves for a long time.

Listening through a whole service, a language does not sit at one level of faithfulness. It moves, in both directions, inside a quarter of an hour: clean through one passage, loose through the next, clean again after that. Two languages running from the same microphone move independently, so the Farsi can be solid while the Armenian slips, and fifteen minutes later they have swapped.

Two traps follow, and churches fall into both. Check once and you may get a clean reading and conclude the system is sound. Check once and catch a bad stretch, and you condemn a system that carries most of the sermon well. Neither church has learned anything, and both think they have.

How do you check live translation quality?

Ask one person in the congregation who speaks both languages to listen on their own phone, the way a listener does, and give them one narrow job: mark the imperatives. Not the whole sermon, not word choice, not style, not the bits that sound slightly stiff. Only the sentences that tell somebody to do something, or not to do something, and whether the "do" and the "don't" survived.

A slip of paper does it: one line per flagged sentence, roughly when it happened. Ten minutes at the start is not enough. Ask them to listen more than once across the service, and above all through the application rather than through the story. If you have one person and one stretch of attention, spend it on the last fifteen minutes.

Then read the slip with the preacher rather than with the tech team, because this is a meaning problem and not a wiring problem. Run the test on whatever you are considering, ours included, and run it again with a preacher whose first language is not the language of the service. That is a different test, and most systems behave differently for them.

What can a church actually control?

The vocabulary layer, and it is worth being honest about the limits. An approved list of names and terms — people, places, your church's own words, the recurring phrases a preacher uses — fixes the errors that come from guessing. A glossary a person on your team approves will stop your pastor's surname arriving as a common noun week after week. It will not stop an inverted condition. Those are two different faults with two different causes, and conflating them is how churches end up disappointed in a word list.

What the vocabulary layer really buys you is a clean signal. When names and terms are right, the one genuine error in a sermon stands out to your bilingual listener instead of being buried under forty small ones. That is what makes the check above workable at all.

The other half belongs to whoever is preaching. Finish sentences. Avoid the stacked double negative from the front, which is hard enough in one language. Say the instruction once in plain form before you ornament it. None of that is a concession to a machine; a visitor with ordinary German benefits from the same discipline.

What should you ask a vendor?

One question is worth more than a feature list: after the service, can we read what the listeners actually heard, in their language? A system that can show you its own output is one you can audit on a Tuesday with a bilingual member and a cup of coffee, over a whole sermon rather than ten sampled minutes. A system that cannot show you is asking you to take accuracy on trust, and accuracy is the one thing in this category nobody should take on trust.

Two more, true of any tool. Can someone on our team approve names and terms without filing a support ticket. And can we add or drop a language on the morning itself, because the week the need changes is the week you find out.

For us that readable record is the same recording everything else comes from: transcript, summary, full message, small-group guide and the drafts, inside the hour, in each language the church serves — which is also what makes a full audit possible instead of a sampled one. The names your team approves are held once and carried into every translation and draft after that. The longer version of how the live side works is here: AI translation in a church service, and the five things worth testing on one ordinary Sunday are here.

This Sunday, before the service, ask one bilingual person to sit with their phone through the last fifteen minutes and mark only the imperatives. Whatever system you are running, that slip of paper will tell you more than any specification sheet.