Is AI translation good enough for a church service?

Yes for the sermon, no for blind trust. AI translation now carries ordinary preaching cleanly and still slips on names, scripture references, quoted opponents and sung text. What separates a service that works every Sunday from one that embarrasses the preacher is a glossary of your names and a test in your own hall.

What “AI translation” means in a church

Three machines run in a chain. The first listens to the microphone and writes down what it hears. The second translates that text. The third either shows the translation as captions or speaks it as a voice. Every vendor on the market uses some version of this chain; the newest ones fold the first two steps into one model that hears speech and produces the translation directly, and some can keep a natural voice that keeps the preacher's tone and rhythm while doing so.

The chain matters because each link fails differently, and a church evaluating the result should know which link it is looking at. A wrong name is usually the listener, not the translator. A sentence that is grammatical but says the opposite of the preacher is the translator. A voice that sounds bored during the most urgent part of the sermon is the third machine.

What it now does well

• Ordinary spoken prose: a story, an explanation, an appeal. This is most of a sermon, and it comes through cleanly in the large languages.
• Long sentences that a human interpreter would have to compress. The machine does not tire in minute forty.
• The second, third and fourth language at once. Each one is a setting, not a person on the rota.
• Consistency: once it knows the church calls the building “the Hall” and the pastor by his first name, it says so every time.

Where it still slips, on a real sermon

Names, places and scripture references

Zechariah, Thessalonica, Nebuchadnezzar: the listening machine has heard these less often than it has heard words that sound like them, and it guesses. A reference such as “First Corinthians thirteen, verse four” can come out as a number sequence in the target language. This is the most common and the most visible failure, because a congregation knows exactly what the preacher said.

Quotation and irony

When the preacher quotes a doubter and then answers him, the machine translates both voices in the same tone, and the listener has to work out which one the preacher meant. A human interpreter changes their voice; the machine does not, yet.

Half a sentence

Live translation has to start speaking before the preacher has finished. German verbs arrive at the end of the clause, Farsi verbs likewise, and a translator that has to commit early sometimes commits to the wrong reading. Good systems wait for a natural pause; a preacher who leaves small pauses between thoughts gets a noticeably better translation, and nobody notices the pauses.

Songs, and the reading of liturgy

Sung text is not recognised reliably by any system we know of, and translating a hymn line by line produces nonsense. Most churches translate the sermon, the announcements and the prayers, and let the songs be songs.

What a glossary fixes, and what it does not

A glossary is a list the church keeps: this name is spelled this way, this phrase is translated that way, this word is left in the original. Every serious vendor has one under some name. It repairs the first failure almost entirely and the second not at all, so ask to see how the glossary is fed, who approves an entry, and whether the system learns from a corrected Sunday or forgets it.

A short note before the service does the rest: the passage, the names that will come up, the one word the preacher will use in an unusual sense. A system that can take such a note translates that sermon better than one that cannot, and the note takes two minutes.

A ten-minute test that beats a demo

• Play ten minutes of one of your own recorded sermons through the vendor's system, from your own laptop, in your own hall, with the room's echo.
• Have a native speaker of the target language listen with earphones and mark every place they would have corrected a human interpreter.
• Count the names and references. Note which ones came through and which did not, then add the failed ones to the glossary and play the same ten minutes again.
• Ask what the listener had to do to hear it. If the answer involves an app store, that is a finding too.

The second pass is the real result. A system that is wrong the first time and right the second is one you can live with. A system that is wrong both times is not ready for your church, whatever the demo sounded like.

How Sanctavo uses AI

Sanctavo runs the chain described above and, where a language allows it, the shorter form that speaks with a natural voice keeping the preacher's tone and rhythm. The congregation hears the translation as a spoken voice on their own phone, with captions alongside, in the browser and without an app. A per-church glossary holds your names and phrases, every entry approved by a person, and a note before the service tells the system what is coming.

The translation is machine-made and we say so; the four failures above apply to us as to everyone, and the glossary and the note are how we close them. We set it up free on one of your real services, run the ten-minute test with you, and let your own listeners give the verdict.