It is a fair question and it deserves an honest answer rather than a sales pitch. Machine translation has improved enormously and it is not magic. Knowing precisely where it is strong, where it fails, and — most importantly — how it fails is the difference between a congregation that trusts it and one that quietly stops using it.
What follows is based on our own measurements on real recordings from real rooms, not on demonstrations.
Where it does well
For clear, conversational speech — the bulk of most talks — modern translation is genuinely good. Ordinary sentences, explanations, illustrations and encouragement come through accurately and naturally in dozens of languages. For someone who would otherwise have understood nothing, the gap between "very good" and "perfect" is much less important than the gap between "very good" and "nothing".
The quality is also stable. It does not get tired at minute thirty-five, it does not lose concentration during the third point, and it does not have an off week.
The failure mode you need to understand
This is the single most important thing on this page, and almost nobody is told it.
When the audio is poor, speech recognition does not go silent or leave gaps. It invents. It produces confident, fluent, grammatical sentences that nobody said. In our own testing, feeding a system non-speech — room noise, a pause, a shuffle — produced fabricated sentences rather than nothing at all.
The consequence for a church is specific: a system failing obviously is safe, because everybody can see it. A system failing fluently is not, because the person reading in Ukrainian has no way to know the paragraph in front of them was never spoken. They will assume the preacher said it.
This is why every piece of practical advice below is really about audio.
Where it struggles
- Quiet or distant audio. By a wide margin the biggest cause of poor results, and the cause of the invention problem above. A microphone across the room is not good enough.
- Proper names. Places and people from scripture and tradition are where recognition most often trips, because they are exactly the words the model has least reason to expect. A mangled name is also the error a congregation notices most.
- Idioms and wordplay. A pun or a culture-specific turn of phrase rarely survives translation intact — in any language, by human or machine. "Bearing fruit" becomes agricultural in half a dozen languages.
- Overlapping speech. Two people at once, or a discussion, is far harder than a monologue.
- Rare language pairs. Translation between two widely-spoken languages is better than between two rare ones. This is worth saying plainly to a congregation that includes speakers of a smaller language.
What actually fixes it, measured
Three things make a measurable difference, and we can put numbers on two of them.
A vocabulary list. Supplying the religious and place-specific terms that will come up took correct renderings on a deliberately difficult recording from 46 of a fixed set to 67, with no term that had been right becoming wrong. In one language the improvement was from six correct out of ten to ten out of ten. On a clean recording the same list changed nothing at all — because there was nothing left to fix.
That last clause is the useful part: a word list rescues difficult audio and does not improve good audio. Which tells you where the real problem is.
Telling the system what languages are in the room. Without it, the model's guess space is unbounded and the results can be bizarre — in our tests one language came back rendered in the script of an entirely unrelated one. Giving the system the actual set of languages in play took a test from 18 correct out of 21 to 21 out of 21.
A warning attached to that, learned the hard way: naming a single expected language is worse than naming none. It biases the model into forcing everything toward that one, and it corrupted material that had previously been correct.
Clean audio from the desk. Everything above is secondary to this. A feed taken from your sound mixer rather than a device across the room is the single highest-value change available, and it costs very little.
How to judge it for yourself
Do not evaluate this by listening to a demonstration, and do not evaluate it with a synthetic voice reading a script. We tried that and it actively misled us: artificial speech is unnaturally clean, and it reversed the ranking we got from real recordings. A setting that looked helpful on synthetic audio turned out to be harmful on a real microphone.
The only test worth running is your own preacher, in your own building, on your own equipment, for five minutes. Then read the transcript — properly, not a skim. If it is accurate, you are fine. If it is fluent but wrong in places, that is your audio talking, not the software.
Set expectations honestly
Tell people what this is: a live, machine-assisted translation that follows the sense of what is said. Not a certified transcript, not a legal record, and not a substitute for a human interpreter where one is genuinely required.
Congregations are remarkably tolerant of an imperfect translation they were told about honestly, and remarkably unforgiving of one that was oversold. Saying "it gets names wrong sometimes" in advance costs nothing and buys you every subsequent mistake for free.
The honest bottom line
Machine translation will not replace a gifted human interpreter where the stakes are diplomatic, legal or contractual. For welcoming a visitor who would otherwise sit through an hour in silence, it clears the bar comfortably — instantly, in many languages at once, and at a cost that finally makes multilingual services possible for ordinary congregations.
Put the effort into the microphone and the word list, be honest about what it is, and it will do more good than almost anything else you could spend the same money on.
Try it at your next service
InMyTongue streams live subtitles and spoken translation to every attendee's phone, in the language they choose, straight from a browser with nothing to install. New accounts include free credit and no card is needed to try it. See how it works.