Churches spend a great deal of energy choosing a translation service and almost none on the thirty centimetres between the speaker's mouth and the microphone. That is backwards, and we can prove it with numbers rather than opinion.
Two services, one difference
On a single day we watched two live broadcasts from the same customer, running the same speech engine, minutes apart. One produced clean subtitles. The other produced garbage — wrong words, wrong language, sometimes nothing at all.
The engine was identical. The settings were identical. Here is what actually differed:
| Broadcast | Speech level reaching the system | Signal-to-noise | Result |
|---|---|---|---|
| First | 1500–4700 | 38 dB | 23 of 26 windows correct |
| Second | 260–710 | 25 dB | Garbage or empty |
The second speaker was roughly five times quieter at the microphone. That was the whole story. We tried other engines on the same recording and every one of them failed, because there is a floor below which no model can recover speech that was never properly captured.
The practical lesson: before blaming the AI, measure the level. If the audio arriving at your system is weak, changing providers changes nothing.
Why quiet audio produces nonsense rather than silence
You would expect a system that cannot hear to output nothing. Several widely used speech models do the opposite: handed audio with no clear speech in it, they invent a sentence — and often not even in the language you asked for.
We tested this deliberately. Eight windows containing no speech, submitted with the language set to Greek, came back as eight fabricated sentences in Japanese, Korean, Hindi, Arabic, Italian, Swedish and Romanian. Turning the model's randomness down to zero changed nothing: eight out of eight, either way.
In a sermon this matters enormously, because every pause is such a window. A speaker who stops to breathe hands the system a gap, and a badly guarded pipeline turns that gap into an invented sentence, translates it into every listener language, and reads it aloud on people's phones.
The fix is not a better model. It is a gate in front of the model that can tell speech from room noise — and that is a property of the service you choose, not of the engine underneath it.
Why a level meter lies to you
The obvious guard is a loudness threshold: if the audio is above X, treat it as speech. We shipped exactly that, and it failed for a specific and instructive reason.
A hall is never silent. Air conditioning, a PA hum, forty people breathing — that ambience has a steady level of its own. A fixed threshold sees the sum. Set it high enough to reject a noisy hall and you also reject a quiet speaker. Set it low enough to catch the quiet speaker and the fans alone will keep the engine busy transcribing nothing.
What works is looking at the shape rather than the level. Real speech has syllable peaks with pauses between them. Steady noise is flat, and it raises its own floor. Measuring in short frames against the window's own quietest tenth separates the two cleanly: speech at full, half and quarter loudness all pass; quiet, loud and very loud ambience and pure silence all produce exactly zero voiced frames.
You do not need to implement any of this. You need to know it exists, and to ask whoever runs your translation whether a pause in the sermon can reach the engine at all.
The seven habits that cost nothing
Everything above is on our side of the microphone. Here is the part that belongs to whoever is holding it — and it is worth more than any equipment purchase.
1. Get closer to the microphone
This single change moves you from the second row of that table to the first. A lapel microphone worn on the chest, or a headset, beats a lectern microphone the speaker walks away from. If you take only one thing from this article, take this one.
2. Finish your sentences
The most common problem is not accent or speed — it is the abandoned sentence. Starting one thought, breaking off, and starting another. A human listener repairs this without noticing. A translation system has to translate what was actually said, and what was actually said was half a sentence.
3. Pause where a full stop would go
A short pause at the end of a thought tells the system the sentence has landed and can be translated. A long pause in the middle of a clause does the opposite — it forces a guess about a sentence that has not finished. One habit, about a week to build, and everything downstream improves.
4. Name the reference before you read it
"In the second chapter of Philippians, Paul writes…" gives the system the context before the difficult words arrive. Reading straight into a passage cold is the hardest thing you can hand it, and it is exactly the moment your listeners care most about.
5. Say figurative points plainly, once
Idioms are the largest single source of confident nonsense. "Bear fruit" and "the ball is in your court" have equivalents in some languages and none at all in others. The output is not gibberish — which would at least be obvious — but a fluent sentence about something else. If a point rests on an image, say it once without the image too.
6. Announce a language switch
If you quote in the original language or greet the room in a second one, say so. A system told the speaker's language is listening for that language. Unannounced switches are where transcripts go strange, and we have watched a single unannounced Greek sentence get transcribed in Latin letters, mislabelled as Spanish, and translated from the wrong language.
7. Repeat the question
During questions, the microphone is at the front and the question came from the back. Repeat it before answering. This is ordinary courtesy for the hard-of-hearing people in the room, and it is the difference between a translated exchange and half of one.
Print this and give it to your speakers
Those seven lines are worth more to your translation quality than any technical setting, and they help human interpreters too. Put them where a guest speaker will see them.
Then check the level. If the speech reaching your system sits where the second broadcast did, no provider on earth will give you good subtitles — and now you know how to tell.
See it on your own audio
InMyTongue takes the clean feed from your existing sound desk and sends subtitles and spoken translation to every phone in the room. New accounts include free credit, so the honest test costs nothing. Try it on a real service.