Church translation quality checklist: test, review and improve AI captions
A practical way to check AI sermon translation: a 10-minute test, bilingual reviewer roles, an error log that separates meaning from style, glossary upkeep, speaker habits and when to call an interpreter.

How do you check the quality of AI church translation?
Check the quality of AI church translation with a short, repeatable routine: run a 10-minute test with a script full of names, references and idioms, have a fluent bilingual reviewer read the captions for each language, log errors by whether they change the meaning, then fix causes in order: audio first, then glossary and slides, then speaker habits. Escalate to a human interpreter when meaning-changing errors keep returning.
Machine translation will make mistakes, and it doesn't need to be flawless to be useful. What you need is to know where it fails for your preacher and your languages, and to catch the errors that matter before a guest acts on them.
Run a 10-minute test before you rely on it
Test with your real preacher, real languages and the real audio path you will use on Sunday. A laptop microphone test tells you little about a mixer feed, and the reverse is also true. Here is a test script that packs the usual trouble spots into one minute:
"Good morning, and welcome to [your church name]. Please turn with me to Second Corinthians chapter five, verses seventeen to twenty-one. Pastor Adaeze Okafor will lead communion, and Deacon Nguyen Van Minh will pray for us. Our marriage course starts on Wednesday the 14th at 7 p.m. in the fellowship hall. Don't worry if you feel like a fish out of water at first. As the old hymn says, I once was lost, but now am found."
- Minutes 0-1: start the room with your real listener languages. Reviewers open the listener page on their own phones, from ordinary seats.
- Minutes 1-3: the preacher reads the script at normal pace.
- Minutes 3-6: the preacher speaks unscripted about last week's sermon, as naturally as possible.
- Minutes 6-8: stress cases: faster speech, turning away from the mic, a second person interrupting, and a few bars of music so the operator practises pausing.
- Minutes 8-10: each reviewer writes down the three biggest problems in their language, using the error log format below.
The test is cheap. Ten minutes of speech in three languages is 10 × 3 ÷ 60 = 0.5 translation hours, so the 4 free hours that come with a new SermonLingo account cover up to eight full test runs.
Who reviews, and what does each person check?
- Language reviewer (one per language). Fluent in both languages, ideally from the community that reads the captions. Reads on a phone during the service and logs errors with the time and severity. Does not try to fix anything live.
- Theology reader. Someone who knows how your tradition says words like grace, repentance, communion and baptism in that language. Often the same person as the language reviewer, sometimes a pastor from a partner congregation.
- Single-language reader. A native speaker who does not understand the preacher's language. This role matters more than it looks: bilingual reviewers quietly fill gaps from the original speech they heard, while this reader shows whether the captions make sense on their own.
- Glossary keeper. Collects names and terms from the reviewers and the sermon notes, and keeps the glossary or the slide list current.
Treat reviewers as volunteers with a real ministry, not as proofreaders on call. Send them the sermon notes ahead of time, rotate the job monthly so nobody carries it alone, and make it clear that their notes are about the system, never a grade on the preacher.
Keep an error log that separates meaning from style
Sort every error into one of three levels, because they need different responses:
- Level 1, meaning-changing: a dropped negation (so "God will not abandon you" loses its "not"), a wrong number, date or time, a wrong scripture book or chapter, a name turned into an ordinary word, or an idiom translated literally in a way that misleads. Fix the cause before next Sunday.
- Level 2, awkward but understandable: unnatural word order, a formality level that feels off in languages that mark respect in grammar, or a church term rendered with an everyday word. Batch these for a monthly review.
- Level 3, cosmetic: punctuation, capitalization, a sentence split in an odd place. Ignore unless they pile up.
| Time | Language | What was said | What appeared | Level | Likely cause and fix |
|---|---|---|---|---|---|
| 10:42 | Spanish | "Pastor Joy Adams will pray" | The word for joy, not the name | 1 | Name heard as a common word: add to glossary, show on slide |
| 11:05 | Portuguese | "Romans 8:28" said quickly | Wrong chapter number | 1 | Fast reference: say "chapter 8, verse 28" and show it on screen |
| 11:12 | Korean | A warm aside to the congregation | Ending style felt too casual | 2 | Formality: raise at the monthly review with Korean reviewers |
On Advanced and Premium plans, transcripts in every language can be exported as TXT, CSV, SRT or WebVTT. The SRT and WebVTT files carry timestamps, so a reviewer can jump from a log entry to the exact moment and read the source text beside the translation at home, rather than squinting at a phone mid-sermon.
How do you keep a glossary useful?
A SermonLingo glossary (on every plan) teaches the system the names, scripture books and ministry vocabulary your church actually uses. It works best as a small, living list rather than a one-time dump.
- Add: staff and lay leaders who are named from the platform, guest speakers, ministry and course names, local place names, sermon series titles and any terms your tradition uses distinctively.
- Weekly: when the sermon notes arrive, add new names before Sunday. After the service, add anything from the Level 1 log.
- Monthly: remove people who have moved on and check entries with each language reviewer.
- Privacy: include only names that are spoken publicly. Never put prayer needs or other personal details into a glossary.
Ask each language group which Bible translation they read. Many Spanish-speaking congregations read the Reina-Valera 1960, for example, but ask yours rather than assuming. AI translation renders what the preacher says; it does not quote a particular Bible version, so when exact verse wording matters, put the verse on a slide in the version your community knows. Sermon notes help with the rest: paste the week's outline and passages into room settings before the service so the translator recognizes the references and names it will hear.
Which speaker habits make translation better?
| Instead of | Try |
|---|---|
| "Turn with me to that famous passage" | "Turn to Romans, chapter 8, verse 28" |
| "First John three sixteen," said fast | "First John, chapter 3, verse 16" |
| "That sermon series knocked it out of the park" | "That sermon series helped a lot of people" |
| Starting a sentence, dropping it, starting another | Finishing each sentence, then pausing |
| A joke that hangs on a pun | Telling the story and skipping the pun, or explaining it |
| Talking over an interview guest | One voice at a time, with questions from the floor repeated into the mic |
None of this means preaching like a robot. Passion, stories and humor still land. Most preachers find that two habits, full scripture references and finished sentences, fix a large share of their Level 1 errors.
When should you bring in a human interpreter?
AI captions are a good fit for the sermon, announcements and readings in many churches. Bring in a human interpreter, whom you recruit or hire yourself, when:
- Level 1 errors keep recurring in one language after you have fixed the audio, the glossary and the speaker habits.
- The moment carries spiritual, legal or pastoral weight: membership vows, wedding vows, baptism questions, congregational votes, safeguarding information, prayer ministry and altar calls.
- The content is poetic or sung, such as song lyrics, a psalm read as poetry or spoken-word pieces.
- A language community tells you captions don't serve them, for example listeners who prefer to listen rather than read in a language that has captions but no AI voice.
- Deaf people who sign attend: they need a sign language interpreter, not captions.
Many churches settle on a mix: AI captions for the sermon and a bilingual volunteer for the prayer team and altar ministry. Our comparison of volunteer interpreters and AI translation covers that decision, and how AI church translation works explains why names and idioms are the usual weak spots. If audio keeps showing up as the cause in your log, start with our church audio setup guide.
A 20-minute monthly quality review
- Count Level 1 errors per language for the month, and note whether the trend is up or down.
- Name the top three causes (audio, names, fast references, idioms, crosstalk) and assign one fix to each.
- Review the batch of Level 2 notes with each language reviewer.
- Decide whether any language or any part of the service needs a human interpreter.
- Thank your reviewers, in person.
Transcripts arrive with the Advanced plan ($299 a month with 100 hours); the glossary and sermon notes come with every plan. A church translating 60 minutes of speech into three languages every Sunday uses about 12 translation hours a month (4 × 60 × 3 ÷ 60), which the Standard plan's 25 hours would cover; upgrading is therefore a decision about review tools such as transcripts, not hours. Compare the tiers on the pricing page, read a scripted service on the live demo, or start free and run your first 10-minute test this week.
Questions
How many bilingual reviewers does a church need to check translation quality?
Start with one fluent reviewer per caption language, plus one native speaker who isn't bilingual and reads only the captions. The bilingual reviewer spots errors against what was said, while the single-language reader shows whether the captions make sense on their own. Rotate reviewers monthly so the job stays light.
What counts as a meaning-changing translation error in church?
A meaning-changing error is one that would lead a reader to believe or do something different: a dropped negation, a wrong date or time, a wrong scripture reference, or a person's name turned into an ordinary word. Log these as top priority and fix their causes before the next service, ahead of awkward wording.
Can a glossary fix every mistranslated name?
No. A SermonLingo glossary, available on every plan, helps the system handle names, scripture books and ministry terms, but it cannot fix audio it never heard clearly. Names said quickly, mumbled or spoken over music can still go wrong, so pair the glossary with a clean mic feed and names on slides.
When should a church use a human interpreter instead of AI captions?
Use a human interpreter for moments with spiritual, legal or pastoral weight, such as vows, membership commitments, baptism questions and prayer ministry, and whenever meaning-changing errors keep recurring in one language after audio and glossary fixes. Deaf signers need a sign language interpreter. Many churches pair AI sermon captions with an interpreter for those moments.






