
Nehemiah 8:8 describes an interpreted Bible reading: the Levites “read in the book in the law of God distinctly, and gave the sense, and caused them to understand the reading.” Live AI translation now sells a version of that work to any church with a microphone, and vendors advertise 100 or more languages delivered to phones through a QR code. The access is real, and Pentecost blesses it. The sense is where the risk sits, and for sermons preached in Luganda, Acholi or Runyankole the published accuracy evidence is thin and dated, so a church should adopt these tools rung by rung and keep a human sense-giver wherever a wrong word would wound.
Key Takeaways
- Live AI sermon translation is a chain of three machines: speech recognition in the preacher’s language, text translation, then captions or a synthetic voice in the listener’s language. Each link can fail on its own.
- The tools are strongest out of English. In 2022 spf.io reported speech recognition for 60 languages against translation for more than 100, so the language the pastor preaches in sets the ceiling.
- The published Luganda tests are 2022 research baselines: a 33% word error rate for Luganda speech recognition, and a documented tendency to turn Luganda’s gender-neutral pronouns into “he” or “she” in English.
- Scripture settles the access question in favor of the mother tongue (Acts 2:8, 1 Corinthians 14:19). Accuracy is a separate question, and it belongs to the church, whatever the vendor promises.
- The Sense-Giving Ladder sorts a service into four rungs, from captions any church can automate this month to vows, confessions and counsel that keep a named human interpreter.
- The interpreter beside the pulpit is a ministry and a training path. A church that automates the role should decide what replaces the formation as well as the voice.
What a live translation tool does between the pulpit and the phone
The products look alike under different names. Wordly, spf.io, Maestra and Glossa all take the preacher’s microphone feed, turn speech into text, translate the text, and push captions or a synthetic voice to phones in the room and on the livestream. Maestra says its church system works live in more than 100 languages and offers listeners the speaker’s cloned voice or a stock AI voice. Glossa markets itself as “AI trained on biblical texts.”
Adoption is early. Wordly told Religion News Service in March 2025 that about 200 U.S. houses of worship used its service, mostly churches, paying per minute across some 60 languages. Wordly told reporters the price varies with users and hours, and the same report put a small church’s annual spend under $5,000. Set that beside a paid interpreter for every language group and the appeal is obvious.
Vendor results are vendor-reported, and we should read them that way. Wordly says Christ Community Church saw 900 attendees use its tool in 36 languages in the first two months. The Timberline Church case in Colorado is better attested because a reporter spoke to the man in the pew. Miguel Flores Robles, a Spanish-speaking drummer in the worship band, told RNS: “It was the reason I feel more at ease in this place, because I can understand the preaching.”
The language menus are lopsided: output lists are long, and input lists are shorter. A Kampala pastor preaching in English can reach Swahili, French or Arabic listeners today, and Wordly lists Swahili among its supported languages. A pastor preaching in Luganda first needs a machine that hears Luganda well, and the evidence for that is much thinner.
Pentecost already answered the access question
Start with the blessing, because the impulse behind these tools is good. At Pentecost the crowd asks, “how is it that each of us hears them in our native language?” (Acts 2:8). The Spirit could have given the whole crowd Aramaic. He gave Parthians, Medes and Elamites the wonders of God in their own speech instead. Paul then makes intelligibility a rule for the gathered church: “in the church I would rather speak five intelligible words to instruct others than ten thousand words in a tongue” (1 Corinthians 14:19). And the story ends in many languages at once, with a multitude “from every nation, tribe, people and language” before the throne (Revelation 7:9).
The church has carried Scripture into mother tongues for centuries, and the engineers building speech models for Luganda and Acholi continue that work, knowingly or otherwise, and the case for AI Bible translation in African languages rests on the same ground. So the builders deserve thanks first. A deaf member reading live captions, a Congolese family following an English sermon in French, a visiting engineer hearing it in Mandarin: each is a neighbor the church can now reach on an ordinary Sunday.
Where the nuance breaks in Luganda, Acholi and Runyankole
Uganda has over 43 indigenous languages and dialects, and Luganda, spoken by more than six million people, is among the better resourced, according to a 2022 Makerere University paper. Better resourced is relative. A companion 2022 Makerere paper trained a Luganda speech recognition model on 300 hours of Mozilla Common Voice recordings and reported a word error rate of 33%. Put plainly, about one word in three came out wrong at the first link of the chain, before any translation happened.
Makerere’s baseline translation models scored BLEU 37 from English to Luganda and 36.8 back, while a related paper reported 26.0 and 24.6 on a different setup. BLEU measures overlap with a reference translation and is hard to compare across test sets, and neither figure tells a pastor how a sermon will land in the ear. The data underneath is uneven too: Makerere’s open dataset, as described by the Lacuna Fund, holds 100,000 Luganda sentences and 40,037 Acoli sentences.
These are research baselines from 2022, and Ugandan and commercial systems have likely moved since. Sunbird AI’s SALT corpus covers Acholi, Lugbara, Luganda, Runyankole and Ateso and sat among the resources for the 2023 WMT shared task on African machine translation, and its Sunflower models were built for Ugandan languages. I looked for a published, independent test of live sermon translation out of Luganda, Runyankole or Acholi and found none. A church buying a tool today buys it on the vendor’s word plus its own testing. We prefer to say that plainly.
Nuance has a second, sharper edge. Luganda pronouns carry no gender, and English pronouns do. The Makerere gender-bias study found “a heavy tendency” for Luganda-to-English systems to produce gendered outputs for gender-neutral pronouns, with doctors drifting male and nurses female in its examples. Picture, as an illustration only, a Luganda sermon on Lydia of Thyatira reaching the English captions with Lydia as “he” for a paragraph. That looks small until the sermon is about women in leadership, and then the machine has taken a side the preacher never took.
Many Kampala pulpits also move between English and Luganda inside one sentence. Before you buy, ask how the tool handles a speaker who switches languages mid-thought, and test it on your own preacher.
The strongest case against AI in the pulpit
Kutter Callaway, associate professor of theology and culture at Fuller Theological Seminary and among the faculty spearheading its guidance on AI, made the case to RNS at full strength. Christianity is “incarnational,” he said: “Our bodies in space and time matter.” On accuracy he went further: “Christian theology historically is basically a whole bunch of Christians arguing about words,” and those arguments can change a church’s understanding of who Jesus was. Katalina Tahaafe-Williams, a multicultural ministry expert who speaks Tongan, added a warning we should hear in Uganda: her community wants the next generation fluent in its tongue and wants input in any use of it, and she fears a tool that de-skills migrants and Indigenous peoples.
We concede most of it. Homoousios and homoiousios differ by one letter, and councils divided over the difference. A sermon is a pastor bearing witness before people who know him. A translation engine bears no witness and can answer for nothing it says. If the question were whether a machine can carry the whole service, Callaway would win.
The question in a Kampala church is narrower. For the Congolese family in row nine, the realistic alternative to a machine may be silence. Callaway himself told RNS he could imagine AI translation “for the occasional larger event or for educational meetings.” His objection tells us where to draw the line, and it leaves the line standing.
The Sense-Giving Ladder
We call the working tool the Sense-Giving Ladder, after Nehemiah 8:8. It has four rungs, and a church climbs only as far as its own testing allows.
- Captions in the preacher’s own language. Live transcription for deaf and hard-of-hearing members, plus a transcript for the week. Members can check the words against what they hear, so the risk is low. Automate it now.
- Out of English, into well-supported languages. Swahili, French, Arabic or Mandarin for visitors and migrants. Automate with two guardrails: a glossary of names, places and doctrinal terms loaded before Sunday, and a bilingual member who scores one recorded sermon each month. Timberline did this, asking bilingual members to test Wordly’s output and using its glossary to fix specific terms.
- Out of Ugandan languages, or into them. The human interpreter leads. The machine supplies backup captions, labeled on screen as machine output. Promote it only after a scored test on your own preacher’s voice.
- The words that bind. Baptismal confession, wedding vows, the words of institution at communion, the altar call, prayer ministry, and anything said to one person in distress. A named human interpreter, always, because these words create obligations and someone must answer for them.
The scored test is simple. Record ten minutes of last Sunday’s sermon and run it through the tool. Two bilingual members mark the output separately for three faults: wrong meaning, wrong person (including he and she flips), and wrong tone. Compare their marks, agree a count, and write down the threshold your elders will accept for each rung. The number is yours to set. Writing it down is what turns a feeling about the tool into a decision.
The interpreter beside the pulpit is a ministry
Briana Southerland, Share Ministries Pastor at First Alliance Church in Calgary, put the gain this way in a testimonial on the spf.io site: “Before, you needed a human translator per language group, but now you need one operator for 60 languages.” For a stretched volunteer team that is a real gift, and she is right to celebrate it.
The idol arrives one step later, when reach becomes the only measure and the interpreter’s chair is booked as a cost line. In many East African churches the person standing next to the preacher, carrying the sermon into Luganda phrase by phrase, is a deacon or a young minister learning to preach by bearing someone else’s words first. The Levites at the Water Gate were teachers among the people, and Nehemiah 8:7 records their names. A machine renders words. A named interpreter can be prayed over, corrected after the service, and one day handed the pulpit.
So here is the decision for this week. Before anyone signs a subscription, the senior pastor and the media lead sit with last Sunday’s recording and sort its minutes into the four rungs. Then they run the scored test on one rung-two segment and one rung-three segment. Keep the interpreter on rungs three and four, and give that person a new title: translation reviewer, owner of the glossary and the monthly score. Automation then becomes a promotion for the person it seemed to threaten, and the church keeps someone accountable for the sense. The earlier piece on whether pastors should use AI for sermons drew the same boundary around authorship; this one draws it around interpretation.
Frequently Asked Questions
Is it biblical to use AI to translate a sermon?
Scripture favors hearing God’s word in your own language. Pentecost delivered the gospel in the crowd’s native tongues (Acts 2:8), and Paul required intelligible speech in worship (1 Corinthians 14:19). The tool itself is permissible. The church still owes its people accuracy, so it should test the tool and keep human interpreters for the words that bind.
How accurate is live AI translation for Luganda sermons?
Published evidence is limited. 2022 Makerere research recorded a 33% word error rate for a baseline Luganda speech recognition model and a heavy tendency to assign gender to Luganda’s gender-neutral pronouns in English. Systems have likely moved since, but no independent sermon benchmark was found, so each church should run its own scored test.
How much does live AI sermon translation cost?
Most vendors charge by the minute or by subscription. A 2025 report on Wordly said the price varies with users and hours, and that a small church could expect to spend under $5,000 a year. Budget also for volunteer time: a glossary to maintain and one bilingual reviewer who scores a recorded sermon each month.
Should a church replace its human interpreter with AI?
Automate captions and translation out of English first, and keep the human interpreter for Ugandan-language preaching and for vows, confessions, communion and counsel. Where your interpreter is a minister in training, give the role a new task: owning the glossary and scoring the tool each month.
What is the Sense-Giving Ladder?
It is a four-rung method named after Nehemiah 8:8, where the Levites “gave the sense” of the Law. Rung one is same-language captions, rung two is translation out of English, rung three is translation out of Ugandan languages with a human lead, and rung four is binding words, which always keep a named human interpreter.
Next Step
If you lead a church or a church network setting a translation policy before 2027 budgets close, book a strategy call with Joshua this month. Bring last Sunday’s recording and your rung sort. We will help you set test thresholds and a reviewer role before a contract sets them for you, which saves a year of guessing.
Related Reading
- Should Pastors Use AI for Sermons? A Third Way
- AI Bible Translation: Africa’s Languages Are the Frontier
- Sunflower: Uganda Built Its Own Local-Language AI
- AI Misquotes the Bible: A Theology of Truth-Telling
Sources and Evidence
- Power of Babel: Real-time AI translation may be coming to church near you (Religion News Service, 18 March 2025): about 200 U.S. houses of worship on Wordly, per-minute pricing across some 60 languages, the Flores Robles quote, Timberline’s bilingual testing and glossary, and the Callaway and Tahaafe-Williams objections.
- Power of Babel, as republished by The Roys Report: the report’s estimate that a small church spends under $5,000 a year, with price varying by users and hours.
- Translation Services for Churches (Wordly blog): vendor-reported 900 attendees in 36 languages at Christ Community Church; Swahili in the supported list; custom glossary support.
- Translation in church technology: new tools for belonging (spf.io, 7 October 2022): speech recognition for 60 languages against translation for more than 100.
- Church Translation and Accessibility (spf.io): the Briana Southerland quote on one operator for 60 languages.
- Church Translation System (Maestra): live translation in 100+ languages, 125+ after saving, cloned or stock AI voices.
- Glossa: AI Church Translation (App Store listing): the “AI trained on biblical texts” claim.
- Gender Bias Evaluation in Luganda-English Machine Translation (Wairagala et al., AMTA 2022): Uganda’s 43+ languages and dialects, Luganda’s six million speakers, gender-neutral pronouns, the heavy tendency to gendered output, BLEU 26.0 and 24.6.
- Building Text and Speech Datasets for Low Resourced Languages: A Case of Languages in East Africa (AfricaNLP at ICLR 2022): Luganda speech recognition at 33% word error rate on 300 hours of Common Voice data; BLEU 37 and 36.8.
- Language datasets (Lacuna Fund): Makerere dataset of 100,000 Luganda and 40,037 Acoli sentences.
- Shared Task: Large-Scale Machine Translation Evaluation for African Languages (WMT 2023): Sunbird’s SALT corpus covering Acholi, Lugbara, Luganda, Runyankole and Ateso; Luganda among evaluated languages.
