How sermon transcription works, from the recording to the transcript.

Between “upload” and “here’s your transcript,” most of us are happy to treat sermon transcription as a black box. You upload forty minutes of preaching and a few minutes later five thousand words come back. This page opens the box: what a speech recognition model is, how sermon audio becomes text, and how the model learns to spell Melchizedek. In plain language, the way we’d walk a media volunteer through it after the service.

First, the line that matters most: speech recognition writes down the words it hears, in the order they were said. Every doctrine, illustration and word of the message comes from the pulpit. The model’s only job is getting those words onto the page.

What a speech recognition model is

The engine behind modern transcription is a speech recognition model (you’ll also see “speech-to-text” or the research term automatic speech recognition). It has been trained on a very large amount of recorded speech paired with accurate transcripts: meetings, podcasts, phone calls, broadcasts. From those examples it learns how sounds map to written words. Nobody writes rules like “this sound means shepherd”; the model learns the patterns itself.

That training is why context works. When something is ambiguous, the model doesn’t match sounds in isolation; it weighs which words are likely given the rest of the sentence. After “saved by,” the word “grace” is far more probable than “grays,” and the model knows it.

It’s also the model’s built-in weakness. Words that are rare in everyday speech, like Habakkuk, propitiation or Zerubbabel, are improbable to it, so it reaches for something common instead. More on the fix below.

From recording to transcript, step by step

Here’s what happens to a sermon in SermonKeep. Full disclosure: SermonKeep is our product.

1. The recording. You upload an audio file (MP3, M4A, WAV and most other formats, up to three hours and 500 MB) or paste a YouTube link to the service. This step matters more than people expect: the model can only transcribe what the microphone picked up, and close, clean audio beats any clever modeling. Our guide to recording sermons on an iPhone covers the practical side.
2. The language. You can tell us the sermon’s language (English, Spanish, Portuguese, French or German) or let it be detected automatically.
3. Speech recognition, with sermon vocabulary. The audio goes to AssemblyAI’s Universal-3.5 Pro model together with a short description of what a sermon sounds like, our theological dictionary and a few spelling rules (more on those below). The model returns the words with their timings, punctuated and capitalized.
4. Paragraphs and timestamps. The text is split into paragraphs, each with the minute it was said, so you can click any paragraph and hear that moment. Speaker labels are there if you switch them on.
5. The review pass. A second pass reads the transcript looking for words that look misheard. Obvious misrecognitions are fixed and listed so you can undo them; anything uncertain waits for you, with the audio ready to play. It only touches single words and short phrases. It never rewrites a sentence.
6. Scripture. Every passage the sermon used is found, including the ones alluded to but never named, and listed with the verse text beside the moment it came up.

That’s the honest answer to “how does sermon transcription work”: recording, recognition, formatting, review. No step decides what the sermon should have said.

The model under the hood

SermonKeep’s transcription runs on AssemblyAI’s Universal-3.5 Pro. We name the vendor on purpose: the same way an app tells you its payments run on Stripe, you should know whose model is listening to your preaching.

How good is it? AssemblyAI’s published benchmarks (2026) report a mean word error rate of 5.6% (median 4.9%) on English audio, roughly 94 correct words out of every 100, measured across 250+ hours of audio, 80,000+ files and 26 datasets. Those are AssemblyAI’s numbers on general audio, not ours on sermons, and the overall figure hides exactly the misses that matter most in church. We break it down honestly in our guide to sermon transcription accuracy.

Teaching the model your vocabulary

Remember the weakness: rare words are improbable. For sermons the fix isn’t a smarter model. It’s telling the model what’s coming, a technique AssemblyAI calls keyterms prompting. Alongside the audio, the request can carry a list of terms to expect (up to 1,000, per AssemblyAI’s documentation). When the model hears “mel-KIZ-eh-dek” and “Melchizedek” is on the list, the right spelling wins over three invented English words.

The mechanism is AssemblyAI’s; what goes on the list is our work. With every English sermon SermonKeep sends three things:

• A one-paragraph description of the recording: a Christian sermon, with Bible references, quotations from Bible translations, biblical names and transliterated Greek and Hebrew.
• The theological dictionary: about 180 hand-picked biblical names, places, doctrinal terms and Greek and Hebrew words. Spanish sermons get their own description and a list of about 130.
• Spelling rules for the variants that still slip through, so “Melchisedec” always comes back as Melchizedek.

The list is short on purpose: every entry is a word that’s rare and that we’ve seen go wrong, because long lists of common words make a model “correct” words that were already right. We publish 50 of them, with why each one trips a generic model, on the theological vocabulary page.

The same “context up front” idea matters even more when a service moves between languages. If your pastor preaches in English and Spanish, or alongside an interpreter, we’ve written a separate guide to bilingual sermon transcription.

What happens to your audio

Your sermons aren’t used to train anyone’s models. AssemblyAI has our account opted out of its model-improvement program, and we delete each transcript from AssemblyAI as soon as we’ve collected it. The transcript then lives in your church’s library, where your team can find it by title, passage, series or preacher, or search inside every transcript for a word or phrase and jump to the minute it was said. The details are in our privacy policy for churches.

What to look for in any sermon transcription service

If you’re evaluating speech recognition for your church, the technology above boils down to four practical questions:

1. Does the recording give the model a chance? Close, clean audio is the biggest improvement available, and it’s free.
2. Does the service know sermon vocabulary? A general tool handles ordinary sentences well and stumbles on exactly the words your congregation listens for. Ask whether it sends the model theological terms.
3. What happens to misheard words? You want corrections you can see and undo, and doubts that wait for a person, not a tool that quietly rewrites.
4. Is a person still in the loop? No model knows your members’ names or your town’s streets. A quick read before publishing is honest workflow, not a failure of the technology.

If you’d like to see it on your own preaching, our sermon transcription service for churches does your first sermon free, up to an hour, with no card. The clearest explanation of how it works is watching it handle your pulpit.

Sources

• AssemblyAI, Benchmarks: word error rate and dataset size (accessed October 2026).
• AssemblyAI, Custom vocabulary: keyterms prompting and its limits (accessed October 2026).