Why generic transcription mishears Scripture, and the dictionary we built to fix it.
How custom vocabulary transcription works, why Melchizedek turns into “milk is a deck,” and 50 words from the theological dictionary SermonKeep sends with every English sermon.
By the SermonKeep team ·
Run a sermon through a general-purpose transcription tool and you’ll meet a strange new cast of characters. These are illustrative examples of the kind of error generic speech recognition makes, not measurements of any particular product, but if you’ve ever transcribed a sermon they’ll feel familiar:
- “Melchizedek, king of Salem” becomes “milk is a deck, king of Salem.”
- “The prophet Habakkuk” becomes “the prophet have a cook.”
- “Christ is the propitiation for our sins” becomes “Christ is the proposition for our sins.”
- “A sermon series on eschatology” becomes “a sermon series on a scatology.”
- “Nebuchadnezzar’s dream” becomes “never can as her dream.”
Each is one misheard word. But it’s never a throwaway word. It’s the king in your text, the doctrine in your main point, the name your small group will look for on Thursday. A transcript that gets the filler right and the theology wrong isn’t much of a transcript.
This page explains why generic models make these mistakes, how custom vocabulary transcription fixes most of them, and shares part of the theological dictionary we built into SermonKeep. Full disclosure: SermonKeep is our product, and the dictionary is our own work, so we’ll describe both the mechanism and its limits plainly.
Why generic speech-to-text mishears Scripture
A speech recognition model learns from enormous amounts of recorded speech: meetings, phone calls, podcasts, videos. When it transcribes, it isn’t looking words up in a dictionary. It’s predicting the most likely sequence of words, given the sounds it hears and the words around them.
That’s why these models are so good at ordinary conversation and so bad at Zerubbabel. In everyday speech, the sounds mel-KIZ-eh-dek almost never turn out to be “Melchizedek,” because almost nobody outside a pulpit says Melchizedek. The sounds do resemble common words, milk, is, a, deck, so the model reaches for the sequence it has seen a million times instead of the name it has barely seen at all.
The model isn’t broken. It’s doing exactly what it was built to do: bet on the probable. The trouble is that a sermon is full of improbable words, and they’re the words that matter most. It’s also why a headline accuracy figure can mislead: a model can get 94 of every 100 words right and miss the six that carry the sermon. We explain that gap, and the “missed entity rate” that exposes it, in our guide to sermon transcription accuracy.
The fix: custom vocabulary transcription
The idea is simple: give the model the vocabulary before it starts listening.
AssemblyAI, whose Universal-3.5 Pro model handles SermonKeep’s transcription, calls this keyterms prompting. According to its documentation, you can send up to 1,000 words or phrases with each transcription, up to six words per phrase. The model uses them in context, so when it hears mel-KIZ-eh-dek and “Melchizedek” is on the list, the improbable word becomes the probable one.
A blank list helps nobody, though. Someone has to decide which words earn a place. That’s the part we built.
Our theological dictionary
With every English sermon, SermonKeep sends three things along with the audio:
- A one-paragraph description of the recording: a Christian sermon, with Bible references by book, chapter and verse, quotations from Bible translations, biblical names and places, theological vocabulary, and transliterated Greek and Hebrew. It tells the model what kind of speech to expect before it hears a word.
- The dictionary itself: about 180 terms in five groups: Bible books that get misheard, Old Testament names and places, New Testament names and places, doctrine and theology, and transliterated Greek and Hebrew.
- A few spelling rules for variants that still slip through. “Melchisedec” always comes back as Melchizedek, “chesed” as hesed, “Gethsemene” as Gethsemane.
Spanish sermons get their own description and their own list of about 130 terms (Melquisedec, Nabucodonosor, propiciación, jésed), because the spelling of a name changes with the language; more on that in our guide to bilingual sermon transcription. Sermons preached in other languages are transcribed without a dictionary for now.
Why 180 terms and not 1,000
The obvious move would be to fill all 1,000 slots. We don’t, on purpose. Long lists of common words make a model start “correcting” words that were already right, and AssemblyAI notes that long words and capitals use up the list’s capacity faster. A sermon about justification doesn’t need the word grace on a list; the model already knows grace. So every entry has to pass two tests: it’s rare in everyday speech, and we’ve seen it go wrong. Short and sharp beats long and noisy.
What it doesn’t fix
A dictionary raises the odds; it doesn’t make transcription perfect. A muddy microphone, the congregation talking over the benediction, or a name we haven’t added yet can still produce errors. After transcription, a review pass reads the text 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 of that moment ready to play. It never rewrites a sentence. Your own people and places, which no list could know, are worth a quick read before anything is published.
And we don’t have a first-party accuracy benchmark on sermons yet. The figures we quote are AssemblyAI’s, on general audio: a 5.6% mean word error rate across 250+ hours, 80,000+ files and 26 datasets (2026). When we’ve measured our own, we’ll publish it.
50 words generic transcription gets wrong in sermons
Here’s a sample from the dictionary: fifty terms that trip up general-purpose models, and why. Every one of them is on the list SermonKeep sends. Feel free to share or link this page, and if you’re building a custom vocabulary on another tool, it’s a fair place to start.
Bible books
| Term | Why it trips a generic model |
|---|---|
| Habakkuk | Sounds like the everyday phrase “have a cook” |
| Haggai | Short, rare, all vowels |
| Zephaniah | Rare; the middle syllables blur |
| Obadiah | Rare; easily split into “Oh, but I” |
| Philemon | Heard as a common surname or “Philly man” |
| Ecclesiastes | Long Greek name with no everyday neighbor |
| Thessalonians | Six syllables; the ending gets dropped |
Old Testament names and places
| Term | Why it trips a generic model |
|---|---|
| Melchizedek | Rare four-syllable name; turns into common-word soup |
| Nebuchadnezzar | Long, rare, no everyday phonetic neighbor |
| Zerubbabel | Extremely rare; the inner syllables get mangled |
| Mephibosheth | Rare even among preachers; no English cue at all |
| Sennacherib | Rare Assyrian name; stress pattern misleads |
| Jehoshaphat | Survives mostly as an exclamation, not a spelling |
| Bathsheba | Splits into “bath” plus a guess |
| Shadrach, Meshach, Abednego | Three rare names in a row, said fast |
| Ahasuerus | Spelling bears no resemblance to the sound |
| Nineveh | Heard as “ninety” or “Ninive” |
New Testament names and places
| Term | Why it trips a generic model |
|---|---|
| Zacchaeus | Rare; the -aeus ending defeats phonetic guessing |
| Nicodemus | Nearby common words pull it apart |
| Epaphroditus | Five syllables of rarity |
| Onesimus | Rare Greek name; heard as “one, Simus” |
| Caiaphas | Rare; vowels collapse into “Cai fuss” |
| Gethsemane | Soft consonants; often misspelled even when heard |
| Golgotha | Guttural cluster with no English cousin |
| Capernaum | Rare place name; stress pattern misleads |
| Sanhedrin | Rare; heard as “San Hedrin” or worse |
| Areopagus | Five syllables of Greek |
Doctrine and theology
| Term | Why it trips a generic model |
|---|---|
| Propitiation | Near-homophone of “proposition” |
| Expiation | Collapses into “expiration” |
| Sanctification | Long; blurs into “certification” on a poor recording |
| Imputation | Heard as “amputation” |
| Eschatology | Painfully close to “a scatology” |
| Soteriology | No everyday phonetic anchor |
| Hermeneutics | Opening syllables suggest “her men” |
| Exegesis | Academic rarity; heard as “exit Jesus” |
| Hypostatic union | Two rare words, rarely said together |
| Substitutionary atonement | Long technical phrase; words get split or dropped |
| Sola scriptura | Latin phrase inside an English sentence |
| Theodicy | Heard as “the odyssey” |
Transliterated Greek and Hebrew
| Term | Why it trips a generic model |
|---|---|
| Hesed | Hebrew; often written as “he said” |
| Agape | Heard as the English “agape” (mouth open) or “a gap” |
| Koinonia | No English spelling cue |
| Kenosis | Drifts toward “keno” plus noise |
| Metanoia | Rare; the syllables reshuffle |
| Shekinah | Heard as a name, spelled half a dozen ways |
| Tetelestai | ”It is finished” in Greek; no English neighbor |
| Ruach | Guttural ending gets swallowed |
| Yahweh | Divine name; models guess “Yahoo way” and worse |
| Maranatha | Rare Aramaic; the syllables reshuffle freely |
| Ebenezer | Heard as the Dickens character, which is at least spelled right |
What this means for your church
Getting these words right isn’t a spelling contest. It’s the difference between a transcript you can use and one a volunteer has to babysit.
When Melchizedek is spelled Melchizedek, the passage he belongs to can be found and listed beside the moment the pastor preached it. The small group guide quotes the sermon accurately instead of comically. The clips and quotes you pull for the week don’t need a correction pass before they face the congregation. That’s what our church transcription service is built around: upload the sermon once, get the word-for-word transcript with its Scripture references, then the small group guide, clips and quotes, and the YouTube and podcast description from the same message.
If you’re comparing tools, most general-purpose services don’t ship a theological vocabulary; you’d have to build and maintain your own list, if the tool accepts one at all. We look at that trade-off in our roundup of the best sermon transcription services.
The first sermon is free, up to an hour, no card required. Upload one this week and see what it does with your pastor’s favorite hard name.
Sources
- AssemblyAI, Custom vocabulary: keyterms prompting, limits and capacity (accessed October 2026).
- AssemblyAI, Benchmarks: word error rate and dataset size (accessed October 2026).
