How accurate is speech-to-text transcription today?

On general English audio, the best speech recognition models are very good. AssemblyAI's published benchmarks (2026) report a mean word error rate of 5.6%, roughly 94 correct words out of every 100, measured across 250+ hours of audio, 80,000+ files and 26 datasets. Real-world accuracy depends heavily on the recording: one clear speaker close to the mic scores far better than a reverberant room with background noise.

What is word error rate (WER)?

Word error rate is the standard measure of transcription accuracy. It counts every substituted, inserted or deleted word and divides by the number of words actually spoken. One wrong word in a 20-word sentence is a 5% word error rate, or 95% word accuracy. Lower is better.

Why do transcripts get biblical names wrong?

General speech recognition models learn from everyday speech, such as meetings, phone calls and podcasts, where names like Melchizedek or Habakkuk almost never appear. Word error rate also weighs every word equally, so a model can score 94% overall and still miss the rare names that matter most. The fix is to tell the model in advance which biblical and theological terms to expect.

Is speech-to-text accurate enough for sermons?

For a searchable archive, accessibility transcripts and small group material, yes, as long as the service is set up for sermon vocabulary. A general-purpose tool handles ordinary sentences well and stumbles on biblical names, doctrinal terms and Scripture quotations. A quick read before anything is published is still wise.

How much does a human transcriber cost compared with speech recognition?

Human transcription services typically charge about $1 to $2 per audio minute (Rev, the best-known, charges $1.99 as of October 2026), so a 45-minute sermon costs roughly $45 to $90 and takes hours to days. Speech recognition does the same sermon in minutes for a small fraction of that, which is why most churches now use it weekly and keep a human transcriber for special projects like a book manuscript.

How accurate is sermon transcription? Speech-to-text accuracy, explained.

Q: Does the review pass change what the pastor said?

No. In SermonKeep the review pass only touches single words or short phrases 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 one click away. It never rewrites a sentence. The transcript stays the words that were preached, in the order they were preached.

What word error rate really measures, why a transcript can be 94% right and still miss Melchizedek, and what actually makes a sermon transcript better.

By the SermonKeep team · October 9, 2026

If you’ve ever turned on YouTube’s automatic captions during a sermon, you already know the honest answer: speech-to-text is impressively good at ordinary sentences and reliably bad at “Melchizedek.”

Both halves are worth understanding, because the gap between them is exactly where a church transcript succeeds or fails. (If you’d like the bigger picture first, start with how sermon transcription works.) On this page we explain how speech-to-text accuracy is measured, what the best models achieve today, why sermons are unusually hard audio, and what closes the gap. In plain language, the way we’d explain it to a media volunteer after the service.

One thing to be clear about first. A speech recognition model does one job: it writes down the words it hears, in the order they were said. It doesn’t summarize the sermon, improve it or put words in the preacher’s mouth. Accuracy is simply a question of how often it hears a word right.

What “accuracy” means: word error rate

When a transcription company says its model is “94% accurate,” there’s a specific measurement behind it: word error rate, or WER. It’s the industry standard, and it’s simple enough to work out on a napkin.

Take the transcript and compare it, word by word, with what was actually said. Count three kinds of mistakes:

Substitutions: the wrong word (“grease” for “grace”)
Deletions: a word that was dropped
Insertions: a word that was never spoken

Add them up and divide by the number of words spoken. That’s the word error rate. Word accuracy is the other side of it: 100% minus the WER.

A worked example. The preacher reads a 20-word sentence:

“By grace you have been saved through faith, and this is not from yourselves; it is the gift of God.”

If the transcript gets 19 of those words right and one wrong, that’s 1 error ÷ 20 words, a 5% word error rate, or 95% word accuracy. One mistake every 20 words. A 40-minute sermon runs roughly 5,000 to 6,000 words, so 95% means around 250 to 300 wrong words across the whole message.

Why one overall number can mislead: missed entities

Here’s the catch, and it matters more for sermons than for almost any other audio: word error rate treats every word as equally important. Getting “the” wrong costs exactly as much as getting “Melchizedek” wrong.

Those two mistakes are not remotely equal. If a transcript writes “a” instead of “the” fifty times, nobody notices. If it writes “have a cook” for “Habakkuk” and “proposition” for “propitiation,” the transcript is wrong in precisely the places your congregation is paying the most attention.

Speech recognition researchers now measure this separately. AssemblyAI, one of the major speech recognition companies, calls it the missed entity rate: instead of scoring every word, you score only the high-value ones, like proper names, organizations and technical terms. Their published figures (2026) make the point: their own leading model, while scoring under 6% WER overall, still missed about 17% of organization names and 13% of specialized medical terms in their tests. Roughly one in six of the words that matter most.

In a sermon, the “entities” are biblical names, places and doctrinal words. A model can honestly claim 94% overall and still get Epaphroditus and Zerubbabel wrong every time, because those names are statistically invisible in the overall score. So when you evaluate a transcription service, ask not only “how accurate is it overall?” but “how accurate is it on the words a general model has never heard?”

Speech-to-text accuracy today

SermonKeep’s transcription runs on AssemblyAI’s Universal-3.5 Pro speech recognition model. We name the vendor for the same reason a checkout page says “payments by Stripe”: you should know whose technology handles your church’s audio.

AssemblyAI’s published benchmarks (2026) report a mean word error rate of 5.6% (median 4.9%) on English audio, measured across 250+ hours of audio, 80,000+ files and 26 datasets. That’s roughly 94 correct words out of every 100.

Two honest caveats. First, those are AssemblyAI’s numbers on AssemblyAI’s test sets, a mix of podcasts, meetings, phone calls and broadcast audio. They are not sermon benchmarks, and we don’t present them as ours. Second, AssemblyAI itself cautions that public benchmarks can flatter any model, because the test sets are public. Benchmarks tell you the ceiling; your own recordings tell you the truth.

Why sermons are hard audio

Preaching combines nearly every condition that makes speech recognition harder:

Biblical names. Mephibosheth, Onesimus, Sennacherib. They barely exist in the everyday speech a general model learns from, so it guesses.
Theological vocabulary. Propitiation, sanctification, ecclesiology, plus transliterated Greek and Hebrew like hesed, agape and kenosis, which sound like other English words.
Scripture quotations. A preacher reading from the King James shifts into 17th-century grammar mid-sentence. “Thou shalt” and “cometh” trip up a model tuned for modern conversation.
Sanctuary acoustics. High ceilings, hard walls and a mic ten rows from the pulpit produce reverb, one of the toughest conditions for any model.
The congregation. Amens, laughter, babies and the worship band bleeding into the recording all compete with the preacher’s voice.
Preaching dynamics. No other kind of speech swings from a shout to a whisper in the same paragraph.

Some of this is in your hands. A clean recording close to the source makes a measurable difference. We cover the practical side in our guide to recording sermons on an iPhone, and if your recording desk can take a feed from the soundboard, use it.

What closes the gap: telling the model what’s coming

The vocabulary problem, the biggest one, has a real solution, and it isn’t a smarter model. It’s telling the model what to expect.

AssemblyAI’s models accept what it calls keyterms prompting: with each transcription you can send a list of domain terms (up to 1,000, per its documentation), and the model uses them when choosing between similar-sounding options. Heard something like “mel-KIZ-eh-dek”? If “Melchizedek” is on the list, the model picks it instead of inventing three English words.

This is the part we built. Full disclosure: SermonKeep is our product. With every English sermon we send three things along with the audio:

A short description of what the recording is: a Christian sermon, with Bible references, quotations from Bible translations, biblical names and transliterated Greek and Hebrew.
Our theological dictionary: about 180 hand-picked terms (Bible books that get misheard, Old and New Testament names and places, doctrinal vocabulary, Greek and Hebrew words). Spanish sermons get their own list of about 130.
Spelling rules for the variants that still slip through, so “Melchisedec” always comes back as “Melchizedek” and “chesed” as “hesed”.

Why 180 and not the full 1,000? On purpose. A long list of common words makes a model “correct” words that were already right, and every entry uses up room on the list. So the list holds only words that are both rare and known to go wrong. We explain how we chose them, and publish a sample, on the theological vocabulary page.

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 or smooths out how the preacher talks: it only touches single words and short phrases.

When you should still read it through

We’d rather earn your trust than overclaim. Even a sermon-tuned transcript deserves a read before you publish it. Watch three places in particular.

Your own people and places. No dictionary contains your members’ names, your town’s streets or the sister church you partner with. Expect guesses there.

Other languages. A preacher who drops a sentence of Spanish, Latin or Greek into an English sermon will usually get an approximation back.

Invented words. Sermon illustrations love made-up names and wordplay. The model maps them to the nearest real words it knows.

What tuning changes is where the errors live. Instead of correcting the whole transcript, you’re checking a handful of predictable spots: minutes of reading instead of the hours a manual transcript takes.

What about a human transcriber?

A person typing every word is still the accuracy ceiling for difficult audio, and it costs accordingly: human transcription services typically charge about $1 to $2 per audio minute, roughly $60 to $120 for an hour of preaching (Rev, the best-known, lists $1.99 a minute as of October 2026), and the transcript arrives hours or days later. For a book manuscript or a denominational archive that has to be perfect, that can be worth it. For a weekly sermon, speech recognition plus a short read gets you a usable transcript the same morning, for a small fraction of the price. We compare the two in detail in human vs automatic transcription.

How we’ll measure accuracy on real sermons

Everything quoted above is AssemblyAI’s data on general audio, clearly attributed, because that’s what honestly exists today. What doesn’t exist yet, from us or anyone we know of, is a published accuracy benchmark on real preaching: sanctuary acoustics, biblical vocabulary, Scripture quotations and all.

We’re building one: a test set of real sermons with carefully corrected reference transcripts, scored for word error rate and for missed biblical terms, with and without the theological dictionary. When the numbers are ready we’ll publish them here, including the ones that don’t flatter us.

The bottom line

Modern speech-to-text is accurate enough to make sermon transcription genuinely useful: AssemblyAI’s benchmarks (2026) put the state of the art at roughly 94 of every 100 words on general audio. But sermons aren’t general audio, and the overall number hides the misses that matter. What closes the gap is sermon vocabulary, a good recording and a quick read over the predictable weak spots.

If you’re comparing tools, that’s the lens we’d use, and it’s the one we apply when we compare sermon transcription services, competitors included. And if you’d like to see a sermon-tuned transcript of your own preaching, our sermon transcription service does your first sermon free, up to an hour, no card required. The fairest benchmark is always your own pulpit.

Sources

AssemblyAI, Benchmarks: word error rate, dataset size and missed entity rate (accessed October 2026).
AssemblyAI, Custom vocabulary: keyterms prompting and its limits (accessed October 2026).
Rev, Pricing: human transcription per minute (accessed October 2026).