# Dictation software gets legal terms wrong on Mac: fixes for lawyers in 2026

> Dictation software gets legal terms wrong because it is trained on everyday speech. Add Apple's custom vocabulary on your Mac, then check every draft.

URL: https://www.lazur.app/for/lawyers/problems/dictation-software-misreads-legal-terms
Updated: 2026-11-06

Dictation software gets legal terms wrong because it is trained mostly on everyday speech. Case names, Latin phrases and statute numbers are rare words, so the software guesses lookalike words. On a Mac, the fix you can use today is Apple's custom vocabulary in Voice Control, which holds up to 1,000 terms. No study measures how often lawyers hit this, so the numbers here come from related fields.

## Quick answer

- Most dictation tools learn from everyday speech, so rare legal words are where errors pile up.
- On a Mac, Apple's Voice Control vocabulary takes up to 1,000 terms. Standard Dictation has no documented word list.
- Dragon for Mac stopped being sold on 22 October 2018, so it is not a current option.
- In a 2018 medical study, raw drafts had 7.4 errors per 100 words. Read every draft before it leaves your Mac.

## What dictation misrecognizing legal vocabulary costs you

- **Time lost to cleanup.** A raw draft is the most expensive version to read. In the 2018 medical study, raw drafts carried 7.4 errors per 100 words. Each one is a word you must find, check and fix. Clio's 2017 Legal Trends Report, as quoted by National Jurist, puts administrative tasks at 48% of the average lawyer's day. Cleanup competes with the work you can bill.
- **Errors that reach a signed document.** Editing cuts word errors a lot, but it does not cut meaning errors. In the same medical study, errors involving clinical information were 15.8% in raw drafts and 25.9% in signed notes. A wrong party name or citation can look right on a quick read. The cost shows up later, as a correction, a question from a client or a reviewer who spots it first.
- **Money and time for a switch.** If your firm used Dragon for Mac, that product is no longer sold. Nuance discontinued it effective 22 October 2018. Any replacement needs its own test, because each tool mishears different words. Budget time to check a switch before you assume it fixes your terms.

## Why it happens

1. **Training data is everyday speech.** General speech models learn from everyday language. That is the material they see most. When a word is rare, the model has less evidence for it, so it picks a common word that sounds similar. A 2024 INTERSPEECH paper tested general English audio and found that combined contextual biasing cut rare-word error rate by 60% relative to no biasing. Rare words are where the model stumbles.
2. **Case names, Latin phrases and proper nouns.** Party names, judges and Latin phrases such as res judicata and mens rea appear rarely in ordinary speech. A surname with an unusual spelling gives the model several plausible options. Spacing and accents make it worse, because a phrase can split into two words or join into one. Each of these is a place where a lookalike word can slip in.
3. **Pronunciation, speed and noise.** Fast speech, soft endings and background noise give the model less to work with. A strong accent can help or hurt depending on the training data. Pause briefly before and after a hard term. Say it at the same pace as the words around it, not louder. These habits do not replace checking, but they reduce the misses you have to catch.
4. **Cloud and on-device models differ.** The ABA Journal says on-device dictation isn't always as accurate as cloud processing. Cloud dictation typically sends your recorded speech to a provider's servers. A setting that keeps audio on your Mac can trade away some accuracy. Test both on the same passage before you choose one for client work.
5. **Vocabulary lives in one place.** Apple documents its custom vocabulary inside Voice Control. Standard Dictation has no documented word list. A term you teach in one place may not help in the other. Check both paths before you assume the fix took, because the same word can behave differently in each.

## How to fix it

- **Build the term list from real matters.** Start with the names and phrases you type most often in your current matters. Put one term per line in a plain text file. Keep the list to terms you actually dictate, because a short list is easier to check than a long one you forget. Apple's custom vocabulary holds up to 1,000 terms, so you have room to grow.
- **Switch Voice Control on only when you need it.** Standard Dictation is unavailable while Voice Control is on. Switch Voice Control off for everyday dictation. Switch it on when you work with the vocabulary itself. This is a workaround, not a fix. It also means you must test whether your terms carry over to the mode you use most.
- **Use placeholder tokens for the hardest terms.** Some terms will not come out right no matter how you say them. Dictate a short, unusual token in their place, then finish the sentence at normal speed. Search for each token when the draft is done, and replace it with the exact spelling from the matter file. This needs no purchase and no special setting.
- **Add exact spellings to Word's dictionary.** In Word, right-click a flagged word and choose Add to Dictionary. The spelling flag goes away for that word in Word. This does not teach Dictation how to hear the word. It does stop a correct name from hiding among real errors, so your eye can find the wrong ones faster.
- **Verify names against the filed record.** Check every party name, judge name and citation against the document that was actually filed, the court's public site or your firm's research database. Public pages are a good place to start. Research run by Lazur happens in its own Chrome tabs and comes back as a summary and report, but you still verify each spelling yourself.
- **Choose the setting that fits the matter.** Decide before you dictate whether the work can go through a cloud service. Built-in dictation can run on the device, which can help when the matter is sensitive, but it may be less accurate. Follow your employer's rules on client data. If a rule is unclear, ask a supervisor or your firm's IT contact before you test with real client material.
- **Keep a miss log and retest after changes.** Write down each wrong word, the term it should have been and the app where it happened. Review the log each week. When macOS updates or your microphone changes, run the same test passage again and compare the misses. A rising count tells you the fix has stopped working.

## Why dictation software misses legal words

Most speech recognition is built for everyday talk. The model learns from a large amount of ordinary sentences, so it expects common words. When you say a term it rarely hears, it picks a common word that sounds close. For example, “tort” can come out as “tart.” Legal work is full of words that rarely show up in everyday speech. Those are the words that break.

Here is the honest gap. I could not find a survey, study or audit that measures how often lawyers see wrong legal terms in dictation. I also found no study that puts a price or a grade on those errors. What exists is indirect. A 2024 benchmark of French legal speech says current systems need further refinement for legal use, but it gives no error figures. Treat the evidence below as signals, not benchmarks for your practice.

![Lawyer typing on a laptop at a desk while checking legal text on screen](https://lfobmih1tpj3sgpo.public.blob.vercel-storage.com/content-images/problem/dictation-software-misreads-legal-terms/dictation-software-gets-legal-terms-wrong-1-791ca4.jpg)

## What the numbers say, and what they do not

- **7.4** errors per 100 words in raw speech-recognition drafts of medical notes (2018) ([source](https://pmc.ncbi.nlm.nih.gov/articles/PMC6203313/))
- **96.3%** of raw dictated medical drafts had at least one error ([source](https://pmc.ncbi.nlm.nih.gov/articles/PMC6203313/))
- **1,000** vocabulary terms you can add to Apple Voice Control ([source](https://support.apple.com/guide/mac-help/use-a-custom-vocabulary-mchl3eb7b79a/mac))
- **60%** relative cut in rare-word errors from combined biasing, tested on general English audio ([source](https://arxiv.org/abs/2407.10303))

The best-measured case is medical. A 2018 JAMA Network Open study looked at 217 clinical notes from 144 dictating physicians. The speech-recognition draft had 7.4 errors per 100 words. After a transcriptionist edited the note, the rate fell to 0.4%. Signed notes reached 0.3%. Most raw drafts, 96.3%, had at least one error.

Medicine is not law, and this study looked at clinical notes, not briefs or contracts. Still, the pattern is useful. Editing removed most of the word-level errors. It did not remove errors in clinical information. Those were 15.8% in raw drafts, 26.9% in edited versions and 25.9% in signed notes. A wrong meaning is harder to spot than a misspelled word.

**Errors per 100 words at each review stage**

- Raw speech-recognition draft: 7.4 per 100 words
- After transcriptionist edit: 0.4 per 100 words
- Physician-signed note: 0.3 per 100 words

Source: https://pmc.ncbi.nlm.nih.gov/articles/PMC6203313/

Read the chart with care. It shows medical notes, and it does not predict how many case names your Mac will miss this week. What it does show is the shape of the problem. A raw draft is a rough first pass. Most of the word errors disappear only when someone reads and corrects the text.

## A worked example with a 1,000-word memo

Take a 1,000-word memo dictated in one session. At the raw draft rate from the study, you would expect about 74 word errors before any edit. At the edited rate of 0.4%, you would expect about 4. These are sums on study rates, not a forecast for your Mac. Your microphone, your accent and your vocabulary change the result.

The arithmetic also shows where to spend your time. Some errors matter far more than others. A case name that is wrong on page one can change how a reader sees the whole memo. Check the names, citations and numbers first, then read for everything else. That order catches the costly misses first.

| Term type | Example | What to check |
| --- | --- | --- |
| Case names | Smith v. Jones | Spelling of each party, against the filed opinion |
| Latin phrases | res judicata, mens rea | Spacing and accents, and whether the software split or joined the words |
| Statute numbers | § 1983 | Digits and the section sign, against the official code |
| Client and judge names | Surnames with unusual spelling | Spelling against your matter file |

_Common legal term types and what to check in each. Examples are illustrative, not from a study._

## Set up a custom vocabulary on your Mac

Apple's custom vocabulary lives in Voice Control. Apple says you can add up to 1,000 terms and teach Voice Control how to say them. The macOS 27 guide puts the setting at System Settings > Accessibility > Voice Control > Vocabulary. Older macOS versions may show a different layout, so check the screen you actually see.

![Professional speaking into a microphone at an office desk while dictating notes](https://lfobmih1tpj3sgpo.public.blob.vercel-storage.com/content-images/problem/dictation-software-misreads-legal-terms/dictation-software-gets-legal-terms-wrong-2-99fc1d.jpg)

1. Turn on Dictation. Apple's guide says Dictation must be on to teach Voice Control how to say a phrase.
2. Open System Settings > Accessibility > Voice Control > Vocabulary.
3. Add each term, or import a plain text file with one term per line.
4. Teach the pronunciation for each term. If Dictation is off, Apple's guide says you can still save the term with Save Without Recording.
5. Export the list as a backup. Voice Control can export its vocabulary to a file.
6. Dictate a test passage in Pages or Word that uses every term, then check the result.

Keep the trade-off in mind. Apple says standard Dictation is unavailable while Voice Control is on. You may need to switch Voice Control off to use standard Dictation for everyday work. Apple's guide does not say whether your vocabulary carries into standard Dictation. Test that before you rely on the fix.

Dragon is not the way out on a Mac. Nuance discontinued Dragon Professional Individual for Mac effective 22 October 2018, and it stopped providing updates. Existing version 6 perpetual license holders could keep using it. If a colleague recommends it, check whether the product is still supported on your Mac before you spend anything.

## Check legal dictation before it leaves your Mac

Placeholders help when the vocabulary does not. Dictate a made-up token in place of the hard term, such as placeholder cite one. Keep the sentence moving at normal speed. When the draft is done, search for the token with ⌘F in Word or Pages. Replace it with the exact term from your matter file, then search again to confirm none are left.

![Lawyer reading printed legal documents at a wooden desk during case review](https://lfobmih1tpj3sgpo.public.blob.vercel-storage.com/content-images/problem/dictation-software-misreads-legal-terms/dictation-software-gets-legal-terms-wrong-3-798aea.jpg)

Apple's Dictation page says uncertain words are underlined in blue. Stop at each blue underline and check it against your source. Those underlines show what the software is unsure about. They do not show the words it got wrong with confidence, so keep reading after you clear them.

- Check every case name and party name against the filed document.
- Check citations and section numbers against the official source.
- Search for every placeholder token, and confirm none are left in the draft.
- Check Latin phrases for spacing, accents and any split or joined words.
- Read the blue underlines in context, not only the red spelling flags.
- Read the final text once more, aloud if you can, before it goes out.

## Mistakes people make with dictated legal text

- Treating the first draft as the final draft. Raw drafts carried the most errors in the study above.
- Assuming a vocabulary list works everywhere without testing it in each app.
- Checking spelling only. A spell checker cannot tell which real word you meant.
- Dictating client names into any tool before checking your firm's policy.
- Forgetting that Voice Control turns off standard Dictation while it is on.
- Reading the text on screen but never checking it against the source record.

## What to do when one term keeps failing

Some terms will keep failing, even after you add them. Work through these steps in order. First, check the spelling in your term file, because a typo there teaches the wrong word. Second, dictate the term alone in a short sentence, with a pause before and after it. Third, if it still fails, stop dictating that term and use a placeholder for every instance.

Keep a short miss log. Write down the term, the wrong output and the app where it happened. After a few weeks, patterns show up, like one judge's surname failing every time. That pattern tells you whether to fix the vocabulary, the placeholder habit or the review step.

## Cloud or on-device dictation and client data

The ABA Journal notes that built-in dictation on Windows and Mac can process speech on the device, so it can work offline. It also says on-device dictation isn't always as accurate as cloud processing. Cloud dictation typically sends your recorded speech to the provider's servers. Apple's macOS 27 guide adds an on-device cleanup for spelling, punctuation and capitalization. It needs English, an M3 or later chip and at least 12GB of unified memory.

The accuracy and privacy trade-off is real, so test both settings on the same passage. Then follow your employer's rules on client data before you dictate client names, case facts or privileged details into any tool. If the rules do not answer your question, ask before you guess. Your firm's answer matters more than any product page.

## Where Lazur helps, and where it does not

Lazur is a voice assistant for the Mac. It lives in the notch. Hold fn to dictate clean, formatted text at the cursor in any app. Dictation includes snippets, a personal dictionary and clipboard slots. Those help with terms you type often. They do not check a case name against a court record, and Lazur cannot tell whether a term is right for your matter.

For Apple's settings, Lazur guides and points. It does not click buttons, edit files or operate other apps for you. Ask Lazur to walk you through the Vocabulary pane, and it waits for each click before the next step. Ask where a setting is, and Lazur circles it and labels it on your screen.

On privacy, Lazur listens and looks at the screen only while you hold the shortcut. Audio and the screenshot go to AI providers to answer you, and they are not stored. Content is never used to train AI models. That does not replace your firm's rules on client data, so check them before you use any assistant with client material.

For the Word side of a legal review, see [Word for lawyers on Mac](https://www.lazur.app/for/lawyers/word), which covers Track Changes and Compare. To test your dictation on a passage, try the [dictation accuracy checker](https://www.lazur.app/tools/dictation-accuracy-checker). For Lazur in Word, read the [Word assistant page](https://www.lazur.app/apps/word).

## With Lazur

Lazur is a voice assistant for the Mac that lives in the notch. Hold ⌥ ⌘ and ask; it sees your screen, answers out loud, points at buttons and walks you through apps. Hold fn to dictate. Free for 3 days, then $18/month: https://www.lazur.app/download

- “Where is the Vocabulary setting?”: Lazur circles the Vocabulary row in System Settings and labels it on screen. You still click it yourself. (Points at it)
- “Walk me through adding a custom vocabulary term in Voice Control.”: Lazur guides you one click at a time, and it waits for each click before the next step.
- “Hold fn and dictate into the open Word document.”: Clean, formatted text lands at the cursor. Check each legal term before filing, because dictation can still mishear.
- “What does this Voice Control setting change for my dictation?”: Lazur explains what is on screen and answers out loud. It does not change the setting for you.
- “Research the official spelling of this court's name and summarize it.”: Lazur runs the research in its own Chrome tabs and returns a summary and a report. You check the source yourself.
- “Draft a term list for this matter from these case names.”: Lazur drafts the list as text you can paste into a plain text file. Check every name against the filed record.

## A plan you can start this week

- **Today: Build your term list.** Write down the case names, party names, statute numbers and Latin phrases you dictate most. Put one term per line in a plain text file. Check with your firm about where working notes should live before you save it anywhere.
- **This week: Import and test the vocabulary.** Turn on Dictation, then import the list under System Settings > Accessibility > Voice Control > Vocabulary. Dictate a test passage in Word or Pages that uses every term. Log each miss, then fix the terms that failed.
- **Every draft: Use placeholders for hard terms.** For a term you cannot get right, dictate a placeholder token and keep writing. When the draft is done, search for each token with ⌘F and replace it with the exact spelling from the matter file. Search again to confirm none are left.
- **Before filing: Read the flagged words.** Check every blue underline in Apple's dictation output. Then check each case name, citation and section number against the source. Read the final text once more before it leaves your Mac.
- **After updates: Retest after macOS changes.** Run the same test passage after a macOS update or a microphone change. Compare the misses with your log. If the count rises, test whether your vocabulary still carries over to the mode you use.

## FAQ

### Does Dragon work for legal dictation on a Mac?

Not as a current product. Nuance discontinued Dragon Professional Individual for Mac effective 22 October 2018, and it stopped providing updates. People who held a version 6 perpetual license could keep using it. Check the current options before you buy anything.

### Can I teach Apple dictation my legal terms?

Apple documents a custom vocabulary inside Voice Control, with up to 1,000 terms. Its guide does not say whether those terms carry into standard Dictation. Test with a passage that uses your hardest terms before you rely on it.

### Does on-device dictation keep my words off the internet?

The ABA Journal says built-in dictation can run on the device and work offline. It also says on-device dictation may be less accurate than cloud processing. Whether a setting keeps data private depends on your setup and your firm's policy, so confirm both.

### What accuracy is acceptable for legal transcription?

Speechmatics says there is no single number. It cites a UK Crown Court contract that requires 99.5% accuracy. Set your own bar by counting misses on your own passages, then decide what your review step must catch.

### Will Lazur catch wrong legal terms for me?

No. Lazur dictates clean, formatted text at the cursor, and it has a personal dictionary and snippets. It does not verify case names or citations against a legal source, so you still check every term against the filed document.

### Which Macs can use Apple's on-device cleanup?

Apple's macOS 27 dictation page says the on-device cleanup for spelling, punctuation and capitalization needs English, an M3 or later chip and at least 12GB of unified memory. It does not say the cleanup fixes legal vocabulary. Test it before you count on it.

### Is a spell checker enough to catch wrong case names?

Usually not. A spell checker flags words it does not know. It can pass a real word that is wrong in context, or a name it already accepts. Compare every name against the source, not against the spell checker.

## Sources

1. [Use a custom vocabulary with Voice Control on Mac](https://support.apple.com/guide/mac-help/use-a-custom-vocabulary-mchl3eb7b79a/mac), Apple Inc., checked 2026-11-07
2. [Use Dictation on Mac (macOS 27)](https://support.apple.com/guide/mac-help/use-dictation-mh40584/mac), Apple Inc., checked 2026-11-07
3. [Analysis of Errors in Dictated Clinical Documents Assisted by Speech Recognition Software and Professional Transcriptionists (JAMA Netw Open 2018)](https://pmc.ncbi.nlm.nih.gov/articles/PMC6203313/), PubMed Central / JAMA Network Open, checked 2026-11-07
4. [Improving Neural Biasing for Contextual Speech Recognition by Early Context Injection and Text Perturbation](https://arxiv.org/abs/2407.10303), arXiv, checked 2026-11-07
5. [The State of Commercial Automatic French Legal Speech Recognition Systems](https://arxiv.org/abs/2408.11940), arXiv, checked 2026-11-07
6. [Nuance Discontinues Dragon Professional Individual for Mac](https://www.macrumors.com/2018/10/24/nuance-discontinues-dragon-mac/), MacRumors, checked 2026-11-07
7. [Lawyers only bill 2.3 hours a day; what happens to the rest of it?](https://www.nationaljurist.com/lawyers-only-bill-23-hours-day-what-happens-rest-it/), National Jurist, checked 2026-11-07
8. [Speech-to-text dictation: A 21st-century twist to a traditional law firm tool](https://www.abajournal.com/web/article/speech-to-text-dictation-a-21st-century-twist-to-a-traditional-law-firm-tool), ABA Journal, checked 2026-11-07
9. [What Word Error Rate Is Acceptable for Legal Transcription?](https://www.speechmatics.com/company/articles-and-news/acceptable-word-error-rate-legal-transcription), Speechmatics, checked 2026-11-07