Custom vocabulary: teach transcription your names and jargon: ThinkScribe

Speech recognition is trained on general language, so a colleague's surname, your product's name or a drug name is exactly what it gets wrong most often. Custom vocabulary for transcription is the fix: ThinkScribe keeps a custom dictionary on the device, and a word you add once is used for every future transcription and dictation. This guide explains why these words go wrong, shows the three ways to add them, and sets out what a dictionary can and cannot do.

Why speech models get names and jargon wrong

A recognizer does not hear letters. It turns sound into the most probable sequence of words, and probable is decided by what it was trained on. Three things make a word improbable however clearly you say it.

It was never in the training vocabulary. Training data is dominated by ordinary language. A surname shared by a few thousand people, a compound place name, a molecule, an internal project code: none appear often enough to be learned, so the model reaches for the nearest thing it knows. That is why an unusual name comes back as two ordinary words that sound roughly the same.

Homophones. Many names are acoustically identical to everyday words: Reed and read, Wright and right, Cole and coal. Nothing in the sound separates them, so only context decides, and a model with no reason to expect a name picks the common word.

Rare words in ordinary sentences. Recognizers weigh each word against the words around it, so a long drug name between two plain English words competes with much likelier alternatives. The same goes for acronyms said as words and for a term from a second language dropped into your main one.

None of this is fixed by speaking more slowly. The word has to be made likely, which is what a dictionary does.

How to add a custom word to transcription

On Mac, open Settings from the sidebar and choose the Vocabulary tab. On iPhone, open Settings and scroll to Custom Dictionary.
Type a word or phrase in the Add Word field. Multi-word phrases are fine: a full name, a product line, a project code name.
Click Add. It appears under Your Words.

Double-click a word on Mac, or tap it on iPhone, to edit it later. Clear All removes the list.

Teach a word from a transcript

The quickest route on Mac, because you are looking at the mistake:

In any transcript, select the wrong word or phrase.
Right-click and choose Teach Correct Word….
Type what it should have been and confirm.

The transcript is corrected on the spot and the correct word joins your dictionary. ThinkScribe also notices when you correct the same word three times by editing and adds it for you.

What vocabulary biasing actually changes

Two things happen with your words:

Hints while recognizing: with the Whisper engine, your words are handed to the recognizer before it decodes, so when the audio is ambiguous it leans toward them. A name it has never seen becomes a likely choice instead of an impossible one.
Correcting afterwards: with every engine, a match-the-whole-word pass fixes the common near-misses, such as a name split into two ordinary words or a product name with the wrong spelling.

Both steps run on your Mac or iPhone. The dictionary is never uploaded.

Biasing recognition versus correcting the text afterwards

These are two different repairs, and they fail differently.

Biasing acts while the audio is decoded, so it can rescue a word that would otherwise never be produced at all. The audio still has to be close: a name pronounced nothing like its spelling stays out of reach however heavily it is weighted.

Correcting acts on the finished text, so it is predictable and works whatever produced the words, but it can only fix what still looks like a mistake. If a surname was heard as a different ordinary sentence, nothing is left to match against.

So prefer distinctive words. "Renner" is a safe entry. "Bill" is not, because half its occurrences are ordinary English and a replacement pass cannot tell which is which.

What is worth adding to a custom dictionary

Good candidates:

People: colleagues, clients, patients, students, the guest you interview every week.
Organisations and products, including your own.
Domain terms the engine mangles: drug names, legal Latin, acronyms said as words, place names.
Terms in a second language that come up inside your main language.

By field:

Medicine: the drug names you dictate most often, procedure and device names, anatomical terms from your specialty, the named signs and scales, and the surnames of referring clinicians. Ward and clinic names count too, being local proper nouns no general model has seen.
Law: case names you cite repeatedly, the Latin that survives in practice, statute short titles, and the names of counsel, judges and firms in your matters.
Engineering and software: component, library and service names, the code names of projects and environments, and part families. Anything that looks like a normal word but is not, such as a service named after an animal, is worth adding, because the model will always prefer the animal.

Not worth adding:

Ordinary words the engine already gets right.
Spelling variants of the same word; add the one you want.
Very short acronyms said letter by letter, which are usually recognized fine as letters.
Words that came up once and you have already fixed by hand.

The cheapest way to build a list is to note the words you correct twice in your first week. Those, and only those, belong in it.

How the same problem is solved elsewhere

Two familiar cousins of vocabulary biasing set useful expectations.

Large language models are steered the same way when you paste context into a prompt: an agenda, a list of attendees, a glossary. Nothing about the model changes, but the words you supplied become far likelier in its output. That is biasing, applied to generated text rather than decoded audio.

Text replacement in the keyboard settings on macOS and iOS is the other half of the idea, applied to typed text. It shares the same weakness, that a rule cannot read context, which is why a good table stays short and unambiguous.

Use it with voice typing and with recordings

The same dictionary applies everywhere: voice typing on Mac, the voice keyboard on iPhone, recordings and imported files. There is no separate list per feature and nothing to sync manually; if Sync Across Devices is on, corrections you make on one device are reflected in transcripts on the others.

For medical and legal work

Dictionaries are where private, on-device recognition pays off. Drug names, procedure codes and case names are the words a general model stumbles on, and the words you least want to send to a server to fix. Private dictation for medical and legal work goes into the workflow; this guide is the mechanics.

Frequently asked questions

Does adding a word guarantee it is recognized?

No. Biasing makes the word much more likely when the audio sounds close to it, and correction catches common near-misses, but a name pronounced very differently from how it is spelled can still be missed. Adding the phonetic spelling as a second entry sometimes helps.

Is there a limit?

The dictionary holds up to 500 entries. In practice, keep it much shorter for the best results.

Can I import a list?

Add entries one at a time in Settings or from a transcript with Teach Correct Word…. A list built from real mistakes over a week or two beats a pasted glossary.

Why does the same word come out right in one recording and wrong in the next?

Because biasing only tips a decision that was already close. Clear speech from a near microphone gives a strong signal and your word wins easily; a noisy room or a fast aside gives a weak one, and the ordinary word the model has heard a million times wins instead.

Does the dictionary leave my device?

No. It lives in the app's storage on your Mac or iPhone. Your transcripts and corrections only go to iCloud if you turn sync on, and then only into your own private iCloud database.