Start Free Trial
← All posts
Teach the engine your words

Fix terminology errors with a custom dictionary

Engines transcribe by probability, so the rare words your session turns on lose to common words that sound like them. The dictionary is the lever you control.

1 lever you control 6 min read Published July 2026

A live caption engine has never seen your agenda. It transcribes by probability: given these sounds, what words does language usually make? For everyday speech, that works startlingly well. For the words your session actually turns on — the drug, the witness, the product, the case number — it fails in one specific, predictable way: the engine replaces words it doesn't know with common words it does. A custom dictionary is the lever that fixes this whole class of error, and it's the lever you control most directly.

Why engines mishear exactly the words that matter

Speech recognition is trained on enormous amounts of general language, so it carries strong expectations about which words follow which sounds. When audio is ambiguous — and live audio always is — the engine resolves the ambiguity toward the statistically likely option. "Xarelto" is not statistically likely. "Zarelto," or worse, "za relto," is what a general-language model reaches for when it hears an anticoagulant it never trained on.

This is why a caption feed can be excellent and useless at the same time. An engine can transcribe 49 of 50 words correctly and still miss the only word the sentence existed to deliver. Overall accuracy is measured across all words; your session's value is concentrated in a few of them. (For how accuracy is actually scored, see our word error rate explainer.) The words that concentrate that value are almost always the same kinds:

  • Proper nouns — participant names, company names, places. "Nguyen" becomes "when."
  • Drug and procedure names — invented pharmaceutical words with no general-language neighbors.
  • Product names and SKUs — deliberately unusual spellings marketing chose for distinctiveness.
  • Case numbers, citations, and codes — strings with no linguistic pattern for the engine to lean on.
  • Field jargon and acronyms — words that are common inside your domain and rare everywhere else.

What a custom dictionary actually does

A custom dictionary tells the engine, before the session starts, that certain low-probability strings are high-probability *here*. It doesn't change how the engine hears — it changes how the engine decides. When the audio could plausibly be your term, the term wins instead of losing to a common near-homophone. The effect on the errors that matter is immediate:

What was saidWhat a general engine hearsWith the term preloaded
Xarelto"zarelto" / "za relto"Xarelto
Farxiga"far sika"Farxiga
voir dire"vwar deer"voir dire
Ms. Nguyen"Miss when"Ms. Nguyen
Kubernetes"cooper netties"Kubernetes

Be clear-eyed about the limits, too. A dictionary can't rescue audio the engine never heard cleanly — a term mumbled under crosstalk stays lost. And it can't referee between two common words that are both plausible in context. It fixes the guessing problem, which happens to be the problem specialist sessions have most.

How to build a dictionary for each assignment

The good news: the terms an assignment will use are almost never a mystery. They're sitting in documents you already have. Ten minutes before the session is usually enough.

Mine the paperwork

Agendas, slide decks, case files, pleadings, discharge summaries, product briefs. Skim for every capitalized word, every italicized term, every string you'd have to look up. Those are your entries.

Collect the names

Every participant, organization, and place likely to be spoken aloud. Names are the highest-frequency terminology errors in most sessions, and the most noticed — people always catch their own name misspelled.

Add the codes

Case numbers, statute citations, SKUs, acronyms. Decide the written form you want once — "24-CV-1187," not three competing renderings — and enter it that way.

Skip the common words

A dictionary is for rare terms. Loading it with everyday vocabulary adds nothing and can nudge the engine toward your entries where ordinary words were right. Rare and specific beats long and thorough.

Keep it and grow it

A glossary built for one client compounds for the next assignment with them. Interpreters have kept term lists forever — the dictionary is that habit, made executable. Our interpreter prep checklist folds this into a full pre-session routine.

The mistakes that quietly undo the work

A typo in the dictionary propagates everywhere. Enter "Farxagi" and the engine will now faithfully produce your misspelling all session. Verify entries against a written source, not memory.
Dumping a whole document in. Hundreds of common words dilute the signal. Curate — the twenty terms that would actually be misheard, not the two thousand that wouldn't.
Building it mid-session. By the time you've noticed the error pattern live, it has already hit the transcript your client is reading. The dictionary is prep, not triage.
Assuming built-in vocabularies cover your niche. Medical and legal term banks are broad, but your specific parties, products, and case numbers are yours alone to add.

Make it a five-minute habit

The pattern that works is boring and repeatable: when the assignment lands, skim the documents, pull the names and terms, enter them, done. Five to ten minutes, once. Every session after that starts with an engine that already knows the words the day depends on — which is the closest thing to a free accuracy upgrade that live captioning has.

Frequently asked

What is a custom dictionary in live transcription?

A custom dictionary is a list of terms you give a speech recognition engine before a session — names, drug names, product names, case numbers, jargon. It tells the engine those rare strings are expected here, so ambiguous audio resolves to your term instead of a common word that sounds similar. In Unicaption, it sits alongside built-in medical, legal, and government terminology.

Why do live captions get names and drug names wrong?

Because engines transcribe by probability learned from general language. A rare word like a drug name or a surname has no strong statistical support, so the engine substitutes the nearest common word — "Xarelto" becomes "zarelto," "Nguyen" becomes "when." Preloading those terms in a custom dictionary is the direct fix; the engine stops guessing once it knows the word is expected.

How many terms should I add to a transcription dictionary?

Curate rather than dump: typically a few dozen terms per assignment — the names, codes, and jargon a general engine would actually mishear. Skip everyday words; they add nothing and can nudge the engine wrongly. Verify spellings against a written source, because a typo in the dictionary reproduces itself in every caption.

Does Unicaption have built-in medical and legal terminology?

Yes. Unicaption includes built-in medical, legal, and government terminology, plus a per-user custom dictionary for the names and terms specific to your assignment. Captions run in real time, and session content is never stored or used to train models — relevant when your glossary comes from case files or patient-adjacent documents.

Load the words before they're spoken

Built-in specialist terminology plus your own dictionary, live in 60+ languages. 30 free minutes every week — no credit card.

Start Free Trial →