Custom Vocabulary for Live Captions: How to Get Names and Jargon Right

Custom vocabulary is a list of names, places, acronyms, and specialized terms that you give a captioning system before a session starts, so it favors those words over similar-sounding common ones. It does not make captions perfect. It fixes the errors your audience notices most: the pastor’s name, the drug name, the product line, the case caption. This guide shows what to put on the list, how to prepare it, how to test it, and where the limits are.

TL;DR

• Custom vocabulary tells the recognizer which rare words to expect. It helps most with proper nouns, acronyms, and domain terms.
• Keep the list short and specific: names, places, and terms people will actually say. Confirm spellings with the speakers.
• Test with a short rehearsal and a printed term list. Count the misses, fix the list, and run it again.
• Medical and legal captions need extra care. AI captions can still misread abbreviations and context, so they are not a substitute for a certified service where one is required.
• Live Caption AI lets a speaker enter up to 100 Custom Terms before going live and adds Medical, Worship & Ministry, and Finance & Earnings models.

Table of Contents

• What custom vocabulary actually does
• What belongs on your term list
• How to prepare the list before an event
• Medical terminology captions: what helps and what still fails
• Legal terminology captions: be honest about the bar
• Speaker labels: what to expect
• How to test your captions before the real event
• Perspective: why we built it around the term list
• Using Live Caption AI for custom vocabulary captions
• Accessibility duties: a short note for U.S. organizations
• FAQ
• Sources

What custom vocabulary actually does

A general speech recognizer is trained on everyday speech. It hears “Nehemiah,” “Kowalczyk,” or “metoprolol” and picks the likeliest common words that sound alike. The result reads like a bad guess.

Speech recognition tools handle this with hints. A typical cloud speech service lets you supply phrase hints and custom word lists that the model should favor, including medical terms, product names, and uncommon proper nouns. That is the whole idea. You are not retraining anything. You are shifting the odds toward words you know are coming.

Two layers work together:

1. A domain model. It is tuned to a field’s vocabulary and phrasing, such as clinical, ministry, or finance language. This is what people mean by domain specific speech recognition.
2. Your own term list. This covers what no model can know: your speakers, your building, your clients, your program names.

The domain model handles the field. The list handles your room.

What belongs on your term list

A list is only useful if it targets words the recognizer would otherwise get wrong. Think about what a reader would be embarrassed to see misspelled.

Names and places

• Speaker and guest names, including unusual spellings
• Staff, board members, and honorees
• Venue, campus, street, and neighborhood names
• Organization and program names

Acronyms and internal shorthand

Acronyms are the biggest trap. If your team says “MOU” or “CCM,” the recognizer may write something that sounds right and means nothing. Add the form you want to see on screen.

Domain terms

• Worship and ministry: book names, theological terms, song titles, ministry names
• Medical: drug names, procedures, department names
• Finance: ticker-style names, product names, metrics your CFO uses
• Legal: case names, statute names, Latin phrases, party names

What to leave out

Skip common words. Padding the list with everyday vocabulary adds noise and no accuracy. Aim for the dozens of words that matter, not hundreds that don’t.

How to prepare the list before an event

You already do some of this for human providers. The City of New York’s accessible events guide tells organizers to give providers a detailed run-of-show, speaker names, other proper nouns, specialized terminology, and any prepared script. See its Accessible Events Guide. The same package works for AI captions. Just turn it into a term list.

A simple routine:

1. Pull the source material. Read the program, slides, script, sermon notes, or agenda. Highlight every name and unusual term.
2. Add likely spoken terms. Think about what people will say, not only what is written. An interpretation glossary template recommends including likely spoken terms and asking speakers to confirm spelling and meaning.
3. Confirm spellings with the person. Names are not guessable. A two-line message to the speaker saves a correction later.
4. Rank by risk. If you are over the limit, keep the words that would confuse or offend if wrong.
5. Enter the list before going live. In Live Caption AI that is a field in session setup, up to 100 terms.

One caution from the same glossary guidance: a list of words cannot tell the system whether a speaker means a proposal, an approved decision, or an exception. Vocabulary fixes spelling. It does not fix meaning. Your speakers still need to speak clearly.

The habit carries over to other fields. Teachers building word lists from learners’ real gaps get better results than teachers using generic lists, as this piece on AI in English language centers shows. Captions work the same way: the list should come from what will really be said.

Medical terminology captions: what helps and what still fails

Medical speech is full of drug names, procedures, and abbreviations. A domain model and a custom list help with the spelling of these words. They do not remove the risk.

Abbreviations are ambiguous. A study on clinical abbreviations notes they often need context to resolve. For example, “MS” can mean multiple sclerosis or mental status. A caption that prints the letters is not wrong, but it may leave the reader unsure. A separate 2021 study of medical abbreviations says normalization remains difficult because of misspellings, spelling variation, plurals, and unresolved recognition and disambiguation challenges.

A high score is not clinical correctness. A 2022 survey of diagnostic captioning found that measures based on word overlap can miss clinical errors. It gives “no pneumothorax” versus “pneumothorax” as an example. One missing word flips the meaning. So when you test medical captions, read for meaning, not only for spelling.

Practical steps for a clinic:

• Put the clinic’s drug names, department names, and clinician names on the term list.
• Ask clinicians to say abbreviations in full when the meaning matters.
• Treat captions as support for the conversation, and confirm key instructions another way, such as in writing.

On the legal side, the ADA expects effective communication in health care. The Department of Justice says that for telehealth, providers are responsible for ensuring effective communication, which may include real-time captioning platforms or sign language interpreters. For more on how we think about clinics, read our guide to medical live captions and live captions for telehealth patients.

Legal terminology captions: be honest about the bar

Legal speech adds case names, Latin phrases, and statute references. A custom list can carry the names and phrases you know in advance: party names, the judge, counsel, the case caption.

But be clear about what is at stake. The National Court Reporters Association’s guidelines for CART say the service requires a skilled steno writer or voicewriter with a total error rate of 98 percent or higher. Live Caption AI is an AI captioning tool. It is not certified or human-reviewed, and we do not offer a legal vocabulary model. For depositions, hearings, or any setting where a record or a certified provider is required, book that provider. For internal briefings, client presentations, and training sessions, a term list plus AI captions can make the room more accessible at a small cost.

Speaker labels: what to expect

Readers often ask for speaker labels, meaning the tag that shows who is talking. The technique behind them is called diarization, and it is harder than it looks. Some research systems improve it with setup. One clinical project, for example, uses a short pre-session audio sample from doctors to help distinguish speakers.

The lesson for organizers is practical, whatever tool you use:

• Labels can go wrong when people talk over each other, share a microphone, or sound alike.
• One speaker per microphone and one voice at a time helps every captioning system.
• If attribution matters, such as a panel with several people, have the host introduce each speaker by name. Add the names to your term list so they appear correctly.

We would rather tell you to plan for this than promise a label feature that is perfect.

How to test your captions before the real event

This is the part most people skip. A ten-minute rehearsal tells you more than any feature list.

1. Write a test script. Include 15 to 20 of your hardest terms in natural sentences. Read it aloud at normal speed.
2. Run it twice. First without your term list, then with it. Use the same speaker and the same microphone.
3. Mark every miss. Count wrong names, wrong terms, and wrong meanings separately. Meaning errors matter most.
4. Fix the list. If a term is still wrong, check your spelling, then try adding a nearby variant.
5. Test in the room. Stand where attendees will sit. Open the captions on a phone. Check that the text is readable and the delay is acceptable.
6. Ask a real reader. If someone in your community uses captions, ask them to read along and tell you what confused them.

A simple scoring table keeps the test honest:

If the list does not move the numbers, shorten it and retest. Fewer, better terms usually win.

Perspective: why we built it around the term list

We built Live Caption AI for people who cannot book a stenographer for every service, class, or meeting. Most of them hit the same wall first: generic speech-to-text garbles the words that matter most, which are names and specialized vocabulary. A church gets a pastor’s name wrong. A clinic gets a drug name wrong. Trust drops fast.

So we put vocabulary up front. The Medical model, the Worship & Ministry model, and the Finance & Earnings model give you a head start in the field. Custom Terms, up to 100 names, places, and terms, cover what only you know. The speaker types them in before going live. No file upload, no setup call.

We are also direct about the limits. Live Caption AI is not certified, not human-reviewed, and it does not make an organization ADA-compliant on its own. It can be one part of your effective-communication plan. Where a certified provider is required, use one. Where you have no captions today, an affordable AI option beats silence.

— Ryan

Using Live Caption AI for custom vocabulary captions

Here is how it works in practice:

• The speaker runs Live Caption AI on a phone or laptop.
• Attendees scan a QR code and read captions on their own phones.
• No app to install. No account to create. No hardware. No stenographer.
• Pick a domain model, then add up to 100 Custom Terms.
• Add live translation into one of 29 languages per session. Viewers can switch between the original and the translation.
• Save transcripts as TXT, SRT, JSON, or PDF. AI summaries are included. Session audio is never stored.

The Professional plan is $19.99 a month with unlimited captioning. The Free plan gives basic on-device captions, and the Business plan starts at $19​9 a month. See every plan on the pricing page, or read about live captioning for medical, legal, and professional settings.

Explore Live Caption AI

For more on keeping costs low, browse our posts on affordable captioning solutions.

Accessibility duties: a short note for U.S. organizations

The ADA requires public accommodations to provide auxiliary aids, such as real-time captioning, when necessary for effective communication, unless doing so causes an undue burden or fundamentally alters the service. What counts as effective depends on the setting, the length and complexity of the exchange, and the person’s needs.

In practice:

• Ask attendees what they need. Do not assume.
• Offer captions in advance and say how to reach them.
• Keep a fallback ready for high-stakes settings, such as a certified captioner or interpreter.
• Talk to your own counsel about your obligations. A tool alone does not settle them.

FAQ

Does custom vocabulary make live captions 100 percent accurate?

No. It raises the odds that names and specialized terms print correctly, but speech recognition can still miss words in noisy rooms or with overlapping speakers. Test before the event and keep a fallback for high-stakes settings.

How many terms should I add?

Start with the 20 to 50 words most likely to be wrong, then add more if the rehearsal shows gaps. In Live Caption AI the limit is 100 Custom Terms per session. A focused list usually works better than a long one.

Do I need a domain model if I have a term list?

The two do different jobs. A domain model covers the general vocabulary of a field, and the term list covers your own names and terms. Use both when your session mixes the two, such as a church service with guest speakers or a clinic with named staff.

Can AI captions replace a certified captioner?

Not where a certified or human-reviewed service is required, such as court proceedings. Live Caption AI is not certified or human-reviewed. For meetings, classes, services, and events, it can be an affordable way to add captions where there were none.

Sources

• Accessible Events Guide, NYC Mayor’s Office for People with Disabilities
• An interpretation glossary for events
• Custom vocabulary and phrases in Cloud Speech-to-Text
• Deciphering clinical abbreviations
• A deep database of medical abbreviations and acronyms
• Diagnostic captioning: a survey
• NCRA Guidelines for CART Captioners
• HealthTalk D3.1 Clinical requirements and baseline systems
• ADA Requirements: Effective Communication
• ADA Telehealth guidance

Recommended

• Speaker Labels in Captions: When to Use Them and How
• Organizers: Teams Live Events Captions, Phone QR and 85–90% Accuracy
• Copy Ready Captioning Style Tokens for Production Teams
• Creators: Use Captions, SDH, or Subtitles Correctly for Accessibility