Back to Blog
AI Tools
Sep 26, 2026
23 min read

Custom Vocabulary for AI Transcription: How We Built Key Terms, and How 18 Apps Compare (2026)

In one benchmark, AI transcription got people's names wrong 3 times as often as words overall. Here is how we built custom vocabulary into MeetingsAI, and how 18 apps compare.

I. M.

iPhone showing the MeetingsAI Key terms screen with the names Rukfaš, Chombe and Fazlinović, while a glowing sound wave and three glass tags float toward it, with the MeetingsAI logo.
#New Feature#AI Tools#Productivity#Private Mode

Introduction

Custom vocabulary is a list of words you give an AI transcription app, such as people's names, company and product names, acronyms, and industry jargon, so it spells them correctly instead of guessing. In MeetingsAI, that list is called key terms. One list works across live transcription, recorded meetings, uploaded files, voice input, translation, and Cheat Mode.

If you have ever searched a transcript for "Priya" and found "Pria," "Prius," and "Maria" instead, you know the problem. The meeting went fine. The notes did not, and now search cannot find them either.

This post covers why speech AI trips over names, how we built key terms into MeetingsAI (with the real limits), how our approach compares with the custom vocabulary features in 18 other apps, and how to set it up in about two minutes.

One disclosure up front: we make MeetingsAI. We checked every competitor claim against that company's own help, pricing, or blog pages on September 25, 2026, and linked each one so you can check too.

Key Takeaways

  • AI transcription misspells names because the model leans toward common words. In a benchmark of real earnings calls, speech systems got people's names wrong about 3 times as often as words overall.
  • MeetingsAI merges three sources into one list: your saved vocabulary (up to 300 terms), your AI Profile, and key terms for one session (up to 100) with a one-line topic.
  • Each speech engine gets the list in the format it accepts, across live transcription, recordings, uploads, voice input, translation, and Cheat Mode.
  • We checked 18 other apps. 16 have a vocabulary feature, set per account or workspace. None of the meeting note takers documents a list for a single meeting.
  • In Private Mode, your terms stay on your phone.

Why AI Transcription Gets Names and Jargon Wrong

Speech recognition models learn from huge amounts of recorded speech. When a sound could be several different words, the model leans toward the words it has seen most often. Common words win. Your words lose.

The research is blunt about it. In Earnings-21, a benchmark of real earnings calls presented at Interspeech 2021, six speech systems got 11 to 18 percent of all words wrong, but 42 to 52 percent of people's names and 31 to 44 percent of company names. In a study of 45 hours of MIT lectures, a widely used speech model got 32.3 percent of names and technical terms wrong, compared with 7.0 percent of other words.

So the mistakes are predictable. Mistakes like these show up in transcripts that were never given a vocabulary:

  • A colleague named "Priya" becomes "Pria."
  • Your product "Nimbus" becomes "nimble us."
  • "SOC 2" becomes "sock two," and "OKRs" becomes "okay ours."
  • A client named "Siobhan" comes out as "Shivon," because the model knows the sound but has never seen the spelling.

One wrong word looks small. It costs more than it looks:

  • Search breaks. You search your notes for "Nimbus," and the meeting where it came up does not appear.
  • Summaries copy the mistake. The summary, the action items, and the AI chat are all built from the transcript. A misspelled name in the transcript becomes a misspelled owner on a to-do.
  • Trust drops. When a client's name is wrong, you start to doubt every other line.

Custom vocabulary fixes the cause, not the symptom. You tell the engine which rare words to expect before it starts listening.

How Custom Vocabulary Works in MeetingsAI

We built key terms as one small service that sits between you and the speech engines the app uses. Here is what happens, step by step.

Step 1: Your terms come from three places

  • My Vocabulary. Your permanent list, in Settings > AI Profile > Words. It holds up to 300 terms: colleagues, clients, your company, your products. Paste many at once, separated by commas. It saves as you type.
  • Your AI Profile. MeetingsAI reads your name, your topics, and the capitalized names and acronyms in your Personal AI Profile, and adds them for you. Write "I run growth at Acme Corp and report OKRs for Project Atlas," and it picks up Acme Corp, OKRs, and Project Atlas. One switch turns this off.
  • Key terms for this session. Before you record, stream, translate, or start Cheat Mode, tap Key terms and add the names for this meeting only (up to 100), plus one line about the topic. The list stays ready for your next session until you clear it.
MeetingsAI AI Profile screen on the Words tab: My words shows Drip, Rizz and Looksmaxing (3 of 300) with an Add a word field, and the Words from my profile switch is on.
Settings > AI Profile > Words: your permanent list, saved as you type.

Step 2: We merge and clean the list

When a session starts, the three sources become one list. Session terms go first, because they are the most specific. Your saved list comes next, then your profile terms.

The list is cleaned on the way. Duplicates are removed no matter how they are capitalized, and your first spelling wins. Extra spaces are collapsed. Anything longer than 64 characters is dropped, because at that length it is a sentence, not a term.

Step 3: Each speech engine gets the list in the shape it accepts

This is where most of our engineering time went. MeetingsAI uses different speech engines for different jobs, and each one has its own rules:

  • Live transcription: up to 100 terms of up to 50 characters each, plus a short description of the situation.
  • Recorded meetings and uploaded files: up to 200 terms of up to 6 words each.
  • Voice input on iPhone: up to 100 terms, plus the description.
  • The on-device recognizer: your terms become bias phrases on the phone itself.
  • Languages our main engines do not cover: a backup cloud engine gets a short glossary of your terms.

Our vocabulary service trims the list to fit each engine, so a long list never breaks a transcription. If the engine for recordings and files rejects the list anyway, the app retries once without it. You still get your transcript.

Step 4: A one-line topic, not instructions

With your terms, the engine gets a short description, such as "A recorded business meeting. Topic: Q3 pricing review." It tells the model what kind of words to expect.

It describes the situation and nothing more. It never includes your name or your job title. If there are no terms at all and no topic, nothing extra is sent, exactly as before this feature existed.

Step 5: Each recording remembers its own terms

Record Meeting saves a copy of the session's key terms with the recording. If the final transcript is made a few minutes later, or you retry after a network drop, it still uses the list you set before the meeting, even if you have changed the list for another session since.

Why a list per meeting? Because short, relevant lists work better. The Contextual Earnings-22 benchmark (March 2026) found that giving speech systems a keyword list improved key-term recognition in every system it tested. It also found that long lists full of terms nobody says can cause false matches, and that concise, relevant lists did best. That is why session terms go first.

Voice Input Gets a New Dictation Engine

Voice input is the short push-to-talk clip you record in AI chat, folder chat, summary chat, and to-dos. On iPhone, it now runs on a speech engine built for exactly that job, and it uses your key terms.

  • Made for short clips. Up to 2 minutes each.
  • Your words, spelled right. It gets your merged key terms and the one-line topic.
  • Mixed languages. It covers 32 languages and listens for up to 3 in one clip: English plus your default language, or the languages set on your phone. A team that switches between Turkish and English mid-sentence still gets one clean line.
  • Clean text. It removes filler words and resolves self-corrections, so "let's meet Tuesday, no, Wednesday" becomes "let's meet Wednesday."
  • Quiet fallback. If your language is not supported or the engine has a problem, the clip goes to our other cloud engine with the same key terms. You never see an error for it.

Translate Conversation on iPhone uses the same engine for short turns when you pick a fixed source language. Names come through right before they are translated, which matters a lot for live translations between two people who use each other's names.

Where Your Key Terms Apply

Key terms follow you through AI transcription in MeetingsAI:

  • Record Meeting: the live preview while you record, and the final transcript with speaker labels.
  • Live Streaming: real-time captions.
  • File upload: in Cloud mode, your saved list and profile terms also apply to audio and video files you import.
  • Voice input: AI chat, folder chat, summary chat, and to-dos.
  • Translate Conversation: each turn is transcribed with your terms before it is translated.
  • Cheat Mode: the live transcript uses your terms, and the AI answer gets a "spell exactly as written" list, so the answer on your screen spells your client's name the way they do. More on that in Cheat Mode: Get Live AI Answers in Meetings.
  • Summaries, action items, and AI chat: they are built from the transcript, so they inherit the right spelling.

Real Scenarios: What Key Terms Fix

A renewal call with a new client

You are meeting Siobhan Keane from Fenwick Robotics about Nimbus, your product. Before the call, you add three session terms and one topic line: "Renewal call for Nimbus with Fenwick Robotics." The transcript, the summary, and the to-do "Send Siobhan the Nimbus pricing" all come out spelled the way she spells them.

An engineering standup

Your team says "gRPC," "Terraform," "p99 latency," and "Kafka" ten times a day. Add them to My Vocabulary once. Standups, voice notes, and Cheat Mode sessions use them from then on.

A university lecture

A biochemistry lecture is full of words a general model rarely hears: "Krebs cycle," "NADH," "Michaelis-Menten." Add the week's terms as session key terms before class, and your notes and quiz questions use the right spelling.

A voice note on the way out of a meeting

You hold the mic button and say, "Remind Ahmed to send the SOC 2 report to Acme Corp by Friday." With Ahmed, SOC 2, and Acme Corp in your list, the to-do reads exactly like that, instead of "Remind Ahmad to send the sock two report to Acne Corp."

Custom Vocabulary in 18 AI Transcription and Note-Taking Apps, Compared

Most AI note takers now offer some kind of custom vocabulary. The details differ a lot: how many terms, who can edit the list, which plan you need, and what the list affects. We checked the official help and pricing pages of 18 apps on September 25, 2026. Limits change often, so each entry links to its source.

AI meeting note takers

  • Otter: "Custom vocabulary," split into names and other terms. Basic (free) gets 5 terms, Pro gets 100 names plus 100 other terms, and Business gets team lists of up to 800 plus 800 (pricing). Lists are personal or team-wide, and Otter's help center shows how to edit them on the web (help).
  • Fireflies: a custom vocabulary setting under Settings > AI Settings > Customization (help). Its help center does not publish a term limit, and the current pricing page does not say which plans include it.
  • Notta: "My Vocabulary" allows 3 English terms on Free, 50 on Pro, and 200 on Business, per its help center. The pricing page shows higher combined numbers. It is one list per account, it applies to new transcriptions only, and the English list can only be edited on the web.
  • Granola: "personal jargon" holds up to 30 terms, and "workspace jargon" holds up to 50 that only admins can edit (docs). Granola says jargon boosts recognition but cannot replace one word with another.
  • Fathom: the "Transcription Dictionary" holds 50 terms for the whole organization, and only admins can edit it (help).
  • tl;dv: the "Team Custom Dictionary" is one admin-managed list for the whole team, and it applies to future recordings (help). tl;dv's own blog lists it as a Business plan feature.
  • Read.ai: up to 100 terms, each tagged as a person, product, or jargon, on paid plans. There is one list for the whole workspace, and any member can edit it (help).
  • Krisp: 750 words shared across the workspace, for future meetings. You can add a word by clicking a low-confidence word in a transcript (help).
  • Jamie: 50 terms on Free, Plus, and Pro, 100 on Team, and 250 on Enterprise (pricing).
  • Plaud: "Custom Vocabulary" holds up to 500 case-sensitive words, with presets for legal, medical, and tech. You set it in the mobile app or on the web, and it applies to new recordings. Older ones need a manual re-transcribe (help).

Video meeting platforms

  • Zoom: custom dictionaries in AI Studio hold up to 1,000 terms, US English only. Only admins can set them up, and they need a paid Zoom Workplace plan plus a ZoomMate license (Zoom support).
  • Microsoft Teams: the Copilot Custom Dictionary is a CSV file of up to 1,000 entries per language that an admin uploads for the whole organization, with optional "sounds like" spellings. It covers scheduled meetings (not 1:1 calls), the organizer needs a Copilot license, and it can take up to 24 hours to switch on (Microsoft Learn).
  • Google Meet: we found no vocabulary setting in the help pages for transcripts or "Take notes for me".

Dictation and transcription tools

  • Apple (iPhone dictation, Voice Memos, Notes): no word list you can edit. Apple says dictation personalizes on the device and may use your contact names (Apple privacy).
  • Wispr Flow: a "Dictionary" that learns words automatically, can map a misspelling to the right spelling, and syncs across devices, on the free plan (help).
  • Superwhisper: "Vocabulary" hints plus "Replacements," exact find-and-replace rules that run after transcription. The free plan runs local models on your device (docs).
  • Descript: the "Transcription Glossary" holds 30 terms per Drive, English only, and adds a word by itself after you correct it 3 times (help).
  • Rev: the "AI Glossary" holds up to 1,000 phrases of up to 5 words each, English letters only, on every plan including Free (help).

What the comparison shows

  • One list for everything. The lists are set per account or per workspace and apply to everything you transcribe. None of the meeting note takers documents a list for a single meeting.
  • Admins and paid plans. Fathom, tl;dv, Zoom, and Microsoft Teams put the list in an admin's hands. On free plans, Notta allows 3 terms and Otter allows 5.
  • Future recordings only. No app updates past transcripts by itself. Plaud at least lets you re-transcribe.
  • Transcripts first. Plaud and Zoom say the list also improves summaries. None of these apps documents one list that covers live meetings, uploaded files, and dictation together.
  • Web first. Otter's help center shows how to edit the list on the web, and Notta's English list is web-only. Plaud and Wispr Flow let you manage it in the phone app.
  • Privacy is rarely mentioned. None of the meeting note takers says the list stays on your device. Among the dictation tools, Superwhisper keeps everything on the device when you use its local models, and Apple personalizes on the device but has no list you can edit. Microsoft even advises against putting personal or sensitive information in its dictionary.

Where MeetingsAI is different

  • A list for this meeting, not just your account. Session key terms and a one-line topic, added on the setup screen seconds before you start. That matches what the research says works best: short and relevant.
  • One list, six places. Live transcription, recordings, uploaded files, voice input, translation, and Cheat Mode answers all read the same merged list.
  • Terms you do not have to type. Your AI Profile fills in your name and the capitalized names and acronyms you already wrote about your work.
  • Set it on your phone, right before the meeting. No admin, no web dashboard.
  • Private by design. Your lists are saved on your phone. In Private Mode, they only guide the on-device recognizer.
  • Same limits for everyone. 300 saved terms and 100 per session, with no bigger plan to buy for more.

What other apps do that we do not (yet)

Fair is fair. Here is what we saw elsewhere and do not offer today:

  • Learning from your transcripts: Descript and Wispr Flow add words when you correct them, and Krisp lets you add a word with one click from the transcript.
  • Find-and-replace rules: Superwhisper and Wispr Flow can swap one word for another after transcription.
  • "Sounds like" spellings: Microsoft Teams lets admins say how a term is pronounced.
  • Re-transcribing old recordings: Plaud lets you run an older recording again with your new list.
  • Bigger lists: Rev, Zoom, and Teams allow up to 1,000 terms. Given the research above, bigger is not always better, but it is an option we do not have.

For a wider look at how MeetingsAI compares beyond vocabulary, read MeetingsAI vs Otter, Fireflies, Fathom, tl;dv and Notta.

Key Terms in Private Mode: Stored and Used on Your Phone

Private Mode runs transcription on your phone, and your key terms follow the same rule: they only guide the on-device recognizer, even offline on a plane. They work in Record Meeting, Live Streaming, Translate Conversation, Cheat Mode, and voice input.

In Cloud mode, the list travels with the audio for that one transcription. Your lists are saved on your phone, not in your MeetingsAI account, and MeetingsAI does not store them on a server.

Two more details we built in on purpose, in both modes:

  • If your lists are empty and you set no topic, the app sends nothing new, exactly as before the feature existed.
  • The topic line never includes your name or your job title.

Cheat Mode answers come from a cloud AI in both modes. With Private Mode on, only the terms you typed for that session are added to the answer prompt. Your saved list and your profile terms stay on the phone.

What Key Terms Do Not Do (Yet)

We would rather you know the limits up front:

  • Key terms guide the engine. They do not find and replace. A noisy room or a strong accent can still cause a miss, just less often for the words you listed.
  • They apply when a transcript is made. Adding a term today does not rewrite last week's transcript.
  • Voice input on Android uses our cloud engine with your key terms as a glossary. The new dictation engine is iPhone only for now.
  • Live transcription on Android needs Android 13 or later to use key terms. In Cloud mode, recordings and uploads use them on any version.
  • Two paths skip key terms today: files you transcribe in Private Mode, and voice input when Apple Intelligence is your transcription engine.
  • Long lists get trimmed per session. The engines take 100 to 200 terms at a time, and session terms go first.

How to Set Up Custom Vocabulary in MeetingsAI

It takes about two minutes.

  1. Open Settings > AI Profile > Words. No profile yet? The My Vocabulary list is right on the AI Profile screen. New users also see it during onboarding, on the "Words We Should Know" screen.
  2. Add your permanent terms. Type names, companies, products, and acronyms. Separate them with commas to add many at once.
  3. Check "Words from my profile." Leave the switch on to add your name and the capitalized terms from your profile automatically.
  4. Before your next meeting, tap Key terms. You will find it on the setup screen of Record Meeting, Live Streaming, and Translate Conversation, and in the Knowledge Base card in Cheat Mode. Add the names for this meeting and one line about the topic.
  5. Record, then check the names. Open the transcript and search for one of your terms.
MeetingsAI Key terms sheet titled Spell it right, with the session terms Rukfaš, Chombe and Fazlinović, the topic line It's about a Bosnian hit TV series, and 3 terms added automatically.
Tap Key terms before a session to add names for that meeting and a one-line topic.

Write the topic line like a calendar title: "Q3 pricing review with the Nimbus team." Describe the meeting. Do not give instructions.

Tips for a list that works

  • Spell each term exactly as you want it written, capitals included: "SOC 2," "iOS," "McKinsey."
  • Add rare words, not common ones. "Budget" does not need help. "Ahmed Al-Rashid" does.
  • Keep terms short. One to three words is ideal. Some engines drop long phrases.
  • Put one-off names in session key terms, not in your permanent list. Session terms go first, so they are never the ones that get cut.
  • Keep the permanent list focused. Terms nobody says in your meetings can cause false matches.
  • Clear session terms after a one-off meeting, so they do not crowd the next one.

Frequently Asked Questions

What is custom vocabulary in transcription?

Custom vocabulary is a list of words you give a speech-to-text engine before it transcribes, such as names, company and product names, acronyms, and jargon. The engine uses the list to prefer those spellings when it hears a matching sound. It does not edit the transcript afterwards. It changes what the engine expects to hear, which is why it can fix errors that editing and search miss.

How do I get voice to text to spell a name correctly?

Give the app the spelling before you speak. In MeetingsAI, add the name to My Vocabulary or to the session's key terms, and voice input sends it along with your audio. The iPhone keyboard's own dictation has no word list you can edit. Apple says it personalizes on the device and may use your contact names, so saving the person as a contact is the closest option there.

How many custom vocabulary terms can I add in MeetingsAI?

You can save up to 300 terms in My Vocabulary and add up to 100 key terms per session, each up to 64 characters. MeetingsAI also adds terms from your AI Profile. When a session starts, it merges the lists and trims them to what each engine accepts, 100 to 200 terms, with session terms first, so the names for today's meeting are never the ones cut.

Does custom vocabulary work offline in Private Mode?

Yes. In Private Mode, MeetingsAI transcribes on your phone, and your key terms guide the on-device recognizer without leaving the device, even with no internet connection. This covers Record Meeting, Live Streaming, Translate Conversation, Cheat Mode, and voice input. On Android, it needs Android 13 or later. Files you transcribe in Private Mode do not use key terms yet.

Does adding a term fix transcripts I already recorded?

No. Key terms work while a transcript is being made, because they change what the engine expects to hear. A term you add today helps new recordings, live sessions, uploads, and voice notes, but it does not rewrite finished transcripts. A recording does keep its own copy of the session's key terms, so a delayed or retried transcription still uses them.

Which AI note takers have custom vocabulary?

Otter, Fireflies, Notta, Granola, Fathom, tl;dv, Read.ai, Krisp, Jamie, and Plaud all offer some form of it, and so do Zoom and Microsoft Teams. Google Meet does not, based on its help pages. Most use one list per account or workspace, often on paid plans or for admins only. MeetingsAI adds per-meeting key terms, automatic profile terms, dictation, and on-device use in Private Mode.

Conclusion

A transcript is only as useful as the names in it. Speech engines guess at words they rarely hear, and names, products, and acronyms are exactly those words. Custom vocabulary tells the engine what to expect.

We built key terms so the right list reaches each engine in the right shape: a saved list, your AI Profile, and a short list for the meeting in front of you. In Private Mode, it all stays on your phone.

Tired of fixing the same names in every transcript? Download MeetingsAI, add five names to Settings > AI Profile > Words, and record your next meeting.

Share this article

I. M.

Related Articles