Can You Teach an AI Notetaker Your Company's Names and Jargon?

Yes, but the length of the list is the wrong thing to compare

Every serious notetaker now has somewhere to type the words it keeps getting wrong. Granola calls them jargon, while Otter and Fireflies call it custom vocabulary. When people compare tools on this, they usually compare how many terms each one allows.

That number matters less than a plainer question: who keeps the list up to date? A transcript rarely fails on the words you already knew to add. It fails on the surname of a prospect you met for the first time this morning, the company their CFO came from, and the codename someone made up in last Tuesday's planning call. By the time anyone thinks to add those, the call that needed them is over.

What each tool documents

Granola has two lists. Personal jargon holds up to 30 terms and applies to your own meetings. Workspace jargon holds up to 50 terms, applies to everyone in the workspace, and only admins can edit it, although members can see it. Granola's docs also say you do not need to add the names of people attending your meetings, so attendee names appear to be handled for you.

Otter is the most generous on size. Its help page says the free plan includes up to 5 words or phrases, Pro gives a personal vocabulary of 100 names and 100 other terms, and Business adds a team vocabulary of 800 names and 800 other terms on top of a larger personal one. Team terms apply to every member's conversations.

Fireflies has a Custom Vocabulary section in its settings, next to an industry setting that tunes the speech model to the kind of meetings you have. Its own guide says custom vocabulary is limited to Pro and Business accounts. That guide was first published in 2021, so confirm the plan detail on the current pricing page.

We could not find a vocabulary setting in Fathom's public help pages while writing this. That does not prove there is none, so ask Fathom directly if this matters to you.

A list is a chore with an owner

A word list is accurate as of the last time someone edited it. On a sales team that opens fifteen new accounts a month, that is a weekly job, and it belongs to whoever is admin of the notetaker, who is rarely the person on the call with the new account.

Caps make it worse. A 30 or 50 term list fills up fast once you add clients, products, competitors and internal projects. From then on, every new name means deleting an old one, and nobody can tell which old one will matter next quarter.

So when you evaluate this feature, ask where the words come from as well as how many fit. A list filled from your calendar invites and the people you already work with stays current without anyone remembering to update it. A list typed by hand is only as good as the person keeping it.

A word list does not repair the past

Granola's docs are direct about this: the terms boost recognition when the word is spoken, and they cannot replace one word with another in the transcript. In practice a vocabulary list helps calls that happen after you add the term. The call last week where the client's name came out as two common words stays that way.

That matters more for search than for reading. When you read a transcript, you can work out that a misheard phrase meant your client. When you search your meetings for that client's name three months later, the transcript with the wrong spelling does not come back, and you will not know it is missing. If finding past calls by client or project name is why you are buying, the spelling of names is a search problem as much as a transcription one.

Your team list is also a client list

A shared vocabulary is useful because everyone's calls get the same names right. It also means the list is visible to the whole workspace. Granola's docs say members can see the workspace terms even though they cannot edit them.

Think about what a good list contains: every client, every prospect in a live deal, and the codenames of projects that have not been announced. You would hesitate to post that in a company wide channel. Before you paste your CRM's account list into a team vocabulary, decide who should be able to read it, and keep sensitive names in personal lists where the tool offers them.

A ten minute test

Write down ten words your transcripts get wrong today: three people's names with unusual spellings, two client or partner companies, two product or project names, two acronyms your team says out loud, and the name of your own company.

Record a short call with a colleague in which each word comes up twice in normal sentences. Count how many of the ten come out right without any setup. Then add the words to the tool's list, record the same call again, and count again. Finally, search the tool for each word and see which of the two calls comes back.

The first count tells you what the tool does on a call with someone new. The second tells you what the list buys you. The search tells you what you will be able to find later, which is usually the reason you wanted the transcript.

Where Driffle stands

Driffle has this problem too. In one of our own recorded meetings, the transcript heard the word Driffle as trifle. Our product name is now always on the list.

Driffle does not have a settings box where you type your own word list in the current release. Instead, for each call it builds a list of up to 20 terms for the speech model, most specific first: the names of attendees on the meeting's calendar invite, then people and organisations already in your workspace that you are allowed to see, then project and topic names from recent meetings. A name that only appears in a colleague's private notes is never used. The list is checked again during long calls, so it can change as the conversation does.

Know the limits. Twenty terms is a hard ceiling set by the speech engine, so a large call will not get every name. Someone who is not on the invite and has never come up before gets no help on the first call. And you cannot yet pin a term yourself. Driffle runs on Macs with Apple Silicon, captures your Mac's microphone and system audio with no bot in the call, and does its processing in the cloud.

If you want to run the ten minute test on Driffle, it is at driffle.ai/download.