Teaching a Notetaker Your Names, Products and Jargon
If your colleague's name comes out spelled four different ways and your product name keeps turning into an ordinary English word, there is a setting for that. Most transcription tools let you supply a list of terms they should expect, and per Otter's own documentation (checked 2026-08-04), that covers proper names, jargon, acronyms, and unusual spellings of names. The feature is easy to find once you know it exists. The part nobody explains is what to put in the list, and the counterintuitive rule that a longer list makes your transcripts worse rather than better.
What a custom vocabulary actually changes
A speech model works on probability. Hearing an ambiguous sound, it picks the most likely word given everything around it. Common words win by default, which is why an unusual product name loses to an ordinary word that sounds similar, and why a surname it has never encountered becomes the nearest familiar one.
A custom vocabulary puts a thumb on the scale. You are telling the tool that these specific unusual terms are plausible in your meetings, so it stops discarding them in favor of common words.
Two things follow, and both matter later. It biases toward the terms you supply, which is exactly why an over-long list causes harm. And it only helps when the audio was good enough to be ambiguous rather than unintelligible. If a word was never audible, no vocabulary list recovers it, which is a job for the fixes in why your AI meeting transcript is inaccurate.
Where the setting lives
The feature usually sits in account settings rather than inside an individual meeting, which is part of why people miss it. In Otter, per its own documentation, the path is Settings, then Manage Vocabulary, and Otter documents custom vocabulary as a paid feature available on its Premium and Teams plans.
Other tools offer equivalents under names like custom vocabulary, keywords, or a dictionary, and availability and capacity vary by product and plan. Check your own vendor's current plan page rather than assuming, since what is included changes.
Where a team feature exists, use it. A vocabulary maintained once for everyone beats each colleague separately teaching the same tool the same six product names.
Building the list from your own transcripts
Do not brainstorm this list. You will produce a tidy set of terms that were never the problem, and miss the ones costing you most.
Instead, mine what you already have. Open your last month of transcripts and search for the terms that recur in your work: your product names, your top clients, your colleagues' surnames, the acronyms your team says constantly. Look at what the tool actually produced.
Ten minutes of this gives you a better list than an hour of thinking, because it is evidence rather than guessing. You will usually find the same six or eight terms mangled over and over, and those are your list.
A useful trick when a term is hard to search for: search for what the tool heard rather than what was said. If your product "Nexa" keeps becoming "next a," searching your archive for "next a" finds every instance fast.
What to add, in priority order
Work down this order, because the return drops sharply after the first two groups.
- Names of people you meet with weekly. Highest frequency, highest cost when wrong. A misspelled name in a summary you forward is the error most likely to embarrass you.
- Your product, service and project names. Especially anything that is also an ordinary word, which is the hardest case for the model and the most common in practice.
- Client and company names. Same reasoning, particularly for anything you will send back to that client.
- Acronyms your team uses constantly. These are usually rendered as separate letters or as a similar-sounding word.
- Genuine domain jargon. Technical terms specific to your field that a general model has little reason to expect.
Stop there. Anything appearing once a quarter is not worth a slot, and the next section explains why the slot has a cost.
What not to add, and why more is worse
The instinct is to add everything, on the theory that more context helps. It does the opposite.
Do not add common words. Biasing toward a word the model already knows makes it more likely to appear where it should not, which introduces new errors in place of the old ones.
Do not add anything that sounds like a common word unless you need it. "Kai" as a name will start appearing in place of "kind of," and you have traded a rare error for a frequent one.
Do not add one-off terms. A term from a single project that ended is pure noise once the project is over.
Prune it. A vocabulary list is not an archive. When a client leaves or a product is renamed, remove the old entry. A list that grows for two years and is never cut becomes a slow drag on accuracy that nobody thinks to blame.
The mental model: every entry is a small bet that this unusual term is more likely than the common word it resembles. A few well-chosen bets pay off. A hundred careless ones lose.
Names are the hard case
Names deserve their own treatment because they cause the most damage and behave the worst.
Add the spelling you want to see written, since that is what shows up in your notes. If a name has a common variant, decide which one your notes should use and add only that, because adding both invites the tool to alternate between them within a single transcript.
Nicknames are worth a thought. If your notes are cleaner with "Marcus" throughout, do not also add "Marc." If people genuinely use both and both should appear, add both and accept some inconsistency.
Names also interact with speaker attribution. A tool that recognizes a name spoken aloud has a better chance of connecting it to the right speaker, which is one of the anchors described in why your AI transcript has wrong speaker labels. Getting the vocabulary right helps that job too.
What it will not fix
Set expectations honestly, because people over-invest here after their first success.
A custom vocabulary does not fix bad audio. If someone is far from the microphone, in a noisy room, or talking over a colleague, the sound was never clear enough for a preference to apply. That is entirely an audio problem.
It does not fix speaker attribution, which is a separate task with separate causes.
It does not fix a summary that missed the point, which is a weighting problem covered in fixing a bad AI meeting summary.
Its actual job is narrow and worth doing: making sure the words you say every week appear correctly in every transcript. Ten minutes once, five minutes a quarter after that. For the tool-specific walkthrough, our guide to using Otter covers setup around this feature, and the hub guide on how to choose an AI meeting note taker covers the wider criteria.
Feature names, locations and plan availability change often. Confirm the current details on your vendor's own documentation before relying on any specific one here.
FAQ
What is custom vocabulary in a transcription tool?
A list of terms you supply so the tool expects them. Per Otter's documentation, that covers proper names, jargon, acronyms and unusual spellings, and it biases the model toward those terms instead of the common words they resemble.
Where do I find the setting?
Usually in account settings rather than in an individual meeting. In Otter the documented path is Settings, then Manage Vocabulary. It is a paid feature there, and availability varies by tool and plan elsewhere.
How do I know what to add?
Search your own past transcripts for the names and terms you use weekly and see what the tool produced. If you cannot find a term, search for what the tool misheard instead.
Can I just add everything?
No, and this is the common mistake. Every entry biases the model, so a long list containing common words or one-off terms introduces new errors. Keep it short and prune it when things change.
Will custom vocabulary fix a bad transcript?
Only the specific problem of known terms being misrendered. It does nothing for poor audio, cross-talk, or wrong speaker labels, which have separate causes and separate fixes.