Custom vocabulary in meeting transcripts: how to set it up

Run a speech model on a normal meeting and the words it gets wrong are predictable: people's names, product codenames, acronyms, client company names. Not because the model is bad — because nobody says "Kubernetes" or "Sreelakshmi" often enough for a general model to nail the spelling. Custom vocabulary is the fix: you tell the tool in advance what these terms are, and it corrects them as it transcribes. Here's what to add, what to expect, and how it works in QuickSpeak.

Why transcripts fail on exactly these words

Speech recognizers are trained on huge amounts of general speech, so common words arrive with strong priors and rare ones don't. A surname that appears mostly in one region, an internal codename that exists only inside your company, an acronym like "RAG" that the model may resolve to "rack" or "wrack" — the audio is ambiguous, and the model guesses toward whatever is statistically likelier in general English. We walk through the full failure list in why transcripts are inaccurate, but jargon and names are the category you can actually fix yourself.

The fix has a cost structure that explains the market. Cloud tools can retrain or hotword-boost a shared model, so they tend to ration the feature — Otter's pricing page (accessed October 27, 2026) lists custom vocabulary at 5 terms on Basic, 100 names plus 100 terms on Pro, and 800 of each on Business. A local tool has no per-term server cost, so there's nothing to ration. Zoom puts custom vocabulary for its live transcription behind paid plans as well; we couldn't verify the current tier boundaries on Zoom's own pages during this writing, so take that as directional rather than quoted.

What to add (and what not to)

Vocabulary pays off in order of how often the term comes up and how wrong the default spelling is. Before your next meeting, spend five minutes on this list:

1. Every attendee's name. The highest-value entries, because mislabeled names poison the whole transcript's credibility. If the model hears "Jon" but writes "John," add it. Speaker labeling is a separate problem — see how speaker names work.
2. Product and project codenames. Anything your company invented: "Atlas," "Project Periwinkle," "the HALO rollout." The model has zero prior on these.
3. Client and partner company names. Especially ones with unusual spellings or that sound like common words — a client called "Vynt" will not transcribe itself.
4. Domain acronyms and their spoken forms. Add the acronym and how people actually say it: "S O P" → SOP, "kates" → K8s, "postgres" → PostgreSQL if that's your canonical spelling.
5. Recurring technical terms your field mangles: drug names, legal terms of art, finance jargon, non-English loanwords used in English conversation.

What not to add: one-off words from a single meeting (correct them in a manual pass instead), and ordinary vocabulary the model already handles — bloating the list with terms the recognizer gets right anyway just adds match work with no benefit. Treat the list like a living thing: add a term when you see it wrong twice, prune terms from ended projects.

How QuickSpeak does it

QuickSpeak's Vocabulary page — free, no term cap — takes each entry as a canonical term plus optional aliases. If your team says "Q-S" and "Quickspeak" interchangeably, both aliases map to one canonical spelling, and multi-word aliases work too ("quick speak" → QuickSpeak).

Mechanically, corrections run on-device as part of the transcription pipeline, not as a server round-trip after upload. The matcher works in layers: exact single-word matches get replaced with the canonical spelling, alias phrases are matched longest-first so multi-word variants win over partial ones, and acronym-shaped terms (all caps, two or more characters) are kept in their canonical form. Because it's the same local pipeline that produces the transcript, corrections apply live while you record and identically when you're offline — the feature has no network dependency to break.

What this design gives up, honestly: it's a spelling-correction layer, not model retraining. If the recognizer hears a soundalike, vocabulary fixes the transcript; if it hears something too far from any known word, nothing recovers a word that wasn't recognized at all. In practice the first case is the common one — "coober netees" is recognizable audio that just needs the right spelling — which is why a correction list delivers most of what hotword boosting would.

A workflow that survives contact with real meetings

1. Seed the list once with your team's names, codenames, and clients — the four steps above take about ten minutes.
2. Record your next meeting normally. No setup per meeting; the vocabulary applies automatically.
3. Skim for recurring errors at the end. Each mispelled term you spot twice earns a vocabulary entry; one-offs get a manual fix.
4. Let the corrected transcript feed the next step — minutes, action items, or a search across meetings. Garbage spellings in, garbage summaries out, which is why vocabulary earns its ten minutes first.