Skip to content
Call wiki

Can AI fix a garbled meeting transcript, and can your tool learn your team's vocabulary?

Last updated 2026-08-17 · Confidence: documented — Granola’s, Fireflies’, and OpenAI’s docs; the Claude-cleanup practice is tried (Alejo’s, dated 2026-08-17).

Yes, on both ends: teach the transcriber your vocabulary before the call, and run the transcript through an LLM after. The two stack — the vocabulary list cuts the errors, the LLM pass fixes the rest.

Before the call: custom vocabulary

Section titled “Before the call: custom vocabulary”
  • Granola: under Settings → Preferences → Language you paste a comma-separated list of terms; keywords “boost recognition of terms when they are spoken”, so a product name stops transcribing as a common word. It’s recognition-only: no find-and-replace, no acronym expansion, and long phrases only match verbatim. Granola says attendee names and your own company are already handled — add external companies and people instead.
  • Fireflies: a custom-vocabulary box under Personal Settings → AI Settings → Customization — “add names, acronyms, or technical terms your team uses regularly.”

After the call: an LLM cleanup pass

Section titled “After the call: an LLM cleanup pass”

Alejo takes each Granola transcript and runs it through a Claude prompt that infers from context which words are garbled — in his experience it “usually does a pretty good job”, and it needs no setup in the transcription tool at all.

OpenAI documents the same pattern for its own transcription models: a post-processing pass with a text model corrects misspellings and “handle[s] longer terminology lists” than a transcriber’s prompt window (224 tokens for Whisper), with one caveat — check corrections against what was actually said, since a confident model can rewrite rather than repair.

Feeding it your team’s names

Section titled “Feeding it your team’s names”

The cleanup prompt works better with the vocabulary in it. Alejo’s suggestion: harvest the list from data you already have — calendar guests, the editors of your Drive documents — via connectors, and paste it into the cleaning prompt (the same list seeds the vendor settings above).

Getting the transcript out to clean it is its own question — see Granola agent access; for what Granola captures at all, see Granola auto-transcribe.

Sources

Section titled “Sources”