Making Granola's notes trustworthy: fixing misheard words at the source

Thinking

Granola's summary, action items and Chat answers all come from the transcript, so one wrong word is wrong everywhere. And I can't fix it: selecting a line offers Copy, Delete and Report, and Report leaves my notes wrong. Transcripts can't be edited;⁵ editing is something Granola is "keeping in mind for future improvements".⁶

Other reviews say the same,⁷ but most come from competitors, so I count them as signals and nothing more.

Opportunity

Granola already cares. Its jargon lists (April 2026)¹ take names and terms before a meeting, but its docs list the limits:² jargon "does not replace one word with another," caps at 30 personal terms, is English only and doesn't learn. With no audio kept,⁴ there's nothing to check a mistake against.

Language is the bigger gap. Granola transcribes 31 languages on desktop, but not Nigerian Pidgin or Nigerian English, and its docs don't mention switching languages mid-sentence.⁸ Speech models struggle too: on African-accented English, one major study found Whisper got about 31% of words wrong and Google's speech API about 58%, and that models 'fail woefully' at African names.⁹

The gap: Granola can try to prevent errors, but can't correct one once it lands or learn what a local word means.

Success metrics

Set before designing, to judge the ideas against.

Constraints

• No audio. Corrections have to work from context alone.
• Trust. An edit must never make someone appear to say what they didn't.
• Reuse what exists: the selection bar and jargon lists.
• Free plan, desktop first.

Job to be done

Ideas

• 

  A · Fix it in the transcript

  Edit a line; "Fix all" updates every mention and the summary. For slang or Pidgin, keep the words and add their meaning.

• 

  B · Catch words before

  Before a meeting, suggest names and terms from the invite and past fixes for my word list.

• 

  C · Review queue after

  After a meeting, list low-confidence words to confirm or fix.

Prioritisation

I scored each idea from 1 to 10 on six factors: effort to build and launch, and impact on people and Granola. These are outside estimates, to check with the team.

Decision: A. A and B are both quick wins, but A has more impact (24 against 18), and B can't fix a word that still comes out wrong. B is a good follow-up on the same jargon list. From C, I kept one light piece: a quiet underline on words the model isn't sure of.

The design

Fix once, everywhere, and teach Granola what you mean

• Edit joins Copy, Delete and Report in the selection bar.
• "a bio" → "Adebayo". Fix 3 more, and the summary? Matches are listed first and can be unticked.
• Keep words, add meaning. Slang and Pidgin stay as spoken; I add the meaning, so the summary reads them right.
• Each fix and its meaning is saved to my word list (Granola's jargon list), so the word comes out right next time.
• Edited lines are marked; hover shows the original.

Trade-offs and learnings

• Not every 'wrong' word should be replaced. Pidgin and slang are real language, so the fix keeps their words and teaches Granola what they mean.
• "Fix all" can over-correct. One sound can be two words, so it always asks first.
• No playback. Granola keeps no audio by design, so nearby lines carry the context.
• Checking the docs changed the design. I first sketched a new vocabulary feature, then found jargon lists already existed. Connecting fixes to them is simpler and fits the product.
• Next: an unmoderated Maze test: five users fix a planted misheard name and a Pidgin phrase, measuring completion rate and time on task.