How to fix mis-heard words in a transcript

A transcript can get almost every word right and still be useless, because the few words it got wrong are the ones that carried the meaning. Speech recognition is very good at ordinary English. It is bad at exactly the words that make your recording yours: the names of people and places, the jargon of your trade, the product and part names. This guide explains why, what does not fix it, and what does.

Why the important words come out wrong

A speech model turns sound into the most likely sequence of words it knows. When it hears a term it has rarely or never seen, it does not leave a gap. It picks the everyday word that sounds closest, and it does so confidently and consistently, so the same term is wrong every time it comes up.

Here is what that looked like on a real 17-minute phone call between a property owner and a sewer specialist, transcribed with our largest model. These are corrections the people involved confirmed, or that the transcript itself proves:

Notice that none of these is a bad recording. A better microphone would not have helped, because the model heard the sounds fine. It just did not know the words.

What does not fix it

Find and replace

Find and replace only works when you know every wrong form in advance, and you rarely do. The same term came out two different ways on one call, and a blind replace of a common word like "handle" would wreck every sentence where someone really said handle.

Matching by sound

MeetingScribe already corrects close mis-hearings on your device: you describe the recording in a sentence, and terms from it are matched against what was heard by sound and spelling. On an earlier real estate call, nine supplied terms produced seven corrections, every one of them right. On this call it fixed none, and the reason is instructive. Scored from 0 to 1 for how alike they sound, "hand date" and "hand dig" reach only 0.65, "felon" and "belly" 0.54. The matcher needs 0.82 before it changes anything, and lowering the bar does not work: an unrelated pair like "another" and "Andrea" scores 0.48, almost the same. Sound alone cannot separate a mis-hearing from an ordinary word.

Giving the speech model the terms in advance

Whisper, the speech model we use, can be primed with a list of words before it listens. We tested that on the same call. It did get "cleanout" and "gasket" right where it had not before, but it also skipped about one sentence in seven of the conversation entirely. Losing whole sentences to gain a few words is a bad trade, so we did not ship it.

A small AI model on your own computer, used naively

Recent versions of Chrome include a small AI model that runs on the device, which is the ideal private option. Asked to rewrite each section of the call with only a summary of the subject, it missed every one of the errors above and made about fourteen changes that were wrong: "trains running through town" became "roots running through town", and "that's a little crazy" became "that's a little closer". A small model guesses, and on a transcript full of near-misses it guesses wrong with confidence.

Two changes made it useful. Giving it the whole transcript to read first, so it sees the terms people use correctly elsewhere, produced its first real corrections, in about two and a half minutes. And a strict filter keeps only short changes that sound like what was heard, because most of its other suggestions were summaries rather than corrections. On this call that left five suggestions: three right ("hand dig", "gaskets" for "gas kits", and "4-inch pipe" for "4-inch play") and two wrong. Useful as a quick private first pass, with every change checked, but nowhere near a large model.

What does fix it: a large model, with the whole conversation

Every error in the table is obvious to anyone who knows the trade and reads the sentence around it. Nobody digs a pipe by "hand date", and a plumber talking about where water sits in a line means a belly, not a felon. That is the kind of knowledge large general AI models have, and given this transcript, a large model picked out each of the flagged errors from context. The trick is to give it the whole conversation, so it can work out the subject first, and to stop it doing anything except correcting.

A good correction prompt tells the model to:

• work out what the conversation is about before changing anything;
• change only mis-heard words, garbled sayings, and phrases repeated twice by mistake;
• never rephrase, summarize, shorten, or fix grammar, and keep filler words;
• leave speaker labels and timestamps exactly as they are;
• leave a word alone when unsure;
• return the whole transcript in the same format, and nothing else.

Then check every change

Even a good model, told clearly, sometimes rewrites a phrase it should have left, or stops halfway through a long transcript. So the step that matters most is the last one: compare its version with yours, word by word, and decide on each difference. Done by eye on a long transcript that is tedious, which is why most people skip it.

How MeetingScribe does this for you

On the MeetingScribe tool, two buttons sit next to Transcribe.

Quick refine runs Chrome's built-in model on your computer, with the whole transcript read first and the strict filter above. One click, free, nothing sent anywhere. It appears only in browsers that have the built-in model, and the first time, Chrome downloads it.

Refine with your AI is the thorough one: a large model corrects words using the context of the whole conversation, which is what the small built-in model cannot do. It turns the steps above into a few clicks:

1. Transcribe your recording as usual.
2. Click Refine with your AI, then Copy the prompt. It contains your transcript, your one-line description of the recording if you wrote one, and the instructions for the AI.
3. Paste it into ChatGPT, Claude, Gemini or another AI chat and send. If an AI chat sent you to MeetingScribe in the first place, paste it into that same conversation: it already knows what your recording is about, which makes its corrections better.
4. Copy the corrected transcript it returns and paste it back into the page.
5. Review the list of changes. Each shows the old and new words with a few words either side, and a choice of Keep, Edit or Reject. Edit is for when the AI spotted the right place but chose the wrong word: type the right one. Small changes start as kept; any change over eight words starts as rejected, because that is where rewording hides.
6. If the AI missed a word, add it yourself under the list: type what the transcript says and what it should say, and every place it occurs is added as a change you can still reject one by one.
7. Apply. Only the changes you kept are written into your transcript, and they carry into the SRT and VTT subtitle files as well, at the same timings.

Quick refine ends in the same review list. With Refine with your AI, the page also warns you when the reply is much shorter than your transcript, which usually means the AI stopped early, and when it changed an unusually large share of the words.

What it will not fix

• Words that were never captured. If the audio is unintelligible, no model can recover it, and a model that tries is inventing. See how to record a meeting for a transcript.
• Who said what. Text alone cannot tell voices apart, so speaker labels are left as they are. See how to correct wrong speaker labels for that.
• Terms the model has no way to know, like the private name of an internal project. Put those in the "What is this recording about?" box before you transcribe.

A note on privacy

MeetingScribe transcribes on your device, and neither button changes that: the site sends nothing. Quick refine never leaves your computer at all. With Refine with your AI, the correction happens in an AI chat, so the transcript text goes to whichever service you paste it into. For most recordings that is a fair trade for getting the names right. For a recording that must not reach any outside service, use Quick refine and fix the rest by hand; the search box in the transcript editor finds every occurrence quickly.

Questions