Why meeting transcripts get names and numbers wrong

It writes the most likely sequence of words it knows, and a colleague's surname, a product code or a figure said quickly is exactly where "most likely" and "what was said" part ways. The quality of the audio decides how often that happens more than most people expect.

This guide explains where the errors come from, which ones matter, and what you can change. For the practical fixes, see getting better meeting transcripts.

How a recogniser decides what you said

A modern recogniser does not look words up one sound at a time. It takes a stretch of audio and produces the text that best fits both the sounds and the patterns of language it learned in training. That second part is why transcripts read smoothly: when a sound is ambiguous, the model picks the word that makes the sentence sensible.

It is also why errors look plausible. A misheard word is rarely nonsense; it is usually a real word that fits the sentence and sounds close to the one spoken. That makes errors easy to read past, which is the real risk in a meeting transcript.

The hard words

Names

Personal names, company names and product names are the classic failure. Many are rare in the text a model learned from, some are not in it at all, and they are spelled in ways the sound does not predict. "Priya Shah" and "Marc Dubois" may come out right; a less common surname may come out as an ordinary word. The same name can also be spelled differently in different parts of one transcript.

Numbers

Numbers fail in quieter ways. "Fifteen" and "fifty" differ by one stressed syllable. "Two to three weeks" can become "223 weeks". A figure can be written as a word in one place and digits in another. Because a number rarely makes a sentence ungrammatical, the language model gives little help in catching it.

Jargon and acronyms

Internal project names, abbreviations and technical terms have the same problem as names. An acronym spoken as letters may be written as letters, as a word, or as a similar-sounding phrase. If your team says "the Harbour migration" twenty times a meeting, expect the first few to be wrong until the context makes it likely.

Short words that carry meaning

"Can" and "can't", "do" and "don't", "is" and "isn't" can flip the meaning of a sentence. A dropped "not" is one error by any count and the most important one in the meeting.

What matters more than the engine

The choice of recogniser matters, but for meetings the signal usually matters more.

• Distance from the microphone. A voice close to the microphone is loud against the room. The further away, the more the room's echo and noise are mixed in.
• The microphone itself. A laptop's built-in microphone sits next to its own keyboard and fans, and at arm's length from your mouth. A headset microphone near the mouth gives a much cleaner track.
• Speakers instead of headphones. On speakers, your microphone hears the other side as well as you. Their words end up in your track, faint and with the room's echo, and are transcribed twice or badly. See fixing echo in call recordings.
• Cross-talk. When two people talk at once, a recogniser usually follows one voice and drops or garbles the other.
• The other side's audio. What arrives from the call has already been compressed by the meeting app, and may carry their room, their microphone and a weak connection. You cannot fix their end, but you can ask someone who keeps breaking up to repeat a figure.
• Accent and speaking rate. Recognisers are not equally accurate for every accent. The Earnings-22 benchmark, a set of 119 hours of earnings calls from companies around the world, was published in 2022 precisely to show that commercial systems perform differently depending on where speakers are from (Del Rio et al., 2022, checked 25 September 2026).

Which errors to worry about

Most transcription errors do not matter. A wrong "the" or a missing "uh" changes nothing. The ones worth checking are the words someone will act on:

1. Names attached to commitments. "Tom will send the contract" is only useful if it is Tom.
2. Numbers. Prices, dates, deadlines, quantities.
3. Negations. Anything where "not" changes the decision.
4. Terms that will be searched later. If you will look for "Northwind" in a month, a transcript that wrote it three different ways will not find all three.

This is also why a single accuracy percentage tells you less than it seems. Word error rate counts every wrong word the same, so a transcript can have a good score and still get the one number that mattered wrong.

How errors carry into AI notes

A summary written from a transcript inherits its mistakes. If the transcript says "fifty" where someone said "fifteen", the summary will state fifty with confidence. The language model has no way to know the transcript is wrong; it is working from the text alone.

So when an AI-written action item or figure looks odd, go back to the transcript line, and if the transcript looks odd too, go back to the audio. A tool that keeps the recording and lets you play the moment a line was said makes that check take seconds rather than a rewatch.

How Notey handles this

Notey transcribes on your Mac with Apple's speech recognition, live, as the meeting happens. On-device transcription covers how that works and what it is good for. Three things in Notey bear on accuracy:

• Two tracks. Your microphone and what the Mac plays are recorded and transcribed separately. Each line is attributed to "You" or "Them" by the track it arrived on, not by guessing from the voice. If your speakers put someone on your side, you can move the line.
• The recording is kept, on your disk. Click a timestamp in the transcript to hear that moment, and correct the line if it is wrong.
• A second pass for English. After an English meeting, Notey reads the recording again on your Mac and keeps the words the recognisers agree on. Why Notey re-transcribes a meeting after it ends explains what it changes and when it does not run.

None of this makes names or numbers safe to take on trust. Check the words someone will act on.

Frequently asked questions

Why does my transcript spell people's names wrong?

A recogniser writes the most likely word sequence it has learned. An unusual name is rare or absent in what it learned from, so it is replaced by a common word that sounds similar. Spell names out in the meeting, or correct them in the transcript afterwards.

Are numbers in a meeting transcript reliable?

Check them. Numbers are short, often spoken quickly, and can be written several ways ("fifteen", "15", "50"). Prices, dates and quantities are worth confirming against the recording or in the follow-up email.

Does a better transcription engine fix these mistakes?

Partly. A stronger model makes fewer errors overall, but names, jargon and cross-talk stay hard for every engine. A clear signal from a good microphone often makes a bigger difference than the choice of engine.

Is one accuracy percentage enough to judge a transcript?

No. A single figure counts every wrong word the same, so a transcript can score well and still get the one name or number that mattered wrong. Look at which words are wrong, not only how many.