# Why transcription gets names and jargon wrong, and how to fix it

> For anyone who sends notes from Google Meet or Microsoft Teams calls and keeps correcting the same words: why speech recognition gets surnames, product names, acronyms and ticket ids wrong, which words fail most, why AI notes spread the mistake, and workarounds that work with any notetaker, including a per-client glossary. Ends with correction rules in TrackHeros.

Canonical: https://trackheros.com/blog/why-transcription-gets-names-and-jargon-wrong/  
Author: Sami Kanafani  
Updated: 2026-10-22

**In short:** Transcription gets names and jargon wrong because speech recognition picks the most likely words it knows, and surnames, product names, acronyms and ticket ids are rare or new to it. Spell names in writing, agree how acronyms are said, keep a per-client glossary and check names before sending. TrackHeros adds correction rules per client for Google Meet and Microsoft Teams meetings.

Transcription gets names and jargon wrong because speech recognition chooses between words it already knows, and a client's surnames, product names, acronyms and ticket ids are exactly the words it knows least. This guide is for anyone who sends notes from Google Meet or Microsoft Teams calls, with any notetaker or the platforms' own transcripts, and keeps correcting the same words. It covers why it happens, which words fail most, why AI-written notes spread the mistake, and the workarounds that work in any tool.

## Why does speech recognition get names wrong?

Because it guesses from what it expects to hear. A recogniser turns sound into candidate words and then picks the candidates that are most likely in that language. Common words win. A surname it has never seen, or a product name invented last year, cannot be picked at all, so it writes the nearest word it does know.

That nearest word is usually fluent and confident. "Siobhán" becomes "Shivon". A project codename becomes a common noun. Nothing in the transcript marks the guess, which is why these errors survive into notes.

Accent and audio make it worse but are not the root cause. A clear speaker on a good microphone still gets an unknown name wrong; they just get it wrong the same way every time.

## Which words go wrong most often?

The ones that are rare, new, or written differently from how they sound.

| Kind of word                 | Said                 | Typical transcript         | Why                                                 |
| ---------------------------- | -------------------- | -------------------------- | --------------------------------------------------- |
| Names from another language  | Siobhán, Tadhg       | Shivon, Tige               | Spelling and sound follow different rules           |
| Names with several spellings | Stephen, Mohammed    | Steven, Muhammad           | All are real; the model picks the common one        |
| Product and project names    | A product called UBA | UBVA, "you be a"           | New words, often close to ordinary ones             |
| Technical terms              | Postgres, Kubernetes | post grass, cooper netties | Invented words with no everyday equivalent          |
| Acronyms                     | S-Q-L, "sequel"      | sequel, SQL, S Q L         | Said as letters by one person, as a word by another |
| Ticket ids and versions      | PROJ-142, v2.3       | project 142, V 2 3         | Format exists only in writing                       |

Two patterns stand out. Errors cluster on the words that matter most, because the important words in a client meeting are the client's own names for things. And they repeat: the same mishearing comes back in every meeting until something changes.

## Why do acronyms and ticket ids come out mangled?

Because they are written forms with no single spoken form. One developer says "S-Q-L", another says "sequel", and both are right. A ticket id like PROJ-142 is said as "proj one forty-two", "project one-four-two" or "ticket 142", and the hyphen and capitals exist only on the screen. A recogniser hears the sounds and writes words; it has no way to know the team writes them as an id.

Version numbers, dates and figures said quickly have the same problem: "two point three" might be a version, a price or a deadline.

## Why are names harder in a meeting held in another language?

Because every name is foreign to somebody in the call. In a meeting held in Spanish or German, the English product names sit inside sentences in another language, and the colleagues' names may come from a third. A recogniser expecting one language writes the other language's words as the nearest ones it knows, and names suffer first because nothing around them helps.

Translation adds a second trap. When the transcript or the notes are rendered into English, a product name that is also an ordinary word can be translated as if it were a description: a product called Ledger Line, mentioned in a German sentence, can come back as "the ledger line", a phrase rather than a name. Note in the glossary which names must never be translated, and check them first. Which languages a meeting can be held in, for TrackHeros at least, is on [the languages page](/languages/).

## Why does it matter more once an AI writes the notes?

Because the notes are built on the transcript, and a summary hides where each word came from. A misheard name can be copied into the decisions and action items. Worse, a model writing notes can tidy a misheard term into a plausible word that was never said, and once it is in a fluent sentence nobody questions it.

The cost lands in the minutes. A client who sees their own product misspelled, or their head of product's name wrong, in a document you sent, reads it as carelessness, however good the rest is.

## What can you do before and during the meeting?

Put the words in writing before the call, and say them clearly once during it.

- **Spell names in the invitation or agenda.** Whether a given notetaker reads the invitation varies, but the person checking the notes always can.
- **Introduce new people with their role.** "This is Siobhán, the client's head of product" gives the reader, and any model writing notes, something to anchor the name to.
- **Agree how acronyms are said** within a team, and stick to it.
- **Say ids slowly**, and use the same words for them each time: "ticket one-four-two".
- **Repeat key figures and dates** when they are agreed: "so, the twelfth of March, and fifteen thousand".
- **Use the tool's custom vocabulary** if it has one, kept per client.

## How do you keep a glossary for each client?

One short table per client, started on the first call and added to whenever you correct the same word twice. Keep it with that client's other files, not in a shared list, because the same abbreviation can mean a bank to one client and a build server to another.

| Term as written | How people say it | What it is                    | Mishearings seen |
| --------------- | ----------------- | ----------------------------- | ---------------- |
| UBA             | "you-bee-ay"      | The client's banking partner  | UBVA, you be a   |
| Siobhán Byrne   | "shiv-AWN"        | Head of product               | Shivon, Chevonne |
| PROJ-           | "proj", "project" | Ticket prefix                 | project 142      |
| Ledger Line     | "ledger line"     | The product; never translated | ledger lane      |

The last column is the useful one. It is your find-and-replace list before sending notes, and it is exactly what a custom vocabulary feature needs.

## What should you check before sending notes?

The words a client will notice, in this order:

1. **Every person's name**, especially in action items and the attendee list.
2. **The client's own names**: company, products, projects, systems.
3. **Ids, versions and figures**, against the recording if they matter.
4. **Each term on the glossary's mishearings list**, with find-and-replace.

Who said something is a different question from how a word is spelled; [why transcripts get speakers wrong](/blog/why-transcripts-get-speakers-wrong/) covers that half.

## How does TrackHeros handle names and terms?

With correction rules, kept per client. A vocabulary rule says what the transcript writes and what it should be: "UBVA" should be "UBA", with optional context such as "the bank". An instruction rule states a convention in one sentence, such as "Ticket ids are always PROJ- followed by a number", and is added to every transcription and summary prompt for that client. Rules are created on the Rules page, by asking the assistant in that client's chat, or from "Propose a correction" on a meeting.

Rules apply to every meeting summarised from then on. To fix a past meeting, ask the assistant in that client's chat to apply your vocabulary rules to it or to rewrite its notes; both wait for your approval. A rule belongs to one client and is never used for another client's meetings. The full description is on [correction rules](/features/correction-rules/), the most common case is walked through in [fixing a name the transcript keeps getting wrong](/answers/fix-a-name-the-transcript-gets-wrong/), and how corrected terms reach the client is on [meeting minutes](/features/meeting-minutes/).

## Frequently asked

**Why does my transcript spell a colleague's name differently every time?**

Because the recogniser has no record of that name, so each time it writes the nearest spelling it does know, and small differences in how the name is said produce different guesses. Names from a language the model hears less often are the most affected. A glossary of correct spellings, or a custom vocabulary entry in your tool, fixes it.

**Does spelling a name out loud help AI transcription?**

Sometimes, but do not rely on it. Spelling helps a person checking the transcript and can help a model writing notes from it, but the transcription itself may simply record the letters. Putting the name in writing, in the invitation or a shared document, is more dependable.

**What is custom vocabulary in a meeting notetaker?**

A list of terms you give the tool so it writes them correctly: product names, people, acronyms, ticket formats. It works best kept per client, because the same abbreviation can mean different things in two engagements. In TrackHeros it takes the form of correction rules on each client.

**Why do AI meeting notes make transcription errors worse?**

Because the notes are written from the transcript. A misheard term can be copied into decisions and action items, or quietly replaced with a plausible word that was never said, and the minutes then put that error in writing to the client.

---

TrackHeros is in early access: https://trackheros.com/#early-access