# Teach Blabb your words

Add the names, drugs and product codes Blabb keeps getting wrong, and they stay right from then on.

Every job has a handful of words no general speech model has ever seen — a colleague's surname, a
product code, a matter name, a drug your ward uses daily. You teach them to Blabb once.

Settings → **Writing style → Saved words**.

## Adding a word

Three fields, in a row:

| Field | What goes in it |
| --- | --- |
| **When you say** | The word as it sounds. What you actually say out loud. |
| **Use this spelling** | The spelling you want typed. |
| **Add** | Saves it. |

So if Blabb hears "Nguyen" as "win" every time, you put what it hears in the first box and `Nguyen`
in the second.

Worth teaching:

- **People** — surnames, and first names with unusual spellings.
- **Drug names** — brand names in particular, which rarely follow English spelling rules.
- **Products and codes** — internal names, SKUs, model numbers.
- **Matter and client names** — anything that appears in the same document every day.

Each saved word has its own row in the list, and you can change or remove it later.

## One language at a time

A saved word belongs to the language you were dictating in when you taught it.

That is deliberate rather than a limitation. A spelling that is right in one language is wrong in
another — a correction you made while speaking English should not fire while you are speaking
French and reach for a word that sounds similar. So the correction applies in the language you
taught it in, and nowhere else.

The list opens on your current language, because that is the set actually in play. A filter lets
you look at the others.

## Why a taught word sticks

Saved words are used twice, in two different places in the pipeline:

1. **As a hint to the speech model**, before it decides what it heard. This nudges recognition
   towards your spelling in the first place.
2. **As a correction pass over the raw transcription**, before any cleanup style sees the text.
   This one is deterministic — no model is deciding anything, it is a straight replacement.

So by the time a cleanup style is handed your words, your spelling is already in them. What
protects it from there is the check Blabb runs on every rewrite: if the cleanup model comes back
having changed a term you taught it, that counts as the rewrite having gone wrong, and the whole
rewrite is thrown away in favour of the text that went in. Either way, the spelling you asked for
is the spelling that gets typed.

> [!NOTE]
> **Reset all settings** leaves your saved words alone — it puts settings back to default and does
> not touch your history or the words you taught. **Delete Blabb data**, in Settings → Privacy,
> does remove them, along with your history and logs.

## When a word still comes out wrong

If a word is wrong in a way teaching does not fix — a whole phrase mangled, or words dropped
entirely — the cause is usually audio rather than vocabulary. Work through
[When words come out wrong](/docs/accuracy/when-words-come-out-wrong/).

## See also

- [Learning from your corrections](/docs/cleanup-styles/learn-from-corrections/) — the automatic
  version of this page.
- [Dictating in another language](/docs/cleanup-styles/language/) — and what that means for your
  saved words.
- [Medical and legal](/docs/editions/overview/) — tens of thousands of specialist terms you do not
  have to type in by hand.