# Dictation is slow

What normal looks like, why the first dictation after a break is slower, and the settings that actually change the wait.

Before changing anything, check what you are seeing is actually slow. Blabb is doing speech
recognition and an AI rewrite on your own machine, and both take real time.

## What normal looks like

From the moment you finish to the moment text appears, expect a second or two for a sentence and
longer for a paragraph. Roughly, that time goes:

| Stage | What it costs |
| --- | --- |
| Closing off the last segment | About 0.7 seconds of continuous silence |
| Transcription | The bulk of the wait on Original |
| Cleanup | Roughly as much again, more on the heavier styles |
| Typing the text in | Depends on length and text entry method |

If that is what you are seeing, nothing is wrong. If you want it shorter, the levers are below.

## The first dictation after a break

The cleanup model is unloaded after 15 minutes of not dictating, and loaded again on your next
dictation. That first result after a break is noticeably slower than the ones that follow it, and
then everything is quick again.

The speech model stays loaded, so this only affects the cleanup half of the work.

Leaving Blabb running does not prevent this — 15 minutes idle is 15 minutes idle whether the app
is open or not. What it does prevent is the app starting from cold.

## The first dictation after the PC has been asleep

Same cause, more of it. Coming back from sleep, the models have to be brought back into memory
before anything can happen, so the first dictation after you wake the machine can take a while.
It is a known rough edge. The second one is normal.

## No compatible graphics card

Cleanup runs on the graphics card by default and falls back to the processor when there is no
compatible card. The processor works — it is just slower, and the warm-up is longer.

Blabb does not announce the fallback at the moment it happens. The message lives in Settings
instead: if you switch from **Original** to any cleanup style on a machine that scores poorly for
this work, it appears as you save the change, and it appears every time you make that switch.

> No graphics card was detected, so AI cleanup runs on the processor and is slower — the first
> result takes a moment to warm up. To skip cleanup for instant text, set the style to "Original".
> Your setting is unchanged.

That advice is the real fix on a machine without a graphics card. See
[Lower-spec machines](/docs/performance/lower-spec-machines/).

## Long dictations take a moment to type in

Two of the four text entry methods put your text in one character at a time — **Type fast** and
**Type slowly**. That is what makes them work in awkward applications, and it is also why a long
paragraph visibly types itself in rather than appearing at once.

If you dictate long passages, **Paste** delivers the text in flushes rather than character by
character, which is noticeably quicker to land. It uses the clipboard to do it — your copied text is
put back afterwards, but anything non-text you had copied, such as an image, is not. See
[Text entry methods](/docs/configuration/text-entry-methods/).

## Heavy cleanup styles cost more than light ones

Fastest first:

| Style | Cost |
| --- | --- |
| **Original** | No cleanup model runs at all. As fast as Blabb gets. |
| **Fix typos** | The lightest thing the model can be asked to do. |
| **Clear**, **Professional**, **Friendly**, **Screen match** | More rewriting, more time. |

A specialist **Transcription mode** (Medical or Legal) adds a spelling pass and extra context for
the cleanup model. Worth it if you need the vocabulary; pure cost if you do not.

## Two things that are genuinely silent

**Very low memory cancels the dictation.** If available system memory drops below 500 MB, Blabb
cancels rather than push the machine into thrashing. Nothing appears on screen when this happens —
only the log records it.

**A transcription backlog drops speech.** Blabb transcribes while you are still talking, and it
holds at most 20 segments waiting their turn. On a machine that cannot keep up, anything past that
is discarded and written to the log, so words are missing from the finished text with nothing shown
on screen to say so.

So if a dictation vanishes with no text and no error, or comes back with a stretch of what you said
simply absent, on a machine that is already short of memory or struggling, that is what happened.
Close some applications and try again, and check `%LocalAppData%\Blabb\logs\` if you want to
confirm it.

## Next

- [What affects speed](/docs/performance/what-affects-speed/)
- [Lower-spec machines](/docs/performance/lower-spec-machines/)