How to speak so Blabb hears you

Speak normally. The habits that feel like they should help — slowing down, over-enunciating — are the ones making it worse.

UPDATED 2026-07-28

Most people, the first time they dictate, slow right down and separate every syllable. It is a reasonable instinct and it is the wrong one. The speech models are trained on ordinary people talking at ordinary speed, so the further you get from that, the less familiar you sound to them.

Speak at your normal speed#

Dictate the way you would say the sentence to a colleague standing next to you. Normal pace, normal rhythm, normal contractions — "don't" rather than "do not", if that is what you would say.

Over-enunciating hurts in a specific way. Real speech runs words together; the model expects that and uses it. Pronouncing each word in isolation, like a 1950s telegram, produces something the model has never heard much of, and accuracy drops. Speaking unnaturally slowly does the same thing for the same reason.

The other half of this is phrasing. Say the whole sentence as a sentence, with the emphasis you would normally give it, rather than assembling it word by word. Blabb has more to work with when the shape of the phrase is intact.

You do not have to say the punctuation#

You do not need to dictate it. Say the sentence and move on rather than working "comma" and "full stop" into everything. If you are on Original and the punctuation is not coming out how you want it, a cleanup style is the thing that tidies it — Original runs no cleanup model, so nothing is adjusting punctuation from the sense of what you said.

You can still say a mark when you want a specific one — "question mark", "open quote", "at sign" — and it comes out as the mark rather than the words, provided you clearly meant it as an instruction. That is worth knowing for the symbols that are awkward to get any other way. The full list is in Numbers, symbols and addresses.

Layout is different, because Blabb cannot guess it: "new line", "new paragraph", "bullet point" and "next point" have to be spoken. See Spoken layout commands.

Correct yourself out loud#

If you fumble a sentence, do not stop the recording. Say "scratch that" — or "I mean", or "make that" — and carry straight on with the right version. The words you took back are dropped, and so is the phrase you used to take them back. See Correcting yourself as you speak.

Pauses are fine#

You can stop and think. Blabb ends a segment at real silence, not on a timer, and there is no maximum length — a pause in the middle of a sentence while you work out how to finish it does not cut you off, and a long dictation is not truncated part-way through.

What ends a segment is a genuine stretch of silence. In practice that means the natural gap when you have finished a thought, not the quarter-second hesitation in the middle of one.

Note

Blabb keeps 400 ms of audio from just before it detects you speaking, so a word that starts a fraction early still makes it in. It is a safety margin, not a licence to start talking before you press the key — a whole word spoken before the shortcut is gone.

The three habits that cost words#

Trailing off. The last two or three words of a sentence, said quieter as you lose interest in it, are the ones most likely to come out wrong. Keep your volume up to the full stop.

Mumbling the final word. Same problem, tighter focus. If you consistently lose one word per dictation and it is always the last one, this is why.

Starting before you press. Press the shortcut, then speak. The gap does not have to be long, but "press and talk simultaneously" reliably clips the first word.

When it is not you#

If you are speaking normally, at a sensible distance, and one particular word still comes out wrong every time, that is a different problem and it has a different fix — a name Blabb cannot know, a specialist term, or a homophone. Work through When a specific word keeps coming out wrong.