A rough draft page next to a clean printed one

Questions

Can dictation produce clean writing? Yes, three ways

Say the punctuation yourself, fix it by hand, or let a writing model rewrite it. All three work; only one of them can put words in your mouth.

Last updated

Method 1: say the punctuation

"Full stop", "comma", "question mark", "new paragraph". They appear as you speak. It feels absurd for about a day and then becomes invisible.

This gives you exact control, which matters more than people expect. Your commas control your rhythm, and in anything contractual or instructional a comma changes the meaning. Do not delegate those to a model's guess.

Method 2: dictate loosely, fix by hand

Speak the whole thing, then spend ten seconds deleting the fillers and adding the full stops. For a chat message or a short reply this is genuinely the fastest route.

It stops scaling at about a paragraph. Editing four hundred words of unpunctuated speech is slower than having typed them, which is where people conclude dictation cannot do long-form work.

Method 3: let a writing model do it

A small language model takes the raw dictation and returns what you would have typed: fillers gone, punctuation placed, spoken self-corrections applied, a spoken list laid out as a list.

What you said

so um I think we should push the launch to to Monday because er legal haven't signed off yet and and I don't want to you know ship something we then have to pull

What lands on the page

I think we should push the launch to Monday. Legal haven't signed off yet, and I don't want to ship something we then have to pull.

The same mechanism can target a shape rather than just tidying — the same speech as bullets, as a formal email, or as notes.

The two questions to ask about any rewriting feature

  1. Where does it run? Many apps recognise speech locally and then send the text to a cloud model to tidy it. That is a second transfer, of the actual content of what you said, and it is often on by default. VV's writing model runs on your Mac, which is why it works with no connection.
  2. What happens when it is unsure? A model that rewrites can also invent — a figure you did not say, a commitment you did not make. The behaviour that matters is what it does when it cannot vouch for the result. VV validates the rewrite against what you actually said and pastes your raw words instead.
Transforms in VV Natural, Concise, Professional, Bullets and Email, chosen from the recording overlay. Local by default.

Which method for which task

TaskMethod
A chat messageLoose, fix by hand — or automatic Concise
An email to a colleagueAutomatic, then read it
An email to a clientAutomatic, then read it properly
Meeting notesAutomatic, targeting bullets
Anything contractual or instructionalSay the punctuation. Do not delegate a comma
A long first draftLoose and fast. Structure by keyboard afterwards
A prompt to an AILoose. It does not care about your commas
Anything where your voice is the productTurn rewriting off and edit by hand

Questions

Do I have to say the punctuation?

No. Modern models place it reasonably well and a rewriting step handles it too. Say it yourself when a comma changes the meaning — instructions, contracts, anything where precision is the point.

Can dictation produce bullet points?

Yes — either by saying "new line" between items inside a list the app has already started, or by dictating the list as a sentence and letting a rewrite lay it out. Count the items afterwards; dropping one is the usual failure.

Does the rewriting send my text to a server?

In VV, no — the writing model is local, which is why it works offline. In many apps that stage is a cloud API call. It is the step where a privacy-first setup most often leaks, so test it specifically.

Can I turn rewriting off?

Yes. Some people want their exact words and nothing else, particularly for notes to themselves and for anything where their own voice is the point.