How to Write Twice as Fast With Voice Typing and AI

The shortcut to writing faster with voice input is to separate two steps: dictating a rough draft by voice, and having AI clean it up. Spoken words don't become a manuscript on their own, so the approach is two-stage: generate volume with your voice, and let generative AI shape it into prose.

For freelancers who spend a large part of their week on blog posts, client reports and newsletters, the first draft is usually what eats the most time when everything is typed. Combining voice input with AI cleanup noticeably shortens the time to a finished first draft. Here is a workflow you can run day to day.

Why "just talking" doesn't produce a draft

The most common voice-typing complaint goes like this: "I tried it, but the text was unreadable, so I stopped using it." That's not a recognition accuracy problem. It's the difference between spoken and written language.

When you talk, filler words like "um" and "like," restatements of the same point, and sentences that change direction halfway through inevitably creep in. Modern speech recognition transcribes all of this faithfully, so the more accurate it is, the more unreadable the draft.

So the starting point is to stop treating voice output as a finished manuscript. What you create by voice is a pile of raw material that contains all your points and examples. Reshaping that material into prose by hand would eat up the time you saved by dictating, but reshaping is exactly what AI writing tools are good at.

Split the work so that you talk to create the material, AI produces the clean draft, and you finish by correcting facts and phrasing. Every step stays short.

The workflow: record, transcribe, clean up

The process has three stages: record, transcribe, and clean up.

First, before recording, jot down an outline of headings only. Talking without a plan leads to more restarts, so keep a short list such as "three things this article needs to say" in front of you. That's all the prep there is, but it makes a big difference to the quality of the material.

Next, talk through the outline. Many people find it easiest to speak into a phone while walking, but a desk works just as well. The key is not to re-record when you misspeak. Just keep going. The AI will tidy it up later, so polish at the recording stage doesn't matter.

A good target is 10 to 15 minutes of talking per article. People typically speak somewhere around 130 to 150 words per minute, so 10 minutes gives you roughly 1,300 to 1,500 words of material. It isn't usable as-is, but compared with a day of staring at a blank page for an hour and getting a few hundred words down, starting with material already in hand is a big advantage.

There are two ways to turn speech into text: real-time voice typing, or recording first and transcribing afterward. Here are the main options:

The practical way to choose is by where you'll be talking. At a desk, real-time typing is satisfying because you watch the text pile up on screen. On a walk or a commute, recording and transcribing later is more relaxed. For short notes, your phone's built-in dictation is plenty. For recordings of 20 or 30 minutes, a local Whisper-based transcriber is handy: it's open source, free, and processes the audio on your own computer without sending it anywhere.

Finally, hand the transcript to an AI writing tool for cleanup. A prompt like this works well:

Below is a transcript of me dictating a draft article.
Remove restatements and repetition, put the points in a logical order,
and turn it into a first draft of a blog post in a friendly, clear tone.
Do not add any information I didn't say.

The last line is the important one. Asked to clean up a draft, AI often tries to be helpful and adds generic points you never made. Saying "don't add anything" explicitly gets you a first draft made only of your own words.

A note on cost: if you run the cleanup through a pay-as-you-go API, text in some languages costs more to process than English. Many non-English languages tokenize less efficiently, so the same content turns into more tokens, the unit that API pricing is based on (osFoundry has a short explainer on tokenization). If you're cleaning up long dictations every day, it's worth keeping in mind.

What a human still needs to fix after the first draft

AI cleanup gives you a first draft, not a final one. There are three things worth always checking.

Proper nouns and numbers. Speech recognition tends to turn product and personal names into similar-sounding words, and the AI then keeps those errors looking plausible in the cleaned-up text. Check every proper noun in the draft against an original source such as the official website.

The opening. Dictated drafts tend to start with a long preamble, and cleanup doesn't remove that habit. Cutting the first two or three sentences so the piece starts with the main point is often all it takes.

Your own voice. Cleaned-up text can be so tidy that your personal style fades. Changing a few phrases back to the way you'd normally say them brings the piece closer to the tone your readers know.

Even with these three checks, a first draft usually comes together faster than typing from a blank page. Voice input helps most on days when you know what you want to say but your hands just won't move.

FAQ

Which is faster in the end, voice or keyboard?

For the drafting stage, voice is faster for most people, because most people speak faster than they type. Editing doesn't suit voice, though, so the realistic split is to draft by voice and revise at the keyboard.

Can AI clean up a transcript full of recognition errors?

It often fixes them by inferring from context. Proper nouns, however, tend to get tidied up while still wrong, so listing frequently used names in the prompt in advance improves the result.

Is it okay to dictate content involving a client's confidential information?

First, check the terms of the voice input service and AI tool you use to see where audio and text are sent. If the content shouldn't leave your hands, choose a method that runs entirely on your own device, such as a local Whisper-based transcriber.