AI dictation on a Mac that cleans up your words without the cloud
By Ryleigh Newman · Published · Updated
AI dictation on a Mac can be free and offline. YapToText, the free app I built, hears you with Whisper and tidies your words with a small language model, Phi-3.5 Mini, and both run on your Mac. Nothing you say goes to a server, there is no account, and nothing is locked behind a subscription.
The AI part is the cleanup: dropping ums, fixing grammar and punctuation, shaping an email or a note when you want one. It is also the part people worry about, and fairly, because a language model can change what you meant. So this page is mostly about what the cleanup does, what it is told never to do, and how to turn it down or off.
Download for free. No account, nothing locked, macOS 14 or later on Apple silicon. Source on GitHub.
On this page: 9 sections
What the AI does to your words
Two models run one after the other, and the AI Models page names them by their job. The dictation model, Whisper Large v3 Turbo (Q5), turns your voice into words. The cleanup model, Phi-3.5 Mini Instruct, turns those words into finished text when the mode asks for it. Whisper does the hearing; the cleanup model fixes what it misheard and tidies the rest. Each mode gives it different instructions:
- Raw Transcription. No AI at all: Whisper's words, with basic punctuation, exactly as heard.
- Clean Up. Removes filler words and false starts, fixes grammar, capitalization and punctuation, turns spoken web addresses and email addresses into proper form, and writes spoken numbers as digits without adding units you never said.
- Note. Short headings and bullets, with action items pulled into a list, and nothing added that you did not say.
- Email. An email body, never a Subject line, with a greeting only if you named someone and your name as the sign-off if you gave the app one.
- Message. A light cleanup for chat that keeps the casual tone.
- Code Comment. Short technical phrasing for comments and commit messages, with identifiers kept exactly as spoken.
Every mode is editable, built-ins included: its instructions are plain language, and you can rewrite them or make a mode of your own. The modes guide has the details.
What the AI is told never to do
A cleanup that rewrites you is worse than no cleanup. My hands make typing difficult, so fixing a rewrite by hand is the last thing I want, and the rules lean conservative:
- Your words stay yours. Clean Up keeps your wording, tone and register, blunt or profane language included, and is told never to delete, shorten or summarize a sentence.
- No made-up content. Note and Email are told to add nothing you did not say, and Email is told never to write a placeholder like [Your Name].
- No translation. Clean Up is told to keep the language you spoke and not to translate it.
- No lists on a whim. Auto mode never turns speech into a list on its own; saying "as a bullet list" does.
- No answers. If a dictation sounds like a question, the cleanup is told to tidy it, not to answer it.
A small model on a laptop does not always follow instructions, so the app checks its work. If the cleaned text reads like an assistant's reply, grows well past what you said, echoes reference text or invents an ellipsis, the app throws it away and types the raw transcript instead. The 1.5.2 release notes add one more line: "Cleanup no longer drops words you said." And if a cleanup ever goes too far, History keeps the raw transcript next to the cleaned one, and Copy Raw Transcript gives you exactly what Whisper heard. When the cleanup misbehaves, you get your raw words, not its guess.
Let Auto mode pick the cleanup
Auto mode is on by default, and it decides how much cleanup each dictation gets, cheapest clue first: a spoken instruction at the end, then the shape of the text, then the app you are dictating into, and only if it is still unclear, a one-word check by the cleanup model.
- A letter. A greeting and a sign-off make it an email.
- A quick chat line. Short or casual chat stays exactly as spoken.
- Everything else. A conservative pass that fixes what was misheard and rewrites nothing.
The app counts too: Mail and Outlook lean Email, Messages and Slack lean Message, and code editors lean Code Comment. The floating panel says what Auto detected, so you are never guessing, and while the cleanup model works its status line might read "Polishing the yap", the most honest job title it will ever have.
Say: At the very end of any dictation while Auto mode is on.
make that formalWhat happens: The phrase is not typed. Auto mode rewrites the dictation in a formal, professional tone. Friendly, casual, concise, shorter, as a bullet list, as an email, as a note and as a message work the same way.
Say what you want at the end, and Auto mode does it without a menu. The Auto mode guide has every rule.
Turn the AI down or off
If you would rather have your words exactly as spoken, climb this ladder, smallest change first:
- Look before it lands. Turn on Review before inserting in Settings, Advanced. Finished text opens in a floating editor first: Command Return inserts it, and Esc throws it away. Out of the box it stops only for dictations of 200 characters or more.
- Skip the AI for one dictation. Press Raw Transcription's number key while you talk, which is 2 in the default list with Auto mode on. The pick lasts for that one dictation.
- Drop the AI from one mode. In any mode's editor, switch off Rewrite the transcript with AI.
- Turn Auto mode off. Switch it off on the Home page. Out of the box, the selected mode is then Raw Transcription, which types exactly what Whisper heard.
- Switch off the whole AI stage. Turn off Post-transcription analysis on the Home page: no modes, no Auto mode, no cleanup, just the fastest raw transcription.
None of these delete anything, and History keeps the raw transcript either way. Every rung is a switch, so you can climb back up the same way.
Free and offline, and what that costs you
Free here means free: no paid tier, no word cap and no account. Offline means the models live on your Mac. The App Store build ships both of them, which is most of its 2.9 GB download, so nothing downloads at all unless you pick another model. The only network traffic the app starts is a model download you click, from Hugging Face, carrying nothing about you.
The cost is your Mac's own effort. The cleanup model is the heaviest thing the app runs, and the app shows it as 2.3 GB. By default the models stay loaded for 2 minutes after a dictation, and the next dictation after they unload takes a few extra seconds. The Energy page changes that, and Switch models with the power source can run lighter models on battery. No cloud means no bill, and it also means your Mac does all the work.
The Homebrew build is the exception: it ships without the models, so it cannot dictate until you download a speech model on the AI Models page, and Auto mode and the other AI modes need the cleanup model as well.
How AI dictation apps handle the cloud
Checked October 3, 2026, on each company's own pages. Plans and models change, so check before you choose.
- Wispr Flow. Its data controls page says transcription always occurs in the cloud. The Wispr Flow alternative page compares the rest.
- superwhisper. Its documentation lists cloud language models that run on hosted servers and one local model, S1-mini, which it says is English only and works only in its Message and Mail modes. Its table marks every language model, local or cloud, as Pro. The superwhisper alternative page has more.
- Apple. Apple Dictation types what it hears and adds punctuation in supported languages. On a Mac with Apple Intelligence, Writing Tools can proofread or rewrite text after you select it.
- YapToText. Whisper and the cleanup model both run on your Mac, for free. On macOS 26 with Apple Intelligence available, you can choose it for the cleanup job instead; it is optional and not the default.
Ask every AI dictation app the same thing: where do my words go to be cleaned up. The free dictation app roundup answers it for more apps.
Swap the cleanup model
Phi-3.5 Mini Instruct is the default, and the app's own catalog says it follows rewrite instructions far more reliably than Apple Intelligence. The Cleanup Model Library on the AI Models page has five alternatives, each downloaded once and run on your Mac:
| Cleanup model | Size the app shows | License |
|---|---|---|
| Phi-3.5 Mini Instruct, bundled and the default | 2.3 GB | MIT |
| Llama 3.2 1B Instruct | 808 MB | Llama 3.2 Community |
| Qwen 2.5 1.5B Instruct | 990 MB | Apache 2.0 |
| Gemma 2 2B | 1.7 GB | Gemma |
| Qwen 2.5 3B Instruct | 1.9 GB | Qwen Research |
| Llama 3.2 3B Instruct | 2.0 GB | Llama 3.2 Community |
The catalog suggests Qwen 2.5 3B Instruct especially for non-English dictation. You can also add any llama.cpp GGUF file under Your own models. Quick Edit, which rewrites text you select by voice, uses the same cleanup model unless you pick another one for edits. The AI models guide covers the library, and the Quick Edit guide covers editing by voice.
The short version
- Free and offline. Whisper and Phi-3.5 Mini both run on your Mac, with no account and no subscription.
- Conservative by default. Auto mode fixes what was misheard and rewrites nothing unless you ask.
- Easy to turn off. Raw Transcription, or Post-transcription analysis switched off, gives you exactly what Whisper heard.
For what Apple's own models add to dictation, see Apple Intelligence and dictation on a Mac.
The AI tidies; it does not get to decide what you meant. If a cleanup ever changes your meaning, open an issue on GitHub with the raw transcript from History. Every one gets read.
Download for free. No account, nothing locked, macOS 14 or later on Apple silicon. Source on GitHub.
Related
Sources
- https://wisprflow.ai/data-controls (checked 2026-10-03)
- https://superwhisper.com/docs/models/language (checked 2026-10-03)
- https://support.apple.com/guide/mac-help/writing-tools-write-improve-summarize-mchldcd6c260/mac (checked 2026-10-03)
- https://support.apple.com/guide/mac-help/use-dictation-mh40584/mac (checked 2026-10-03)
- https://github.com/ryleighnewman/YapToText (checked 2026-10-03)
- https://apps.apple.com/us/app/yaptotext/id6786382289 (checked 2026-10-03)
Questions
Is there a free AI dictation app for Mac?
Yes. YapToText is free on the Mac App Store with no paid tier: Whisper turns your voice into words and Phi-3.5 Mini cleans them up, both on your Mac, with no account.
Can AI dictation work offline on a Mac?
Yes, when the models run on the Mac. YapToText's App Store build ships both models, so dictation and cleanup work with no internet connection. The only network use is a model download you start yourself.
Does AI dictation send my voice to the cloud?
It depends on the app. Wispr Flow's data controls page says its transcription always occurs in the cloud. YapToText runs Whisper and the cleanup model on your Mac, so your voice never leaves it.
Will AI dictation change my words?
Only as much as the mode asks. Clean Up is told to keep your wording, tone and register and never to delete, shorten or summarize a sentence, and Raw Transcription uses no AI at all. When a cleanup looks wrong, the app types the raw transcript instead.
What AI model does YapToText use?
Two, both on your Mac: Whisper Large v3 Turbo (Q5) for speech, 574 MB, and Phi-3.5 Mini Instruct for cleanup, shown as 2.3 GB. Both ship in the App Store build, and the AI Models page can swap either one.
What is AI dictation software?
Dictation software that uses AI twice: a speech model turns your voice into words, and a language model cleans up what you said. In YapToText the speech model is Whisper Large v3 Turbo and the cleanup model is Phi-3.5 Mini, and both run on your Mac.