Fillers disappear
um, uh, you know, like: gone, in every language.
Saidso um I think we should uh ship the fix today
WrittenI think we should ship the fix today.
Free · v1.0.0-rc8Free for Windows 11 · v1.0.0-rc7Free · Mac v1.0.0-rc8 · Windows v1.0.0-rc7
Hold fn, talk the way you think, let go. VoxaFlow turns your words into finished text right at your cursor, in any app. Fillers gone, corrections applied, punctuation set. With its own models, trained for dictation, entirely on your Mac.
Hold the right Ctrl key, talk the way you think, let go. VoxaFlow turns your words into finished text right at your cursor, in any app. Fillers gone, corrections applied, punctuation set. All on your PC.
Hold the dictation key (fn on a Mac, right Ctrl on Windows), talk the way you think, let go. VoxaFlow turns your words into finished text right at your cursor, in any app. Fillers gone, corrections applied, punctuation set. All on your computer.
To: Sarah Chen
Subject: Our meeting
hiSarahumcanwemovethemeetingtoTuesdayactuallyno,Wednesdaymorningthanks
Its own AI models
0.6s
until the text is there, short dictation on an M4 Mac
VoxaFlow Lektor is trained for one job only: turning speech into clean text, without rephrasing and without making anything up.
On Windows 0.8–1.8 s, measured on a notebook with an AMD Ryzen AI 7, processor only, plugged in.
Meet VoxaFlow Lektor →
Packs for your field
3packs
in English and German
Coding, Medical, Legal & Tax: terms, spellings and formats that plain dictation doesn’t know.
See the packs →
No cloud
0bytes
of your voice go online
Your voice and your text never leave your computer. No account, no subscription, no server listening in.
What VoxaFlow sends →
Measured on November 3, 2026 with VoxaFlow 1.0.0-rc4 on a Mac with M4 and 32 GB: median of short dictations (3–10 s of audio), from the end of the recording to the finished text. Speech recognition by Parakeet TDT v3, 25 European languages.
Beta · free
Dictate in any app with the VoxaFlow keyboard, or press the Action button and speak. The same clean text as on your Mac, processed on your iPhone.
Pro and all packs are free during the beta. How the beta works →
Recording starts right away, in any app, even on the lock screen. Or triple-tap the back of your iPhone.
Pause, correct yourself, switch languages. VoxaFlow listens in the Dynamic Island.
Press again. The clean text is saved as a note in VoxaFlow and, when your iPhone is unlocked, put on the clipboard.
Needs iOS 26, best on iPhone 13 or newer. The Action button is on iPhone 15 Pro and Pro Max, iPhone 16e and the iPhone 16 and 17 series; on other iPhones you dictate with the keyboard.
How it works
01
Put the cursor where the text should go: a prompt in Cursor, a reply in Slack, an email, a note. Hold the fn key.Put the cursor where the text should go: a prompt in Cursor, a reply in Teams, an email, a note. Hold the right Ctrl key.
02
Pause, restart, say “um”, change your mind halfway. Long dictations are processed while you are still speaking.
03
Clean text appears at the cursor about a second later. Punctuated, capitalized, with only what you meant.Clean text appears at the cursor moments later. Punctuated, capitalized, with only what you meant.
VoxaFlow Lektor
Speech recognition alone gives you a transcript. VoxaFlow Lektor, our own language model, turns it into the text you meant. It is trained for dictation only, runs on your computer and keeps your wording.
um, uh, you know, like: gone, in every language.
Saidso um I think we should uh ship the fix today
WrittenI think we should ship the fix today.
Say the wrong thing, correct yourself, keep going. Only the correction is written.
Saidsend the draft to Mark sorry no to Mike by Friday
WrittenSend the draft to Mike by Friday.
Think mid-sentence as long as you like. A pause is not a full stop.
SaidI’ll take care of the … … invoices tomorrow morning
WrittenI’ll take care of the invoices tomorrow morning.
Names, products and jargon from your dictionary come out right every time.
Saidask yannick about the voxa flow release and the q four numbers
WrittenAsk Janik about the VoxaFlow release and the Q4 numbers.
In our tests VoxaFlow Lektor did not make up a single name, number or address.
Change your mind halfway through a list, and only the corrected list is written.
Saidwe need milk, eggs, actually no, bread and butter
WrittenWe need milk, bread and butter.
Phone numbers spoken one digit at a time come out as figures, grouped the way you said them.
Saidmeine Handynummer ist null eins sieben sechs, zwei drei vier fünf sechs sieben acht
WrittenMeine Handynummer ist 0176 2345678.
A German dictation, as in our test.
Most accurate
For Macs with 16 GB of memory or more.
Short dictation, until the text is there
Faster
For Macs with 8 GB. As accurate for everyday dictation, slightly weaker with medical and legal terms.
Short dictation, until the text is there
VoxaFlow picks the right model by itself, on the Mac and on Windows since version 1.0.0-rc4. On Windows the choice depends on memory and processor. On a notebook with an AMD Ryzen AI 7, processor only, a short dictation is done after 0.8–1.8 s with VoxaFlow Lektor and after 0.5–1.0 s with Lektor Kompakt.
Never trained on your dictations.What you say never leaves your computer, so it cannot end up in any training.
Packs · free until January 31, 2027
Plain dictation doesn’t know dosage schedules, case citations or the identifiers in your repository. The packs add word lists, writing rules and formats for your work, in English and German.
Coding pack
Talking is how you explain a bug to a colleague. Now it is how you explain it to Claude Code or Cursor. The Coding pack learns your repository and writes identifiers, file names and commands exactly.
SettingsView, user_id, MAX_RETRIESMedical pack
metoprolol fifty milligrams twice daily blood pressure one forty over ninety
Metoprolol 50 mg twice daily, blood pressure 140/90 mmHg.
Legal & Tax pack
the claim arises under forty two u s c section nineteen eighty three
The claim arises under 42 U.S.C. § 1983.
Only with Pro, one at a time or combined. Each pack €40, at launch €30. Practice and office license: 5 seats for the price of 4. All prices →
Works where you type
VoxaFlow types like a keyboard, so there is nothing to install per app and nothing to copy and paste. If you can click into it, you can dictate into it.
No cloud
Speech recognition and the AI cleanup both run on your Mac’s Neural Engine and GPU. Turn off Wi-Fi and VoxaFlow keeps working. This is the complete list of what it ever sends, all to our own servers in Germany.
| Cloud dictation service | VoxaFlow | |
|---|---|---|
| Where your voice is processed | on the provider’s servers | on your computer |
| Internet needed | for every dictation | once, to download the models |
| Your text | with the provider, at least while it is processed | stays with you |
| Account | required | not needed |
| Cost | usually a monthly subscription | buy once, updates included |
| Service down or prices up | dictation stops | VoxaFlow keeps working |
Applies to services that process speech in the cloud.
Never sent
Never sent. Not once, not anonymized, not “to improve the service”.
Every connection VoxaFlow makes3
Model download
Once, after installing
Loads the speech and language models (about 2.5 GB). Nothing is sent except the request for the files.
Update check
Once a day, can be turned off
Asks whether a new version exists. No system profile, no identifiers. We count these requests anonymously to see how many installations are active.
License check
From version 0.8: to register the free period or trial, then every 30 days
A hashed device ID, the app version and, if you bought one, your license key. No name, no email, no usage data.
Tested, not just promised: an automated network test runs the app, logs every request and fails as soon as any server other than our own appears.
Data protection under the GDPR
Pricing
One license, two devices, Mac or Windows. Pro for everyone who writes a lot, plus packs for code, medicine and law & tax. Pay once, no subscription, updates included.
Right now everything is free. Until January 31, 2027 every feature that is available today is free. Sales start on February 1, 2027, and everyone who uses VoxaFlow before the sales start gets the launch price, guaranteed.
Full quality, for everyday notes.
€0forever
Recommended
For people who write all day.
€39€49once
On sale from February 1, 2027
Launch price for the first 500 buyers
Added to Pro, one at a time or combined.
+€30€40per pack
On sale from February 1, 2027
Launch price for the first 500 buyers
Free until January 31, 2027. No account, no credit card, no email.
At a glance
FAQ
Something missing? Write to hello@voxaflow.ai. A human reads it.
VoxaFlow Lektor is our own language model for cleaning up dictations: it removes fillers, applies your corrections, sets punctuation and keeps names, numbers and addresses as you said them. It is built on Qwen3 (Apache 2.0) and trained for dictation only. There are two sizes: VoxaFlow Lektor for Macs with 16 GB of memory or more, and VoxaFlow Lektor Kompakt, which takes about half as long, for Macs with 8 GB. Both run entirely on your computer. Both are available on Windows too; which one VoxaFlow uses there depends on memory and processor.
No. Your voice and your text never leave your computer, so they cannot end up in any training. VoxaFlow Lektor was trained on synthetic examples only.
Yes. The Medical and Legal & Tax packs write medical terms, dosages, sections of statutes, case numbers and amounts correctly, set headings by voice and protect dictations with a professional secrecy mode without history or sync, which you can turn off if you are not bound by professional secrecy. VoxaFlow also types into Remote Desktop and Citrix sessions. Both packs work in German and in English (US and UK): VoxaFlow detects the language of each dictation and uses the matching word lists and writing rules. Guides at /docs/medical and /docs/legal.
Yes, dictation works offline. Speech recognition (Parakeet TDT v3) and the AI cleanup (VoxaFlow Lektor) run entirely on your Mac. You need an internet connection once, to download the models (about 2.5 GB). The daily update check is optional and can be turned off. The app also checks its license status online from time to time and, if you use a pack, downloads its word lists; your voice and text are never sent.
VoxaFlow needs a Mac with Apple silicon (M1 or newer) and macOS 14 Sonoma or later. Macs with 8 GB of memory automatically use VoxaFlow Lektor Kompakt, the smaller and faster cleanup model; with 16 GB or more VoxaFlow uses VoxaFlow Lektor.
VoxaFlow recognizes 25 European languages, including English, German, French, Spanish, Italian, Dutch and Polish. It detects the language on its own, so you can switch between languages without changing a setting.
macOS dictation writes down what you say, including fillers, false starts and corrections. VoxaFlow adds an on-device language model that removes fillers, applies your corrections, sets punctuation and keeps your vocabulary, so the result is ready to send. It also handles long dictations and has a command mode for rewriting selected text.
Wispr Flow processes your speech in the cloud and is sold as a subscription. VoxaFlow processes everything on your Mac, works offline, needs no account and is a one-time purchase.
Yes, as a free beta that you install through Apple’s TestFlight app: voxaflow.ai/iphone. With the VoxaFlow keyboard you dictate straight into any app. With the Action button or a triple tap on the back of the iPhone, the clean text is saved as a note in VoxaFlow and, when your iPhone is unlocked, put on the clipboard too. It needs iOS 26; the Action button is on iPhone 15 Pro and newer. Pro and all packs are free during the beta.
Yes, for Windows 11 on PCs with an Intel or AMD processor, free on the download page. The computer check at voxaflow.ai/check shows whether your PC qualifies. Version 1.0 follows in January 2027. On Windows too, speech recognition and cleanup run on the computer. One license covers two devices, Mac or Windows.
Yes. VoxaFlow types into any app with a text cursor, including code editors, terminals and AI coding agents. The Coding pack, free until January 31, 2027, adds exact identifiers and file names from your repository, formats commands in the terminal and turns instructions spoken in German into English prompts.
Until January 31, 2027, VoxaFlow is free with every feature. From February 1, 2027 there is Free (2,500 words per week) and Pro for €49 as a one-time purchase, plus the Coding, Medical and Legal & Tax packs for €40 each. At launch Pro costs €39 and each pack €30; everyone who uses VoxaFlow before the sales start is guaranteed the launch price. All prices are on the pricing page.
Until then everything is free. After that, VoxaFlow continues as Free unless you buy Pro. Your dictionary, settings and history stay on your Mac.
Audio is kept in memory only while you dictate and is discarded after the text is written. Nothing is uploaded. The only connections VoxaFlow makes are the model download, the daily update check, the license check and, if you use a pack, the download of its word lists. The license check sends only a hashed device ID, the platform, the app version and, if you bought one, your license key.

Install, allow access once, start talking. Free until January 31, 2027, with every feature available today.
Try the iPhone beta for freeDownload for MacFree · v1.0.0-rc8 · DMGMac: macOS 14 or later · Apple silicon (M1 or newer) · about 3 GB free space. Windows: Windows 11 · 64-bit · about 4 GB free space · Computer check ›