Measured, not promised.
Every number comes from our own test runs on real recordings. Quality is judged blind, in both orders. Hover any bar for the details. Public test sets, Canto and Voice Arena comparisons →
Key release to final text
OpenFlow 1.4: 30 real recordings replayed in real time on an M4 Pro, two runs (55 and 53 ms). Dictations up to 20 seconds get a second, whole-recording pass for accuracy, which is most of the 54 ms. Slowest 10%: 170 ms or more. Wispr Flow: ~520 ms, measured by Elliot Arledge on his own clips, not by us.
Which output reads better?
Accented English, 13 accents
Research result, not shipped yet: our 110M speech model, distilled from NVIDIA's 0.6B recognizer on 155 hours of public audio. Better than the stock model on 12 of 13 accents at the same speed, but it wrote some numbers as unknown tokens, so phones use the stock 110M for now.
Where the speed comes from
Detailed results
Same idea. Your device.
Wispr Flow is excellent, and it runs in the cloud. OpenFlow does the whole job on your own machine, and the code will be yours to read.
| OpenFlow | Wispr Flow | |
|---|---|---|
| Open source | At launch | No |
| Where your voice is processed | On your device | Wispr's servers |
| Works offline | Yes | No |
| Key release to text (median) | 54 ms | ~520 ms (reported) |
| macOS and Windows | Yes | Yes |
| Linux | With the open-source release | No |
| iPhone and Android | Android beta today · iPhone in testing | Yes |
| Your names and jargon | Reads names on your screen, on the device; boosted while it listens; imports a Wispr dictionary | Dictionary and context awareness |
| Price | Free, no word limits | Free with weekly limits · Pro $12/user/mo |
Wispr Flow details from wisprflow.ai/pricing and Wispr's privacy documentation, November 2026.
Works where you work.
One engine, tested on every platform with a real recording — not a mock.
macOS
Menu-bar app. Hold fn, talk, let go. Apple silicon or Intel, macOS 12 and later.
Windows
Hold Right Ctrl, talk, let go. Windows 10 and 11, 64-bit. Nothing else to install.
Linux
Same engine and tests; runs on NVIDIA (CUDA), Vulkan or the CPU. Ships with the open-source release.
Android
A voice keyboard running NVIDIA Parakeet 110M and our clean-up model on the phone. iPhone is in testing.
It finishes while you're still talking.
That's the trick behind the speed: by the time you let go, almost all of the text is already written.
Hold and speak
Audio streams in 100 ms pieces while you hold the key.
Hear
NVIDIA Parakeet recognizes speech on-device, with your own words boosted as it listens.
Clean up
A small model fine-tuned for dictation removes ums and false starts, adds punctuation, and matches the app.
Land
It rewrites every half second as you talk, so the text is ready the moment you let go.
Fully open. Launching soon.
We're opening the code, the apps, the public models and the benchmark suite, so every number on this page can be checked on your own machine. The license will be announced with the release.
GitHub · coming soon- The engineRust core for Windows, Linux, Intel Macs and phones; native Apple-GPU engine for Apple silicon.
- The appsmacOS and Windows today; Linux and Android with the release; iOS on the way.
- Public modelsThree rewriters (1.7B, 1.2B, 350M) trained only on public data, plus our distilled 110M speech model (research preview).
- BenchmarksThe test sets, the blind-judge harness and the latency suite behind every chart.
- A seven-part build guideHow it was built, step by step, including what didn't work.
Free. One click.
Download, open, hold a key and talk. Everything runs on your own computer or phone; no account needed.
Mac
- Open the downloaded file and drag OpenFlow into Applications.
- Open OpenFlow. The first time, macOS says “Apple could not verify ‘OpenFlow’ is free of malware”. That’s because OpenFlow isn’t notarized by Apple yet (our developer account is waiting for Apple’s approval). Click Done, then open System Settings › Privacy & Security, scroll down and click Open Anyway. You only do this once.
- Allow the microphone, Accessibility and Input Monitoring when asked. Then hold fn, talk, and let go.
Prefer Terminal? One line installs it, with no Open Anyway step.
curl -fsSL https://rajatdandekar--openflow.modal.run/install.sh | bashWindows
- Open the downloaded file, click Extract all, then open openflow.exe.
- If Windows shows a blue SmartScreen box, click More info › Run anyway (first time only).
- When it says OpenFlow is ready, click into any text box, hold Right Ctrl, talk, and let go.
Prefer PowerShell? One line installs it with a Start-menu shortcut.
irm https://rajatdandekar--openflow.modal.run/install.ps1 | iexAndroid beta
- Tap the button on your phone and install the app (allow installs from your browser when Android asks).
- Open OpenFlow and follow the four steps: microphone, models, turn on the keyboard, switch to it.
- In any app, switch to the OpenFlow keyboard, hold Hold to talk, speak, and let go.
iPhone soon
- Testers get an email invite to TestFlight, Apple's beta app.
- Install TestFlight, tap the invite, and install OpenFlow.
- Turn on the OpenFlow keyboard in Settings and talk in any app.
Click, hold, talk. It types where your cursor is.
Works in any app with a text box: Slack, Gmail, Notion, ChatGPT, Notes, Word, your code editor.
Hold fn and talk
Click into any text box, hold fn (🌐) on a Mac or Right Ctrl on Windows, speak naturally, let go. The Vizuara flower turns while you talk.
Hands-free
Tap fn twice quickly to keep listening without holding; tap fn once to stop. Good for long emails and notes.
Names spelled right
Replying to someone? OpenFlow reads the names on your screen (on your Mac only, never saved) and spells them correctly. Turn it off in the ◉ menu.
First launch on a Mac
A setup window walks you through it: allow the microphone and switch OpenFlow on in Accessibility. After an update, if it's already on, switch it off and on once.
Something not working? Click ◉ in the menu bar: it lists anything that needs fixing with a one-click button. If text didn't appear, it's on your clipboard: press ⌘V. Restart engine in the same menu clears anything stuck. On Android, switch to the OpenFlow keyboard, hold the bar to talk, or type with the keys. Everything runs on your device; your voice never leaves it.
Built on the shoulders of phonon.
OpenFlow started from Elliot Arledge's open-source phonon. Here's what it showed, and what we added.
Elliot Arledge's phonon showed that NVIDIA Parakeet plus a small fine-tuned Liquid AI rewriter can beat Wispr Flow on a Mac — he measured 65 ms for phonon against about 520 ms for Wispr Flow on the same clips.
We rebuilt his recipe from scratch. In blind tests on our own dictations it matched or beat his released 1.7B model. Then we kept going. Thank you, Elliot.
- Finishes while you talk54 ms median from key release to text on an M4 Pro, including a whole-recording accuracy pass (phonon reported 65 ms on an M4 Max).
- Cleaner restarts and repeats"Tuesday, no sorry, Wednesday" becomes "Wednesday": +15% on a restart test set.
- Names from your screen, spelled rightNew in 1.4 on Mac, Windows and the Android keyboard: OpenFlow reads the names around where you dictate (on the device, never saved) and uses them while it listens and cleans up. On our benchmark, names right go from 74% to 93% when they are on screen.
- Beyond Apple siliconA portable Rust engine for Windows, Linux, Intel Macs and Android.
- Our own speech model for phones (research)Distilled from Parakeet 0.6B on public audio; better on 12 of 13 accents. It ships once it writes numbers reliably.
- Public models, open releaseModels trained only on public data; the full code is coming soon.
Where we're not there yet.
Long recordings
On long real recordings, Wispr Flow's cloud recognizer still hears better and wins end to end in our blind tests, mostly on misheard words. Closing that gap is our next milestone.
English only, for now
OpenFlow is English-only today; Wispr Flow supports over 100 languages. On older Intel Macs and PCs, the portable engine takes one to two seconds.
How we measured quality
The quality results use a model personalized on the author's own dictations, which stays private. The downloadable public model is trained only on public data and scores lower on the same transcripts.
Your voice. Your device. Two milliseconds.
OpenFlow is in preview. The open-source release is coming soon.
Download for free