Measured, not promised.
Every number below comes from our own test runs on real recordings, blind-judged where quality is concerned. Hover any bar for details.
Time from key release to text
OpenFlow: 30 real recordings replayed in real time, Apple M4 Pro. Wispr Flow: ~520 ms reported in a public same-clip comparison. If you release the key mid-word, OpenFlow's slowest 10% take up to 142 ms — still 3.7× faster.
Which output reads better?
Accented English (EdAcc, 13 accents)
OpenFlow's phone speech model is our own: we distilled NVIDIA's 0.6B recognizer into its 110M sibling on 155 hours of public audio. Better than the stock model on 12 of 13 accents, at the same speed.
Where the speed comes from
Same idea. Your device, your rules.
Wispr Flow is excellent, and it runs in the cloud. OpenFlow does the whole job on your own machine.
| OpenFlow | Wispr Flow | |
|---|---|---|
| Open source | ✓ At launch | ✕ |
| Where your voice is processed | ✓ On your device | Wispr's servers |
| Works offline | ✓ | ✕ |
| Key release → text (median) | 2 ms | ~520 ms (reported) |
| 🍎 macOS · 🪟 Windows · 🤖 Android | ✓ | ✓ |
| 🐧 Linux | ✓ | ✕ |
| 📱 iOS | Coming soon | ✓ |
| Your names and jargon | ✓ Boosted while it listens; imports your Wispr dictionary | ✓ Dictionary |
| Price | Free, no word limits | Free with weekly limits · Pro $12/user/mo |
Wispr Flow details from wisprflow.ai/pricing and Wispr's privacy docs, November 2026.
Works where you work.
One engine, tested on every platform with a real recording — not a mock.
macOS Ready
Native menu-bar app. Hold fn, talk, let go. Live preview while you speak.
Windows Ready
Hotkey, audio and paste in a small Rust app. A real dictation runs on every code change.
Linux Ready
Same engine and tests. Runs on NVIDIA (CUDA), Vulkan or plain CPU.
Android Ready
A voice keyboard: hold the bar, speak, release. Uses our own 110M speech model.
iOS Soon
Keyboard plus companion app (iOS keyboards can't use the mic, so the app records and hands the text back).
Optional cloud Tested
For devices too slow to rewrite locally. Sends text only — never audio — with automatic local fallback.
It finishes while you're still talking.
That's the whole trick behind 2 ms: by the time you let go, the text is usually already written.
1. Hold and speak
Audio streams in 100 ms pieces while you hold the key.
2. Hear
NVIDIA Parakeet recognizes speech on-device, with your own words boosted as it listens.
3. Clean up
A small model fine-tuned for dictation removes ums and false starts, adds punctuation, and matches the app's style.
4. Land
It rewrites every half second as you talk, so the final text is ready the moment you let go.
Fully open. Launching soon.
Everything you need to run it, change it, or build your own.
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.
★ Star on GitHub · coming soon- ⚙️The engineRust core for Windows, Linux and phones; native Apple-GPU engine for Mac.
- 📦The appsmacOS, Windows, Linux and Android, with iOS on the way.
- 🧠Public modelsOur distilled 110M speech model and three rewriters (1.7B, 1.2B, 350M), trained only on public data.
- 📊BenchmarksThe test sets, the blind-judge harness and the latency suite behind every chart.
- 🎓A 7-part build guideHow it was built, step by step, including what didn't work.
Where we're not there yet.
Long recordings
On long real recordings, Wispr Flow's cloud recognizer still hears better: it wins end to end in our blind tests. Most of its wins come down to misheard words. Closing that gap is our next milestone.
English only
OpenFlow is English-only today; Wispr Flow supports 100+ languages.
Accuracy numbers
The quality results use a model personalized on the author's own dictations, which stays private. The public model, trained only on public data, scores lower on the same transcripts.
Your voice. Your device. 2 ms.
OpenFlow is in preview. The open-source release is coming soon.