Runs entirely on your machine
Local recognizers turn speech into text on-device: Parakeet for English, Whisper for ten more languages. No cloud, no API keys. Pull the Wi-Fi and it still works.
Local voice input for Mac & Windows
A local voice keyboard for your computer, set up by one installer. It listens, tidies up what you said, and drops the text right where your cursor is. Mac is the main platform, still an unsigned preview; Windows is an earlier, unverified preview.
Whisper Face types into the text field you are using on your Mac.
Not another transcript box. Every part is built to earn your trust with what you actually said.
Described as they work on a Mac with a default install: Parakeet and Whisper Tiny. The larger Whisper model and the cleanup model are optional downloads. Windows shares capture, recognition, cleanup, snippets and tones, but not the verified insertion transaction, the Outbox, correction learning, or the results view — see how it compares.
Local recognizers turn speech into text on-device: Parakeet for English, Whisper for ten more languages. No cloud, no API keys. Pull the Wi-Fi and it still works.
It leases the field under your cursor and checks it is still there before pasting once. If focus moved, the text parks safely in your Outbox instead of the wrong box.
Models stay warm and long sentences are recognized while you are still talking, so only the last stretch is left to decode when you let go of the key. We have not published a latency measurement yet.
Cleanup removes filler and follows spoken structure. One rule waits for you to switch it on: "Tuesday, actually Wednesday" then keeps only Wednesday. It is off by default because "I thought it was Tuesday, actually Wednesday" would lose Tuesday too. App names and your dictionary can steer recognition but never replace a word you said, and the optional cleanup model may only make edits it can prove.
A word you fix the same way twice becomes a small, local rule you can read and forget. It never maps one everyday word to another or flips a negation, and there is no cloud transcript following you around.
Faithful fallbacks, exact-once insertion, and a recoverable Outbox mean a bad moment never eats your sentence.
Turn on Flight Recorder and the microphone keeps the last twenty seconds in memory, never on disk. Say the thing, realise you wanted it, tap the key and keep it.
Hold a modifier with your key to dictate code with real punctuation or run an editing command. Compose, reply, and edit-the-selection need the optional cleanup model (about 3.4 GB) and English; without it they change nothing and tell you so.
Every result can explain itself — what it protected, which edits it accepted or refused, what it was unsure about. No transcript dossier required.
Five steps, all on your machine.
Press and hold your hotkey, then just start talking. No window to open, no button to find.
Warm local recognizers turn sound into text, starting while you are still speaking. Nothing leaves the Mac.
A small set of written rules removes filler and fixes formatting. The optional cleanup model may only make edits it can prove.
Whisper Face leases the field under your cursor and double-checks it is still the right one.
One clean paste. If the field moved, the text waits safe in your Outbox instead.
Choose who greets you up top. It sits quietly, then opens its mouth the moment you start talking. Hover to say hi — click to send it up to the top of the page.
No cloud calls. No telemetry. Listening, recognizing, cleaning up, and typing all happen on your Mac — even with the Wi-Fi off. Your dictation history stays in the install folder, so keep that folder out of iCloud Drive. When Advanced writes a support bundle, it stays private on your disk and contains no transcripts.
The app sets everything up and checks its own work. Then hold your key and talk.
Open the disk image and drag Whisper Face onto Applications. No account, no sign-up wall.
Whisper Face is not yet signed with an Apple Developer ID, so macOS blocks the first open. On macOS 15 or later, choose Done, then go to System Settings → Privacy & Security, choose Open Anyway for Whisper Face, and enter your password. On macOS 14, Control-click the app in Applications, choose Open, then confirm.
A setup window downloads about 1 GB: the speech model, a small preview model, and the parts they run on. If your Mac lacks Apple’s command line tools, it asks macOS to install them first. It ticks off each step, and if one fails it says why and offers Try Again.
When setup finishes, Whisper Face asks for them. In System Settings → Privacy & Security, turn on Whisper Face under Input Monitoring (so it notices you holding the key) and Accessibility (so it can type for you). macOS asks about the Microphone the first time you dictate.
This build is not signed or notarized yet, so macOS blocks it the first time. Step 2 shows the way through.
Apple Silicon, macOS 14 or newer, about 5 GB free. The app sets itself up and downloads the speech models and the parts they run on. Check the download against SHA256SUMS (Mac). Release notes · Source on GitHub
Unzip it into a folder of its own, such as C:\Users\you\Whisper Face. Your dictation history lives in that folder, so avoid Documents or Desktop if OneDrive syncs them. The installer will not run from inside the zip.
Double-click it — it starts the bundled setup script, which Windows blocks if you launch it directly. One repeatable flow sets up the app, locked dependencies, models, a login task, and health checks, and it is safe to run again later. It needs winget (App Installer, from the Microsoft Store) the first time.
The build is not code-signed, so Windows may say it protected your PC. Choose More info, then Run anyway — or check the zip against SHA256SUMS-windows first.
Windows uses Right Alt in place of Right Option, and Whisper Face lives in the tray rather than the menu bar.
Windows is an unverified preview: neither its installer nor a full dictation has been run on real Windows hardware, and it may not install. It is also not code-signed, so SmartScreen will warn you.
Windows 10 or 11, 64-bit, winget, about 8 GB free. The installer downloads the speech models, a local cleanup model, and tools. Check the download against SHA256SUMS-windows. Release notes · Source on GitHub
# the setup window, on a first open
Set up Whisper Face
✓ Getting ready
✓ Downloading the parts it runs on
✓ Downloading the speech model (565 MB)
✓ Building the speech engine
✓ Downloading the preview model (75 MB)
✓ Starting Whisper Face
Then hold Right Option and start talking → :: the steps setup runs (preview, unverified on hardware)
C:\Users\you\Whisper Face> Install.cmd
== installing uv, ffmpeg, Ollama
== downloading qwen3.5:4b (~3.4 GB)
== installing the locked Windows Python environment
== downloading Whisper Tiny and large-v3-turbo
== installing the Windows login task
Then hold Right Alt and start talking → Yes. Core dictation, privacy, accessibility, the installers, and local use are free and stay that way. No account, no word cap, no paywalled accuracy.
Never. There is no sign-up, no login, no license key, and no email field anywhere — not to download, not to install, not to use it. Nothing to create means nothing to leak, sell, or lock you out of.
No. Recognition runs on your Mac. You can turn off Wi-Fi and keep dictating. Nothing about what you say is uploaded. A local history of your dictations (at most 500, none older than 30 days) stays in the install folder.
Apple Silicon Macs on macOS 14 or newer. Setup needs Apple’s command line tools; if your Mac lacks them, the setup window asks macOS to install them before it downloads anything else.
The app you dragged to Applications, and in ~/Library/Application Support/Whisper Face a copy of its own source with a Python environment, the Parakeet and Whisper Tiny speech models (about 650 MB), a speech helper it builds with Apple’s command line tools, and a login service. No Homebrew. The larger Whisper model and the Ollama cleanup model are added only if you ask (setup.sh --models in that folder, about 5 GB). setup.sh --uninstall shows everything it would remove before removing anything.
Allow about 5 GB free on a default install and 8 GB with the optional models. The setup window downloads about 1 GB, plus Apple’s command line tools if your Mac lacks them. These are estimates: we have not yet measured a clean-machine install, how long it takes, or the memory it uses.
Whisper Face is not signed with an Apple Developer ID or notarized by Apple yet, so macOS blocks the first open. On macOS 15 or later, open the app once and choose Done, then choose Open Anyway in System Settings → Privacy & Security. On macOS 14, Control-click the app and choose Open. You can check the download against the release’s SHA256SUMS first.
Windows shares the core pipeline and has its own installer, but it is an unverified preview: neither the installer nor a full dictation has been run on real Windows hardware yet. A native iPhone product is a future workstream, not a setup promise today.
We have not measured it against anything yet. What it does by design: names, numbers, code, and commands are protected from the optional cleanup model, and app names, window titles, and your dictionary can steer recognition but never replace a word. Other text on screen can only repair a name the recognizer was unsure of (a visible "Qwen" fixes "Gwen"), never an ordinary word. The repo includes a dated, neutral competitor evaluation protocol with no results in it yet.
Not unless you turn that on. The correction is off by default; on a Mac the switch is Settings → Controls → “Actually” corrections. With it on, "We meet Tuesday, actually Wednesday" is typed as "We meet Wednesday" — it works between two numbers, days, months, or times. What it gets wrong is a contrast: "I thought it was Tuesday, actually Wednesday" would come out as "I thought it was Wednesday". Either way, saying "scratch that" takes back what you just said.
Nothing is typed. Those three modes are rewrites done by the optional cleanup model, and what you say to them is an instruction, not text to paste. On a default install, which has no cleanup model, they leave your field and your selection as they were and show "Compose mode needs the cleanup model — nothing changed" (or Reply, or Edit). Ordinary dictation and code dictation do not need the model.
Yes — AGPL-3.0, with a separate commercial license option. The full license and third-party notices live in the repository.