100% local
Audio is transcribed on-device with Parakeet. Your voice never touches a server.
⠿⠶ Open source · Free · Local-first
Speak. It types.
Nothing leaves your computer.
Whispy is a private voice-dictation utility for macOS and Linux. Hold
your push-to-talk key, talk, and release — or flip on toggle mode and
press once to start, once to stop — your words land in whatever
app you're using. All transcription runs locally with NVIDIA's
parakeet-tdt-0.6b-v3. No cloud, no account, no telemetry.
Free & open source · macOS (Apple Silicon & Intel) and Linux (X11)
Whispy, live in your menu bar
See it work
Hold your push-to-talk key, speak, release — Whispy types into whatever field has focus. Here's the whole loop. Don't want to hold it? Turn on toggle mode: one press starts, the next stops. The trigger is configurable; it defaults to Fn on macOS and Right Ctrl on Linux. On macOS, pick it straight from the menu — Fn, Right Command, Right Option, or F13.
Why Whispy
One key, anywhere
Hold your push-to-talk key to record, release to insert text into any focused field — terminal, browser, editor. On macOS the text is handed over through the clipboard, so accents and punctuation arrive intact whatever your keyboard layout.
Hold or toggle
Not every dictation fits under a held-down key. Flip on toggle mode and one press starts recording, the next press stops it — composable with any trigger.
macOS & Linux
Native on macOS and Linux (X11). OS-specific seams sit behind a ports-and-adapters layer bound at runtime.
Knows what you're speaking
25 European languages, detected automatically — even when you switch mid-sentence. Nothing to configure.
Lives in the tray
Runs quietly as a background daemon in the menu bar (macOS) or system tray (Linux). No window to manage — flip Start at login on macOS, and on Linux the installer enables a user service, so it's there every boot either way.
Fast enough to forget
A two-second phrase transcribes in about 75 ms on an M1 Pro. Text lands as you release the key. In toggle mode, text is typed after each pause while you're still speaking.
Quiet when you are
Tap the key by accident, or hold it in a noisy room, and nothing gets typed. Two gates — loudness, then voice detection — keep non-speech away from the model instead of letting it answer with a stray “Yeah.”
Your names, spelled right
List the names, brands and jargon you actually say in custom_vocabulary, and a near miss is corrected after transcription — wispy comes out as Whispy. Still on-device, like everything else.
How it works
-
1
Hold your key
A soft chime confirms Whispy is recording — Fn on macOS, Right Ctrl on Linux by default.
-
2
Speak
Whispy listens — the tray shows it's live (macOS adds an on-screen waveform).
-
3
Release
Whispy transcribes locally and types the text into the active field.
Prefer not to hold the key? Turn on Toggle mode in Settings — one press starts, the next stops. Text then appears as you speak, pause by pause.
Free, open source, and 100% yours
Whispy exists for one reason: dictation that never ships your voice to the cloud. No servers, no accounts, no subscriptions — just open tools running on your own machine. Zero data leaves your computer — by design, not by promise.
Open source
GPLv3-licensed and fully auditable. Read every line, fork it, make it yours.
Free forever
No subscription, no paywall, no “pro” tier. Install it once and it's yours.
Runs on your machine
Self-hosted by default on macOS or Linux — local inference, nothing to upload, nothing to trust.
One model, no choices
Whispy ships a single model and downloads it on first use. There is no size to pick and no language to set.
| Model | nvidia/parakeet-tdt-0.6b-v3, int8 ONNX, CPU |
| Download | 639 MB, automatic on first run |
| Speed | ~75 ms for a 2-second phrase (Apple M1 Pro) |
| Memory | ~1.4 GB resident while loaded |
| Languages | 25, detected automatically — Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian, Ukrainian |
| Non-speech | Discarded before the model — needs both loudness and detected voice |
| Vocabulary | Names and terms you list in custom_vocabulary are corrected after transcription |
Speech outside those 25 languages will not transcribe. The list comes from the model card.
Install in one line
Runs on macOS and Linux (X11). Audio capture uses
sounddevice/PortAudio — no sox needed.
curl -fsSL https://raw.githubusercontent.com/albrones/whispy/main/scripts/bootstrap.sh | bashOn macOS this builds and installs the signed
Whispy.app (needs the Xcode Command Line Tools). Autostart
is the in-app "Start at login" toggle.
On Linux, install the system deps first (X11 session required — Wayland users pick an "Xorg"/"X11" session at login):
# Debian/Ubuntu — xdotool (text injection), xclip (clipboard delivery), PortAudio
sudo apt install xdotool xclip libportaudio2
curl -fsSL https://raw.githubusercontent.com/albrones/whispy/main/scripts/bootstrap.sh | bash