Hold a key,speak, let go.
Free dictation for macOS. Warble types what you say into any app, and NVIDIA Parakeet transcribes it on your Mac.
Can we move the design review to Thursday afternoon? I want the new onboarding flow in front of everyone before we lock the release.
Hold the key, then let go. Sample text from Warble’s preview mode; no microphone is used.

One window
Searchable and grouped by day, with the app each dictation went into, its length and optional audio replay.
The parts you touch



While you hold the key
| Recording | 16 kHz mono audio, from your input | while held |
|---|---|---|
| A brushed key | is ignored, nothing is typed | under 0.25 s |
After you let go
| Parakeet TDT v3 | transcribes on the Neural Engine | on your Mac |
|---|---|---|
| Spoken punctuation | turns “comma” into a comma | if switched on |
| Your dictionary | fixes names, expands shortcuts | on your Mac |
| The text | is pasted at the cursor | in the app in front |
| The pill | shows a check mark | for 1.2 s |
| Your clipboard | is put back as it was | after 1 s |
| History | keeps the text and the app’s name | Application Support |
Nothing leaves your Mac
- No accountNothing to sign up for or log in to.
- No subscriptionFree, with no word limit.
- OfflineAfter one model download from Hugging Face.
- Mic on, key downAudio is recorded only while you hold the key.
- Plain JSONHistory and dictionary in Application Support.
- No analyticsNo crash reporting either.
| Compared | Warble | Typical cloud dictation |
|---|---|---|
| Price | Free | Subscription, or a capped free plan |
| Word limit | None | Weekly cap on free plans |
| Where speech is processed | On your Mac | Remote servers |
| Account | None | Required |
| Works offline | Yes | No |
| Source code | Open, MIT | Closed |
Warble is not the only app that keeps speech on your Mac. Handy, VoiceInk, Superwhisper and MacWhisper run local models too. Warble’s take is a native macOS 26 app on the Neural Engine, free and MIT licensed, that does dictation and nothing else. Read the privacy page

The model is called Parakeet
NVIDIA Parakeet TDT v3 runs on the Neural Engine through FluidAudio. It downloads once, about 500 MB, and then works offline.
Warble began as a fork of open-wispr by human37, with one narrow goal: fast, local dictation in Polish. Its config still carries over on first launch.
For whole calls and meetings with the speakers told apart, there is EchoPad, built on the same model.
25 languages
or Auto-Detect- ENEnglish
- PLPolish
- DEGerman
- FRFrench
- ESSpanish
- ITItalian
- PTPortuguese
- NLDutch
- SVSwedish
- CSCzech
- +18All 28 codes
Install in one command
git clone https://github.com/pieralukasz/warble.git First, clone it
Needs macOS 26 and Xcode 26 or its Command Line Tools (Swift 6.2).
cd warble scripts/install.shThen run the script
It builds a release, puts Warble.app in ~/Applications and links the warble command.

Setup does the rest
Microphone and Accessibility, the 500 MB model, your key, a practice dictation.
Using a coding agent? Point Claude Code or Codex at the repository and ask it to follow the install steps for agents. Only the Microphone and Accessibility prompts need you. Coming from open-wispr? Warble imports its config on first launch.
Install Warble

