Parakeet
The fastest, most accurate
for English
meeting-notes.m4a1:00
Transcribed in
0.00s
Speak your mind.Kaze writes it down.
Hold a key, talk, let go. Kaze turns your voice into clean, polished text in any app, privately and entirely on your Mac.
Free & open source · macOS 26 · Apple silicon
0:00
Today 9:12
Just landed back from Lisbon 🇵🇹
Dinner this week? I owe you pastéis
Yes!! Welcome back
iMessage
To:Design Team
Subject:Launch update
Groceries9:41 AM
Trip ideasYesterday
Book listMonday
Q4 goals12/10/26
15 October 2026 at 9:41
Studio
# general
# design
# launch
# design
A
Ana 9:30
Where are we with the new onboarding? 👀
Message #design
If you can type there,
you can talk there.
One shortcut, every text field on your Mac. Kaze pastes where your cursor is, then puts your clipboard back the way it was.
Email
Messages
Notes
Docs
Code
AI Agents
0:00
Hold fn and say something.
Speak
One key.
That’s the interface.
There’s no window to open. Kaze waits in your notch until you press fn, or any shortcut you pick.
- Hold to talkLet go and it’s typed.
- Tap for hands-freeTap again when you’re done.
- Esc to cancelNothing gets typed.
Clean Up
Say it messy.
Read it polished.
Clean Up drops the ums and false starts, keeps the correction when you change your mind, and writes numbers, dates, times and emails the way you would. It runs on your Mac with S1-mini, or with your own ChatGPT, Claude or Gemini.
- Fillers removed
- Corrections kept
- Numbers & emails
- Casual → Formal
You said
so um can we move the sync to friday no wait thursday at like three thirty
Kaze typed
Can we move the sync to Thursday at 3:30?
- S1-mini
- ChatGPT
- Claude
- Gemini
On device
Speech recognition runs on Apple silicon through Core ML on the Neural Engine. No account, no uploads, no round trip to a server. Pick the model that fits how you talk.
NVIDIA · v2
Whisper
99 languages,
one shortcut
fnJust talkAuto-detect
駅はどこですか？
Apple Speech
Built in. Watch your words
appear as you speak
ListeningLive
Pick up the cake from the bakery on the way home
Download size
0 MB
New
Phonon-2
Parakeet, packed light. About a third
smaller, just as sharp for English
- Fermion Research
- Core ML · Neural Engine
Same Wi‑Fi, two downloadsEnglish
Parakeet v2480 / 480 MB
Phonon-2345 / 345 MB
Accuracy
- audio uploaded
- 0 bytes
- for a minute of audio
- ~0.5s
- European languages
- 25
- accounts needed
- None
Details
Vocabulary
Names and jargon, spelled your way
Add
KazeParakeetSupabaseFayaz
Replacements
Say a phrase, get exactly what you want
“gonna”going to
“my email”hi@studio.co
“shrug emoji”¯\_(ツ)_/¯
“brb”be right back
History
Every dictation, one click from your clipboard
0words per minute
0saved vs. typing
- Quick update: the launch moves to Thursday.2 min ago · Mail
- Hey, are we still on for dinner at 8?9 min ago · Messages
- Groceries: eggs, oat milk, sourdough…1 hr ago · Notes
Native
Feels like it shipped with your Mac
Built in SwiftUI for macOS 26. Liquid Glass, springy little transitions, and settings you’ll actually enjoy opening.
Is Kaze really free?
Yes. Kaze is free and open source under the MIT license. Download the latest build from GitHub Releases, or build it yourself from source.
Does my voice ever leave my Mac?
No. Speech recognition always runs on your Mac. Clean Up runs locally too, with S1-mini. The only exception is if you choose a cloud engine for Clean Up, by signing in with your ChatGPT plan or adding your own Claude or Gemini API key: then the transcript text (never your audio) is sent to OpenAI, Anthropic or Google for that one request.
What do I need to run it?
A Mac with Apple silicon running macOS 26. Models download once, on demand. After that, transcription works without an internet connection.
Which languages does it understand?
Parakeet v2 is the fastest for English, Phonon-2 matches it at about a third smaller, Parakeet v3 covers 25 European languages, Whisper handles 99 languages, and Apple Speech supports around 30 with nothing to download. Clean Up with S1-mini is English only; ChatGPT, Claude and Gemini handle any language.
How do I use it?
Hold your shortcut (fn by default), talk, and let go. Kaze types what you said wherever your cursor is. Tap the shortcut once instead to go hands-free, and press esc any time to cancel.
How is this different from cloud dictation apps?
Most dictation apps record your voice and send it to a server. Kaze does the work on your Mac’s Neural Engine, so it’s quick, private, and doesn’t need an account or a subscription.
0:00
Give your keyboard a break.
Free, open source and private by design. Download Kaze, hold fn, and start talking.
macOS 26 · Apple silicon · MIT licensed