Tamka is a free, open-source dictation app for macOS that works offline once a speech model is downloaded.
Press a shortcut, speak, then press the shortcut again, and the text appears at your cursor. Speech recognition runs on the Neural Engine with NVIDIA’s Parakeet models, and your audio never leaves the Mac. On a MacBook Pro with an M5 Max, the text arrives about 0.2 seconds after you stop, even after 50 seconds of speech.
Before typing, Tamka removes “um” and “uh”, collapses stutters, and turns spoken commands such as “comma”, “new line” or “open paren” into the characters they name. With Parakeet Ultra loaded, it uses about 0.5 GB of memory, most of it model weights held by the Neural Engine.
Dictating
Click where the text should go: an email, a chat, a code editor or a search box.
Press ⌃V and speak. An indicator with a timer shows that Tamka is listening.
Press ⌃V again. Tamka types what you said at the cursor and keeps a copy on the clipboard.
Press Esc while Tamka is listening to throw the recording away. ⌃V is the default shortcut. You can pick any combination that includes ⌘, ⌃ or ⌥ in Settings › General.
The indicator
The indicator never takes focus from the app you’re typing in. Drag it anywhere, including onto another display, and it appears in the same place for every dictation after that. If that display is disconnected, it goes back to the bottom centre of the screen with the pointer.
Recent dictations in the menu
Click a recent dictation to copy it again. Recent dictations are kept in memory only and are gone when Tamka quits.
Start Dictation becomes Stop and Paste while you’re speaking.
Transcribe Last Recording Again runs your last recording through the current model, which is useful after switching models.
Problems such as a missing permission or a model that hasn’t been downloaded appear as menu items that open the fix.
The menu bar icon turns red while Tamka is listening.
Example dictations
Pick a speech model
The app doesn’t include a speech model. On first launch, Settings › Models opens with Parakeet Ultra marked Recommended. Models work offline after the download, and removing a model moves it to the Trash.
Size covers the download and the memory the model uses while loaded. Speed is how soon text appears after you stop speaking. Accuracy is the word error rate on everyday speech. Hover over a grade in the app to see the measurements behind it.
Model
Size
Speed
Accuracy
Best for
Parakeet Ultra (recommended)
B
A
A
Everyday dictation, with the best punctuation
Parakeet Phonon-2
A
A
A−
Casual, unscripted speech
Parakeet v3
A
A
A−
Speaking any of 25 European languages
Parakeet v2
A
A+
B+
The fastest results
Apple Speech
A+
B
B
No separate model download; macOS may fetch its language files first
Spoken commands and cleanup
Say a command and Tamka types the symbol with the spacing fixed: “call foo open paren bar close paren” becomes call foo (bar). The table below shows 10 of the 112 default commands.
In Settings › Text you can turn filler removal and stutter cleanup on or off, write numbers, prices and dates as digits (experimental), and search, add, edit or remove commands. ⌘Z undoes changes, and Restore Defaults… brings back the original list.
Say
Get
Say
Get
“comma”
,
“new line”
a line break
“question mark”
?
“new paragraph”
a blank line
“open quote” … “close quote”
" … "
“open paren” … “close paren”
( … )
“at sign”
@
“em dash”
—
“percent sign”
%
“smiley face”
:)
Settings
Settings › General holds the shortcut, a recording time limit, the start and stop sounds, the indicator position, whether Tamka pastes the text and keeps it on the clipboard, launch at login, the microphone and Accessibility permissions, and the activity log.
Changes apply immediately.
Accuracy and speed
Parakeet Ultra was tested against Whisper large-v3-turbo on 117 public recordings (32 minutes) of read speech, meetings and earnings calls with human-checked transcripts, on a MacBook Pro with an M5 Max.
Measure
Whisper large-v3-turbo
Tamka with Parakeet Ultra
Words wrong, all recordings
9.9%
7.7%
Words wrong, without meeting recordings
6.9%
6.2%
Memory
1.2–1.6 GB
about 0.5 GB
Wait after a 50-second dictation
471 ms
170 ms
Without the meeting recordings, the difference in accuracy is within the margin of error. The full benchmark compares 11 model configurations. To compare models on your own voice and vocabulary, see eval/README.md.
Get Tamka
Download the zip from the latest release. It’s signed and notarized by Apple.
Drag Tamka into your Applications folder and open it.
Build from source
You need Xcode 26 or later and a current stable Rust toolchain from rustup.
git clone https://github.com/piedol/tamka.git
cd tamka && scripts/bundle.sh --install
The script builds the app and its speech engine, signs both, copies Tamka.app to /Applications and opens it. It signs with TAMKA_SIGN_IDENTITY if that’s set, or else with the first Developer ID Application certificate in your keychain. Without either, it uses an ad-hoc signature, and macOS asks for permissions again after every build.
When Tamka first opens
Settings › Models opens. Click Download next to Parakeet Ultra (632 MB).
Allow Accessibility so Tamka can paste the text. Without it, Tamka copies the text and you paste it yourself.
Allow the microphone the first time you dictate.
Tamka has no Dock icon. To start it when you log in, turn on Launch Tamka at login in Settings › General. To remove it, follow Uninstall in the README.
Privacy
Speech recognition runs entirely on your Mac.
Tamka connects to the internet only to download a model. Parakeet models come from Hugging Face, and Apple Speech’s language files come from Apple.
The microphone is on only while you’re dictating.
Tamka keeps your last recording at ~/Library/Caches/Tamka/last.wav for Transcribe Last Recording Again. Each new dictation replaces it.
The log at ~/Library/Logs/Tamka.log records events and timings, never your words.
Frequently asked questions
Is Tamka free and open source?
Yes. Tamka is free to download and use, and its source code is on GitHub under the GNU General Public License v3.0 or later.
Does Tamka work offline?
Yes. Speech recognition runs entirely on your Mac. Tamka connects to the internet only to download a speech model. Parakeet models come from Hugging Face, and Apple Speech’s language files come from Apple.
Which Macs does Tamka support?
Macs with Apple silicon on macOS 26 or later. The Apple Speech engine uses SpeechAnalyzer, which is new in macOS 26.
How accurate and fast is Tamka?
On a MacBook Pro with an M5 Max, Parakeet Ultra got 7.7% of words wrong across 117 public recordings, against 9.9% for Whisper large-v3-turbo, and returned text 170 ms after a 50-second dictation.
How do I install Tamka?
Download Tamka-macOS-arm64.zip, unzip it, drag Tamka into Applications and open it. On first launch, download Parakeet Ultra (632 MB) and allow Accessibility and the microphone.
How do I uninstall Tamka?
Quit Tamka, move Tamka.app to the Trash, then move ~/Library/Application Support/Tamka, ~/Library/Caches/Tamka and ~/Library/Logs/Tamka.log to the Trash. Downloaded models are in ~/Library/Application Support/FluidAudio/Models.
Use it anywhere, including at work, and read, change and share the source. If you distribute a modified version, release its source under the same license.
Speech runtime: FluidAudio (Apache-2.0). Models: NVIDIA Parakeet (CC-BY-4.0); Parakeet Ultra by Moondream; Phonon-2 by Fermion Research. Each model’s authors set its license.
