Your words stay on your Mac.

No account, no cloud relay, no sign-in. The transcription model, the voices, and everything they produce run locally on Apple Silicon, so they keep working with the network switched off.

Everything lives inside the machine

Four jobs, one boundary. Nothing in this box has a reason to reach the network.

Your Mac

DictationSPEECH → TEXT

Your speech is transcribed locally and dropped into the focused field. The audio is processed on-device.

Text cleanupTEXT → TEXT

A local model tidies your dictation as it lands: filler words, false starts, and stray punctuation, cleaned up on your Mac. The text never leaves it.

Read aloudTEXT → SPEECH

The full Kokoro voice library renders on your Mac. Choosing an accent never costs you privacy.

Clipboard historyON-DEVICE STORE

Recent dictations and copies stay in a scrollable history, held locally on your Mac and never synced to a server.

Two models, both on your Mac

Narrato runs its own speech models through Apple’s MLX framework on the GPU. They’re downloaded once on first launch, then live on your Mac. Every transcription and every spoken word is computed locally, with no inference call ever sent to a server.

SPEECH → TEXT

Parakeet TDT 0.6B v3

NVIDIA’s speech-recognition model, 4-bit quantized to about 607 MB. It turns your dictation into text entirely on-device.

ARCHITECTUREFastConformer · TDT

ON DISK~607 MB · 4-bit

RUNTIMEMLX · Apple GPU

TEXT → SPEECH

Kokoro 82M

An open-weight text-to-speech model with 82 million parameters. It renders all 54 voices locally, with no cloud voice API.

ARCHITECTUREStyleTTS 2 · ISTFTNet

PARAMETERS82M · 24 kHz

RUNTIMEMLX · Apple GPU

After that one-time download, each launch loads the models straight from disk, never from the network. The slowest moment is that first local load.

DON’T TAKE OUR WORD FOR IT

Verify it yourself

These are Apple’s own tools. Run them while Narrato transcribes and reads aloud, and watch your audio and text stay on the machine.

you@mac · -zsh · 74×12

you@mac ~ % networksetup -setairportpower "$(networksetup -listallhardwareports | awk '/Hardware Port: Wi-Fi/{getline; print $2}')" off

Wi-Fi off. Dictation and read-aloud keep working.

you@mac ~ % nettop -p Narrato

Watch the process while you dictate or read aloud. Local sessions add no connection; optional update and diagnostic calls may still appear. Press q to exit.

you@mac ~ %

Turn off Wi-Fi

Disconnect entirely. Read-aloud and dictation keep working, unchanged.

Watch the network

Point Little Snitch or the macOS firewall at it. Your words never generate a single byte; turn off diagnostics in Settings and it goes fully quiet.

Read the policy

Our privacy policy says the same thing in plain language.

Narrato does make a few network calls: anonymous diagnostics and crash reports (on by default in the beta, both switch-off-able), an optional update check, and a one-time model download on first launch. None of them carry your audio or text.

The two permissions, explained

macOS asks for these the first time you use the feature that needs them, never on launch, and never for anything else. You can review or revoke both in System Settings › Privacy & Security.

Microphone

Requested the first time you dictate. Narrato captures audio only while a dictation session is active; it’s transcribed on your Mac, and the recording is deleted the moment the text appears.

Accessibility

macOS files text insertion and keyboard control under Accessibility. Narrato uses it for one job: placing your transcribed words into the field you’ve focused. It doesn’t read your screen or watch other apps.

Private by architecture, not by promise.

Because your audio and text never leave your Mac, the sensitive part of your work stays yours.

Try Narrato free$29.99 USD