Parakeet

Live captions for everything your Mac plays: videos, calls, podcasts. Korean, Japanese, English and about 30 more languages, detected on their own. Speech recognition runs entirely on your Mac's Neural Engine, and captions can be translated into about 20 languages, on your Mac too; no audio or text leaves it.

Download · parakeet.lost.plus · Apple silicon, macOS 15+

Using it

• Captions appear in a floating box whenever something speaks. Drag it anywhere; it remembers where. New words type in and scroll up line by line. Hovering over it shows buttons to turn captions off, copy the transcript, and pick a language to translate into.

• With a language picked, the translation shows large and the original small underneath. That language is taken as your own: while the audio is in it, the box stays hidden. macOS asks once to download each language it needs.

• ⇧⌘1 translates the screen: it shrinks into a picture, every paragraph on it is read and translated (into the language captions translate into, or your Mac's language when that's off), and the picture glows once it's done. Hover a paragraph to see it translated in place; drag over an area to keep it translated, or to read an area again where nothing was found. Pinch to zoom, two fingers to move around. Esc, ⇧⌘1 again or a click beside the picture puts the screen back. Languages not downloaded yet are offered in a bar under the picture. It needs Screen Recording: the welcome window (or the first ⇧⌘1) opens System Settings at that list, with a panel to drag Parakeet into it.

• The menu bar icon has a Captions on/off switch (⌘L), Copy Transcript, Translate To, Translate Screen (⇧⌘1), Open at Login and Check for Updates. It checks for updates on every launch (and daily while running) and offers to install them.

• The same switch works from a terminal:

  alias parakeet=/Applications/Parakeet.app/Contents/MacOS/Parakeet
  parakeet status
  parakeet captions on|off

The app isn't notarized by Apple, so the first launch needs System Settings → Privacy & Security → Open Anyway. A welcome window then asks for the audio permission and downloads the speech model (~640 MB) from Hugging Face.

How it works

• SystemAudioTap.swift captures all system audio with a Core Audio process tap and converts it to 16 kHz mono. It's rebuilt whenever the output device changes (headphones in or out), and it doesn't keep the Mac from sleeping.
• NemotronEngine.swift streams it through NVIDIA Nemotron 3.5 ASR (via FluidAudio's Core ML build). Each 0.56 s of sound is processed once and words are never rewritten, so it uses ~10% of one CPU core. For the same reason a sentence locks in as soon as its full stop appears; one that runs on locks at a comma (~3 s) or any word (~5 s). When nothing is playing, the model isn't run at all. Measured on an M2 with macmon: the speech model draws about 0.12 W while captioning (~40 mW CPU, ~80 mW Neural Engine), and Parakeet idles at ~1 mW when nothing is playing.
• CaptionPanel.swift draws the captions. Only a scroll offset is animated; animating the text itself cost ~90% CPU. Its Translator hands locked text to Apple's Translation framework, naming the source language itself (NaturalLanguage detects it) so macOS doesn't stop to ask. Pieces of an unfinished sentence are re-translated together, shown dim until it ends.
• SnapTranslate.swift is ⇧⌘1's work. It captures the screen with ScreenCaptureKit and reads it with Vision on the Neural Engine: whole, then in four overlapping quarters in the languages the whole read found (read whole, Vision drops short lines like a sentence-ending "ん。"). Vision can't read text written top to bottom (a manga's Japanese), so Live Text reads those columns alongside, and says where each one is through VisionKit's private line list; should that change, Vision reads alone. Lines group into paragraphs, split where a line stops short or a bullet starts, so a sentence that wraps is translated whole and a list stays a list; columns group right to left. Every downloaded language translates at once, paragraph by paragraph, nearest the pointer first. Each translation is painted over its original in the colour around it, in the text's own colour, size, weight and line spacing. Columns are translated into rows, centred in their speech bubble, which they may widen into; columns drawn over the art itself are left alone.
• SnapView.swift is what you see: the picture, its glow (the screen's own edges, blurred and stretched outward), the shimmer while it works, hover, drag and zoom. Every motion is the same crisp spring.

Development

Needs Xcode 26 or later (PermissionFlow's package manifest uses Swift tools 6.2).

scripts/build-app.sh          # → build/Parakeet.app
open build/Parakeet.app
build/Parakeet.app/Contents/MacOS/Parakeet --bench clip.wav   # 16 kHz float32 WAV
open -n -W build/Parakeet.app --args --snap in.png out.png ko  # ⇧⌘1 on an image

The interface is translated into the 28 languages Parakeet can caption, besides English. Translations live in app/Resources/Localizable.xcstrings (and InfoPlist.xcstrings for the permission prompt); Xcode can edit them, and scripts/build-app.sh compiles them into the app. Try one with open build/Parakeet.app --args -AppleLanguages "(ko)".

--bench plays a clip into the engine in real time and prints the captions and the CPU used. open build/Parakeet.app --args --rehearse-first-launch opens the welcome window as a new user sees it, with a fake model download and preparation pause, without touching the downloaded model. swift scripts/make-icon.swift redraws the app icon.

Releasing

git tag v1.2.0 && git push origin v1.2.0

.github/workflows/release.yml then builds and signs the app, packs it into Parakeet.dmg (scripts/make-dmg.sh: DMGMaker with the background from Packaging/dmg-background.html; re-render it with scripts/make-dmg-background.sh), publishes a GitHub release, and adds it to docs/appcast.xml, the Sparkle update feed served by GitHub Pages at parakeet.lost.plus (a DNS-only Cloudflare CNAME to lpfchan.github.io, so GitHub can issue its HTTPS certificate; the feed URL doesn't depend on the repo's name). The same site serves Parakeet's homepage, docs/index.html. Keep the repo name lowercase, though: 1.0.0–1.1.1 read their feed from lpfchan.github.io/parakeet/appcast.xml, which GitHub only redirects to parakeet.lost.plus while the name matches exactly (Pages paths are case-sensitive). Version = the tag; build number = commit count.

It needs three repository secrets:

• SPARKLE_PRIVATE_KEY: signs updates; the app only installs updates signed with it. Its public half is SUPublicEDKey in scripts/build-app.sh.
• SIGNING_CERT_P12, SIGNING_CERT_PASSWORD: the "Parakeet Self-Signed" code-signing certificate (base64 .p12). It isn't trusted by Gatekeeper, but keeping the same one means macOS remembers the audio permission across updates.

GitHub can't show secrets again; backups live in passage (folder sparkle) and in the maintainer's login keychain. Losing the Sparkle key means existing installs can never update again.

Credits

FluidAudio (Apache-2.0), Sparkle (MIT), PermissionFlow (MIT), and NVIDIA's Nemotron 3.5 ASR model (OpenMDW-1.1), downloaded at first launch. Parakeet itself is MIT-licensed. The license notices ship inside the app, in Contents/Resources/Licenses.