SpokemarkA word is on screen about half a second after it is spoken
0.45sMedian, speech to screen
Open source · 100% on-device
Live captions for everything your Mac plays
Calls, videos and podcasts, captioned as they're spoken by a speech model running on your Mac. No cloud, no account, no bot in your meeting.
Download for macOSMeasured on a MacBook Air M4, with real speech through the speakers
Errors per word over a 30-minute session of 2,750 words
2.3%Word error rate
Silence produces no text: 621 s of pauses in one session
0Words invented in silence
One pass of the model over a sliding window, on the GPU
179msInference, median
The system audio tap delivers sound to the app quickly
32msTap to app
Even the slow words arrive within about a second and a half
1.3s95th percentile, speech to screen
Whisper engine, standard model (base.en), over a 30-minute session. The new default engine, Parakeet, made fewer errors in our file comparison (0.4 % vs 2.6 %). Methods, raw numbers and what isn't verified yet: RESULTS.md
How it works
From a Zoom call to a YouTube video, Spokemark captions what your Mac plays, on your Mac.
And so my fellow Americans!
Ask not!
What your country can do for you? Ask what you can do for your country
Real output, replayed at its real timing: dim words are provisional and can still change (log).
Why Spokemark?
Built for reading. Captions that keep up with the voice.
Spokemark sits in your menu bar and puts a clean caption over whatever is on screen, with a transcript you can search later.
A CoreAudio process tap hears the Mac's output, so there's nothing to install in Zoom or Meet and playback stays audible.
An overlay above full-screen apps and every Space. Standard, High contrast and Cinema presets; 2–3 lines, top or bottom.
Every session as text, never audio. Full-text search and export to .txt, .srt or .md. Deleted after 30 days by default.
The translation under the original, or alone, by Apple's on-device Translation framework. Needs macOS 26.
Experimental: a separate speaker model tells voices apart on the Neural Engine, with any engine. A new line with a dash starts when someone else talks, and each speaker's translation goes right under their line. Nothing about voices is saved.
TRIGGERSound plays on your Mac⌥⌘C
STEPRecognize speech as it streamsParakeet
STEPTell speakers apartoptional
STEPTranslate each finished sentenceoptional
OUTPUTA caption over every app, a searchable transcript, and an optional on-device translation.
Open source
Built in the open
No paid tier and no telemetry to trust. The app, the audio helper and the speech engine are all in one MIT repository you can read and build.
MMIT License
Use it, change it, ship it, sell it. The Spokemark name and logo stay with the project, so forks pick their own.
Read licenseRMeasured results
Every number on this page comes from a documented run on real hardware, with what hasn't been verified listed too.
See resultsBBuild from source
Clone, npm install, build the Swift and C++ helpers, download the model, package. Pull requests welcome.
Apps & Privacy
Works with anything your Mac plays
Spokemark listens to the system output, not to a specific app. Recognition, translation and history all stay on your Mac; the only network request is the one-time model download.
How capture worksCalls, browsers, players and podcasts: if the Mac plays it, Spokemark can caption it
See all features- Zoom
- Google Meet
- Microsoft Teams
- Discord
- YouTube
- Podcasts
- Safari
- Chrome
- Any other app
Privacy built into how it works, not a setting you turn on
Security policy- No account
- No telemetry
- Audio never saved
- On-device recognition
- On-device translation
- Text-only history
- 30-day retention
- SHA-256 model check
- Open source, MIT
Calls haven't been tested end to end yet; YouTube in Chrome and local playback have.
Menu bar first
Shortcuts
A caption engine in your menu bar.
Start with one click or ⌥⌘C, pause to keep the last captions on screen or stop to hide them, and fix problems like a missing permission or model right from the panel.
Start, pause, stop
⌥⌘C starts or stops captions from anywhere. Pause keeps the last captions on screen; Stop hides them. ⌥⌘O shows or hides them, ⌥⌘E moves them.
Silent when nothing is said
A voice detector gates the model, so pauses stay empty: 0 words in 621 s of silence.
Knows what's playing
The caption shows the icon and name of the app the audio comes from, with waves that follow the voice.
Native on macOS
The same look as this page, dark by default with light and system options; full keyboard navigation and VoiceOver labels; English and Portuguese UI.
Ready to read along?
Download Spokemark, allow System Audio Recording, and play anything. The speech model downloads on first launch.
Builds aren't notarized yet (no Apple Developer ID). On macOS 15 or later: open it, click Done, then System Settings › Privacy & Security › Open Anyway. On macOS 14: right-click › Open.
Version0.4.0 · released 2026-10-15
MacApple silicon (M1 or later)
macOS14.2 or later · translation needs 26
.dmg
Size149 MB · speech model 626 MB
PermissionSystem Audio Recording only
Download for macOSCaptures what the Mac plays, not your microphone. Verify with SHA256SUMS.txt · all releases