This page went up on 2025-06-22, four days after OpenSuperWhisper 0.0.4 shipped. Several claims in that version were wrong, and two of them inverted what the people involved actually said. The corrections come first.
The correction that matters: the developer said “not tested”
The original version listed two items as working features — “Multi-language support – Automatic language detection with support for numerous languages” and “Custom transcription settings – Adjustable parameters for optimal performance” — and framed the page as a review written “after thorough testing.”
The project’s Readme.md as it stood at commit 6718352 on 2025-06-21, the state of the repository the day before this page was published, said this:
- 🌍 Support for multiple languages with auto-detection (not tested, but probably works)
- 🎛️ Advanced transcription settings (not tested)The developer marked both items as untested. This page removed the annotations, promoted both to confirmed capabilities, and attributed the upgrade to testing that never took place here. Every accuracy claim that followed — “thorough testing”, “extensive testing”, “exceptional recognition accuracy”, “rivals paid alternatives” — has been deleted rather than re-sourced, because no measurement sat behind any of them.
The annotations are gone from the current Readme.md, which now lists multi-language auto-detection without qualification. That was the developer’s call to make about the developer’s own software, and it was made in February 2026 — eight months after this page had already made it on the developer’s behalf.
Will it run on your Mac?
The Readme’s Requirements section says only “macOS (Apple Silicon/ARM64)”. It names no minimum macOS version, and filling in that blank is probably the most useful thing this page does. Two places state the floor:
| Source | What it states |
|---|---|
Homebrew cask opensuperwhisper | Required: arm64 architecture, macOS >= 14 |
Info.plist in the 0.1.0 build | LSMinimumSystemVersion = 14.0 |
macOS 14 Sonoma or later, Apple silicon only. The earlier wording here — “macOS (Apple Silicon/ARM64 optimized)” — implied an Intel Mac would run it, just slower. It will not run at all: the 0.1.0 executable is a thin arm64 Mach-O with no Intel slice, and “Intel macOS compatibility” is still unchecked on the project’s TODO list, filed as issue #15.
Is it real-time?
The Readme’s feature list opens with “Real-time audio recording and transcription”. The contribution TODO list, further down the same file, still reads:
- [ ] Streaming transcriptionBoth lines are in the Readme as of 2026-07-17. They are not contradictory so much as compressed: recording begins the moment you press the shortcut, but text arrives after you stop speaking, not while you speak. The earlier version of this page resolved that ambiguity in the wrong direction — “Process speech as you speak with minimal delay” — then contradicted itself a few hundred words later by listing “No Real-Time Display” as a limitation.
If you want words appearing as you talk, that unchecked box is the clearest signal available. It has stayed unchecked since 0.0.5, the release that introduced the TODO list, through 0.1.0 — the box did not yet exist when this page went up.
The VoiceInk comparison was backwards
This page described VoiceInk as a “Commercial service with subscription pricing”, using it as the expensive foil that made a free app look good. Checked on 2026-07-17, VoiceInk’s pricing page says “$0/mo”, “No Subscription” and “Buy once. Own forever.”, with one-time tiers of $25, $39 and $49 covering one, two and three Macs. Its GitHub repository describes it as “The best open-source alternative to Superwhisper & Wispr Flow. Voice-to-text app for macOS with no subscription”, under a LICENSE file headed GNU General Public License Version 3 — though that file runs to 1,442 bytes against the real license’s roughly 35,000, which is why GitHub declines to name it and reports “Other”.
So it is neither a subscription nor closed source. The error has a traceable origin: OpenSuperWhisper’s own Readme listed VoiceInk under “Free alternative to paid services like:”, and paid — accurate — quietly became subscription, which is not. That is a comparison written to flatter the app being covered. The February 2026 edit that dropped the “(not tested)” annotations deleted that section as well, so the wording this page misread is no longer in the Readme at all.
MacWhisper is named here without a price, the same way the earlier version left it — “One-time purchase with premium features,” no figure attached. That restraint is worth keeping. Its storefront builds the price in the browser rather than writing it into the page, so there is nothing to quote and cite the way VoiceInk’s plain-text pricing can be quoted and cited. VoiceInk gets exact figures above because its own site states them outright. MacWhisper gets none here because this page can’t stand behind one.
The privacy guarantee was invented
“Local processing – All transcription happens on-device, ensuring privacy and data security” appeared on this page. The 2025-06-21 Readme made no such claim. It listed “Local storage of recordings with transcriptions” — storage, a statement about where files land, not a guarantee about what the app does with a network connection.
That distinction matters more in 0.1.0, which downloads models from inside the app and therefore makes network requests by design. The bundle is also not sandboxed: codesign -d --entitlements against the 0.1.0 build reports com.apple.security.app-sandbox set to false, alongside accessibility and microphone entitlements.
None of this suggests the app ships your audio anywhere. Whisper.cpp runs locally, 0.1.0 carries a second engine, Parakeet, alongside it, and the source is MIT-licensed and readable by anyone. It does mean the privacy property rests on reading that source — now for two engines rather than one — instead of on a boundary macOS enforces — and a page stating it as settled fact was asking you to take its word for something it never looked at.
What expired since 2025-06-22
The app moved from 0.0.4 (2025-06-18) to 0.1.0 (2026-03-03). Four instructions here aged out:
- Models: hand-downloading
.binfiles from the Whisper.cpp Hugging Face repository still works and is still documented, but 0.1.0 downloads models from within the app, and has a second engine — Parakeet — next to Whisper. - Shortcut:
cmd + `was the Readme’s wording in June 2025. 0.1.0 binds a key combination or a single modifier key (Left ⌘, Right ⌥, Fn), and holding that shortcut records until you release it. A mouse-button trigger sits on the project’s master branch but is in no released version, 0.1.0 included. - Homebrew:
brew install opensuperwhisperinstalls a cask, not a formula. The claim that this “automatically handles dependencies and keeps the application updated” was wrong twice over — a cask places a prebuilt .app and does neither. - “System Preferences” is System Settings, renamed in macOS 13 — two releases before this page was written.
Where this leaves you
If you are on an Apple silicon Mac running macOS 14 or later and want free local dictation, the project’s Readme is the place to start, and it will always be more current than a page describing it — this one included.
Three things it will not tell you: its feature list and its TODO list disagree about streaming, and the TODO list is the one to believe; the macOS floor is 14, which the Requirements section omits; and the paid app it positions itself against costs $25 once and is also open source. That is the whole of what this page has that the Readme does not.
On accuracy — your voice, your accent, your microphone, your room — this page has nothing, and its earlier attempt to answer that was invented. That one you have to run yourself, and a free MIT-licensed app makes it cheap to find out.