The useful details.

Find your build.

The Mac download is 0.4.0 (build 18), for macOS 26 or later and Apple Silicon. Settings groups controls into General, Writing, Appearance and Diagnostics, with Saved spellings and Phrases in the sidebar. The menu keeps recording, recent text and recovery actions available.

General also has Recording sounds and Launch at login. Appearance has a text font, Override individual fonts to pick a family for each part. Pill style offers Classic, Hi-Fi, Glass, Matrix and Minimal, Pill size is Automatic, Standard or Large, and Hi-Fi has a readout colour.

RC.2 refreshes the interface: lips in the menu bar that open with your voice, a new app icon, a glass setup window, a redesigned Settings window, Glass, Matrix and Minimal pill styles, and a live Appearance preview. Update checks are now on by default. This reference describes 0.4.0. If you’re still using alpha.1–3, download the new app manually once to get the current Settings window and built-in updater.

A key that suits you.

With the default Fn shortcut, hold to record and release to finish. Double-tap Fn or use Fn+Space for hands-free recording; tap Fn to stop. Esc cancels, with a five-second Undo offer.

Open Settings → General and change the dictation shortcut. Click the recorder, press your new key or combination, release all keys, then Save. Cancel keeps the previous choice.

In General, Recording behavior defaults to Hold to talk. Choose Press to toggle to start with one press and stop with the next. Releasing the key keeps recording in toggle mode. Escape cancels in either mode. Changing the mode cancels an active recording.

What the Fn remap does.

While Mouthtype runs with the default Fn shortcut, it remaps Fn at the keyboard level so a hold can start recording. That also hides the emoji picker, and Fn navigation combinations keep working. It’s the Remap Fn to hide the emoji picker toggle in Settings → Diagnostics, also in the menu, and it’s on by default.

A small helper, mouthtype-fn-guard, sits beside the app and watches it. If Mouthtype quits or crashes without restoring the key, the helper removes only Mouthtype’s mapping and leaves any other key mappings you set. Choosing a different dictation shortcut restores Fn’s normal behavior. Settings → Diagnostics shows Fn remapping status and any Globe key conflict.

Choose what listens.

Settings → General → Microphone follows System Default or remembers a named input for Mouthtype. It does not change the Mac’s system input. A missing chosen microphone stays selected and is reported as unavailable. Reconnect it or explicitly choose another input. Changes apply to the next take; use Refresh inputs after connecting a device.

Priority list is an opt-in microphone choice. Order your inputs; the first connected listed microphone is chosen at the start of a take. System Default fallback is a separate opt-in when no listed input is available. This does not switch devices during a take or establish Bluetooth, sleep or clamshell support.

Quiet other audio is on by default from 0.4.0, and stays off if you turned it off before. On supported output devices it mutes the current output during capture, then restores only the state Mouthtype still owns. It does not pause media or prevent other apps from hearing the microphone. Physical device behavior still needs tester feedback.

Live words is off by default and applies to the next recording. It shows provisional words above the pill; they may change or omit earlier words. Only the normal full-take result is cleaned up and inserted once after stopping. Cancelling discards the preview. It uses up to eight seconds of recent audio, not an eight-second processing deadline. Final recognition waits for any already-running preview model work to settle.

Turn an audio file into text.

Choose Transcribe Audio File… in the menu, then select one WAV, AIFF or M4A up to 10 minutes and 512 MB. Recognition uses the speech model already loaded on your Mac. The original audio stays unchanged.

Preparing and Transcribing show the current stage. Choose Cancel or close the window to cancel; it may wait for the current recognition operation to return. The full raw result is selectable. Explicitly use Copy or Save Text… to export UTF-8 text. Saving over the original audio is refused.

Nothing is pasted automatically. Imports do not apply cleanup, saved spellings or phrases, or change dictation history and statistics. Finish any dictation or recovery before importing. New recording, recovery and update installation wait while an import or its cancellation settles.

Cleanup and tone do different jobs.

Level chooses how much processing happens. Off keeps recognition output. Basic uses local formatting rules. Fast adds the optional S1-mini model, roughly 1.4 GB. Thorough uses the optional Gemma model, roughly 5.2 GB. Model downloads need internet access; processing runs locally. When a cleanup model is unavailable or misses its deadline, Mouthtype uses Basic instead.

Tone controls capitalization and ending punctuation. Formal keeps sentence capitals and punctuation, and removes or softens swearing at Basic, Fast and Thorough (Off bypasses cleanup): exclamations, emphasis and dismissals like "fuck off" are dropped, and swear nouns and verbs become a plain word ("asshole" becomes "jerk"). Casual drops the ending full stop. Very Casual also relaxes sentence capitals while protecting names and technical spellings. Questions keep their question marks. Tone does not otherwise authorize paraphrasing.

Settings → General contains Level; Writing contains Tone and per-app and website rules. Each rule can set a tone, a cleanup level or both, and a rule can turn cleanup on for one app or site even when the default is Off. Rules start empty. A matching website domain includes subdomains; the most specific domain wins before the app rule and the default. Safari, Chrome, Edge and Brave are eligible when the focused window exposes an HTTP/HTTPS document URL through Accessibility. Missing URL support falls back to app/default. URLs are not retained; the chosen tone and cleanup level stay with the take.

Clean Selected Text… opens an editable Basic cleanup preview for an accessible selection. Explicitly choose Accept to request replacement only if the original app, field, range and selected text still match. Check the field afterward. Changed or unsupported selections keep Copy result available. Password fields are excluded. There is no AI rewrite, automatic paste or retry.

Cleanup can use limited surrounding field text for boundary formatting and spelling hints, excluding secure fields. It doesn’t read previous messages as conversation history. Names and context can still be wrong, so read the result.

Keep the things you repeat.

Saved spellings: open Saved spellings from the Settings sidebar. “Heard as” is the word or phrase that comes back wrong; “Write as” is the spelling you want. Use specific phrases for ambiguous common words. These are exact whole-word or phrase rules, not a promise that every similar-sounding name will be corrected.

Use Import CSV in Saved spellings for files with name or heard,written headers. Review the preview before confirming. Duplicates are skipped; malformed or conflicting rows block the import. Save or cancel any editor draft first. If the library changes during preview, import again to review the current result.

Phrase replacements: save a spoken trigger and the exact text it should insert. For example, “my work email” can expand to an address you choose. Use Phrases in the Settings sidebar.

The phrase library starts empty. Matching ignores capitalization and prefers the longest whole phrase. Replacements expand after cleanup so their wording and line breaks survive. {date} inserts the current long date; {clipboard} uses the text captured from the clipboard for that take. Cleanup Off also turns replacements off.

Save or cancel a draft before switching entries. If a file is unreadable or changes outside the editor, keep your draft and follow the repair guidance rather than overwriting the original.

Recent dictations and stats.

The menu keeps your latest takes in a collapsed Recent Dictations submenu, which only appears when there are takes. Choosing an item copies its full text. Takes from secure fields are not listed. Stats lives inside Recent. File transcription doesn’t add to either.

Corrections Mouthtype can learn.

When you fix a dictated word in the field, Mouthtype can offer Remember this spelling? Accepting it saves a spelling rule locally, the same kind you can add by hand in Saved spellings. Turn this off with Learn from my corrections in Settings → Writing → Advanced writing.

Advanced writing switches.

Advanced writing in Settings → Writing holds the finer tone controls: Learn from my corrections, Lowercase sentence starts, Drop final full stop and raw dictionary editing. Changing these can show Custom as your default tone. Raw file actions open your Mac’s default editor and keep unreadable originals. Cleanup Off bypasses Lowercase sentence starts and Drop final full stop; Learn from my corrections still applies.

If the text needs another way in.

Paste Last uses ⌃ ⌘ V by default. Copy Last uses ⌃ ⌘ C. These actions keep the latest saved transcript available after restart, unless a conflicting dictation shortcut takes precedence.

If a take didn’t land, inspect the target first. An app’s unknown delivery status is not proof that the field stayed empty. Use the recovery card’s Copy action when needed, and avoid pasting the same text twice.

Saved Recordings lets you recover retained interrupted or failed takes. Recovery doesn’t automatically paste into another app. Recovery pauses new recording until it finishes, and a failed recovery stays available to retry.

Downloads and updates.

Alpha.4 and later include the Sparkle updater. Choose Check for Updates… in the menu or Settings → General. If a newer release is available, the updater offers the download and installation. You choose whether to install.

Automatic checking is on by default and there's no prompt for it. Turn off Automatically check for updates in General if you'd rather check yourself. This controls checking; it does not authorize an automatic installation.

Alpha.1–3 need one manual upgrade. Those builds do not include the updater. Download the Mac release ZIP, quit Mouthtype, expand the ZIP and replace Mouthtype.app in Applications manually. Open the new copy from Applications. Replacing the app does not erase the separate history and recordings described below.

The ZIP contains the app, for macOS 26 or later and Apple Silicon. License and readme files are included inside its Resources. The corresponding source code (GPL-3.0) is available alongside the app download. No GitHub account is needed for these links. For first-time installation, read the installation guide.

What stays on your Mac.

History, saved words and retained audio live in ~/Library/Application Support/Mouthtype. Text history is kept indefinitely. App preferences are saved separately by macOS. In Settings → General → Saved audio, choose 14, 30 or 90 days, or No age limit, with a 2 GB or 10 GB size limit. Defaults remain 90 days and 2 GB. Changes take effect at the next launch, removing expired completed Mouthtype recordings, then the oldest completed audio until it fits. No age limit still has a size limit. Text and timings stay. Unfinished recordings and recovery audio are exempt and may use additional storage. Imported files stay at their original locations. Removed audio cannot be restored.

Removing the app does not remove this data. Quit Mouthtype before removing the application. Delete the data folder only when you intend to erase the history and recordings inside it. Dictation doesn’t need a Mouthtype account.

Recognition and cleanup run locally. Network access is used for model downloads, HTTPS update checks and update archive downloads. Update checks do not upload transcripts or audio. Anonymous system profiling is not enabled by default.

When something needs attention.

Microphone access lets Mouthtype capture your voice. Accessibility access lets it insert the finished text. In System Settings → Privacy & Security, check both permissions for Mouthtype.

Open Settings → Diagnostics for status, model retry, shortcut health and permission repair actions. An attention message in the menu opens Diagnostics.

If nothing appears, confirm a text field is selected and the speech model is ready. Report the exact warning, Mouthtype version, macOS version, destination app, what you said and what appeared. Mention whether it appeared once, twice or not at all. Share only audio and text you choose to disclose.

Browser, terminal and Electron compatibility still need wider testing. A fresh install on another Mac, physical custom-key dictation and device changes remain unverified for 0.4.0. An update from one release to the next was observed installing on one Mac; other macOS versions and fresh installs aren't verified. Real file-recognition quality also needs tester feedback. Linux source testing is described below.

Quick fixes.

Settings → Diagnostics is the first stop. It shows model, permission and shortcut status, and each problem has a button.

Check for Updates… in the menu looks for a new build. Automatic checks are on by default and can be turned off in Settings → General.

What 0.4.0 doesn’t promise.

• Installing on a Mac that has never run Mouthtype wasn’t verified for 0.4.0. That was a maintainer scope decision.
• Bluetooth and USB microphones haven’t been physically verified.
• VoiceOver acceptance is deferred.
• Clear Glass over a dark window on a light wallpaper can be hard to read.
• The Matrix padlock sprite is low-detail at small size.
• The Hi-Fi card falls back to the Classic card for text its pixel font can’t draw.

Linux source testing.

Start with the Linux test-client guide for the source download and installation steps. Coding agents can follow the ordered install checklist.

The source archive includes an experimental, source-only daemon/CLI targeting Linux x86_64, Arch/Omarchy, Hyprland and PipeWire. The Omarchy bar plugin in tools/omarchy-plugin adds the menu, floating pill, paste delivery and settings; there is no packaged desktop release. The source includes a local Arch test-package builder, optional saved spellings and phrase replacements, and deterministic basic cleanup. Fast and Thorough cleanup run local llama.cpp models prepared on explicit request, and one audio file at a time can be transcribed; model choices and time budgets differ from the Mac app.

Read tools/linux-desktop/README.md in the source for prerequisites, launcher installation and explicit speech setup. The setup_speech.py helper downloads and verifies a pinned model archive and installs pinned binary packages in a dedicated environment outside the checkout. It does not record, start a service, set shortcuts or edit compositor configuration. Start the foreground service explicitly afterward.

Cleanup defaults off and needs no extra model. Basic formatting runs before literal saved-spelling and phrase expansion. Raw text stays recoverable if formatting fails. Automatic typing refuses multiline/control-bearing results rather than sending Return or Tab; recover the full formatted transcript explicitly. See the source guide for phrase dates, clipboard-context opt-in and limitations.

Recovery-only delivery is the default, with no clipboard or key changes. Automatic typing is an explicit opt-in with compositor and field limitations documented in that README. Synthetic audio, headless Wayland and remote speech-runtime checks do not certify a physical Arch/Omarchy desktop. That acceptance remains open.