How to type by voice in any app on a Mac, with the speech recognised on the Mac itself

What you need, and setting it up

Deixis runs on a Mac with Apple silicon, on macOS 13 or later. Nothing else needs installing: no runtime and no redistributable.

1. Download the disk image from the download link. It is notarized. Open it and drag Deixis to Applications.
2. Fetch the speech model. The app does not carry the model's weights, so the first launch asks you to press Download once. Until that finishes there is no speech model, and dictation says so rather than failing quietly.
3. Grant two permissions. The first time you dictate, macOS asks for Accessibility and microphone permission. Grant both in System Settings → Privacy & Security.

Updates are checked in the app, under Settings → Updates. The Mac has a release feed of its own, so it ships on its own schedule rather than waiting for Windows.

The one gesture: hold, speak, release

Hold Ctrl+Option, speak, release. The text lands at the caret of whatever has focus, and your clipboard is put back the way you left it. That is the product; everything else is a variation on the same chord. If the chord does not suit you, Settings → Hotkey captures a new one as you press it.

Double-tap the chord for free mode. It keeps recording until you double-tap again, so you can talk through a paragraph you have not finished thinking. The finished note is pasted at the caret, left on the clipboard and saved as a Markdown file.

Tap the chord during a free-mode session for the toolbar, a fan of actions at the cursor. Take the text you have selected, grab a region of the screen, or finish. What you said and what you pointed at come out as one note, interleaved — useful for a bug report, a review comment or instructions for somebody else.

Hold the chord and the mode palette appears above the listening pill, one chip per mode, each with its letter. There is more on the gestures page.

Screen grabs and screen recording

Two more things the Mac build does, both from the same chord.

• Screen grab and annotation. The screen freezes, you drag a region, and you either quick-save it or open the editor: arrow, rectangle, ellipse, pen, highlighter, text, pixelate to redact, and numbered step badges. Every save keeps a sidecar file beside the image, so the grab stays re-editable rather than flattened. See screen capture.
• Screen recording. It records the screen to a video file, and the same chord stops and saves it. While it runs, a small bar with a red dot, the elapsed time, Pause and Stop shows that it is rolling.

What the Mac build does not do

It is worth knowing before you download rather than after.

• No meeting mode on the Mac. Recording and transcribing a meeting is on Windows and on the Android app, not on the Mac today. The platforms page keeps the current state of each build.
• No translate on the Mac. Translating a selection is on Windows and Android.
• English speech models. The two models that ship, base.en and small.en, are English.
• Apple silicon only. There is no build for an Intel Mac.

What stays on your Mac, and what leaves

Speech recognition runs inside the app, on your Mac. The recording, the model and the transcript stay local, and nothing uploads the audio a dictation records, under any setting or on any plan. There is no telemetry.

A fresh install makes two calls on its own, and neither carries anything but the request: the update check at launch, and a check for an announcement from us when the Settings window opens, which has an off switch of its own. The speech model download happens once, when you press the button.

Cleanup can use a language model, and that route is off until you configure a destination. When cleanup does send something, it is text, never audio: the transcript, your vocabulary terms, the name of the app and field it is going into, and your previous sentence in that field — not the window title, which is where document names and customer names live. The one thing that uploads a file is Share a link, and only for the file you picked, when you pressed it. The full list, row by row, is on what leaves your machine.

Teach it your words

The dictionary is the biggest accuracy lever there is. Without any vocabulary, the standard model writes shadcn as "shadkin" and useEffect as "use effect", and one line of vocabulary fixes both. The app's Dictionary tab holds four things:

• Words to expect — names and jargon the recogniser should listen for. Only the first 36 reach the speech model itself, because past that accuracy measurably degrades; the editor tells you when you are over.
• Fixes — find-and-replace corrections applied after recognition, at word boundaries. These are unbounded, so once you are past the cap, put new corrections here.
• Filler words — extra words to strip, beyond the built-in "um" and "uh".
• Vocabulary libraries — five packs ship: software development (on by default), accounting and finance, business and product, medical, and legal. You can add your own.

One more lever needs no software at all: a headset beats a laptop's built-in microphone in exactly the conditions that break dictation, such as a room with other people in it. The dictionary page has the details.

Cleanup: rules, or a cloud model when you want one

Cleanup is the second pass that punctuates and structures what you said. With nothing configured it runs on rules on your Mac, and no text leaves. Settings → Cleanup decides when a language model is asked instead:

• When it helps, the default. Rules handle the sentence locally unless you ask for a transform ("make this a bullet list") or dictate a long stretch with no punctuation.
• Every time, for better prose on every sentence, at the cost of about a second on each sentence, and each sentence leaving the machine.
• Never, which stops dictation calling a model at all.

With Deixis's own endpoint, which is what Pro pays for, requests route only to model endpoints that keep nothing and do not train on your text. The cleanup page covers the settings, and free and Pro the other ways to run the cloud features.

One thing cleanup cannot do is fix what the recogniser misheard. It rewrites prose, not hearing, so a misheard name stays misheard. Mishearings are the dictionary's job.