Docs

Setup, shortcuts, models, and the CLI. If something here is wrong or missing, tell us at hello@spok3n.com.

Requirements

macOS 14 (Sonoma) or later.
Apple silicon recommended. Intel Macs work, but transcription is noticeably slower because there is no Neural Engine to run the model on.
About 150 MB of free disk space for the default speech model, more if you switch to a larger variant.

Install

Download the DMG, open it, and drag Spok3n to Applications.
Open Spok3n. The app is Developer ID-signed and notarized by Apple, so it launches normally — no right-click workaround needed.
The speech model downloads on first launch. The default is about 142 MB and usually takes well under a minute.
Grant Accessibility permission when prompted.
Grant Microphone permission when prompted.
Look for the waveform icon in your menu bar. That means it is ready.

The two permissions, and why

Spok3n asks for exactly two, and both are load-bearing. Microphone is how it hears you. Accessibility is how it types the result into whatever app your cursor is in — macOS treats synthetic keystrokes as a privileged operation, so there is no way to insert text without it.

Neither permission involves the network. Transcription runs locally; the walkthrough of the on-device pipeline covers what that means in detail.

Dictating

Action	How
Hold to dictate	Hold Globe, speak, release
Toggle dictation	Tap Globe, speak, tap Globe again

Hold mode suits short bursts. Toggle mode suits anything long enough that holding a key gets tiring. A small bar appears at the bottom of the screen while this is happening:

Red — recording.
Orange — transcribing.
Green — done, text inserted.

Speech models

Spok3n ships base.en, an English-only Whisper model of about 142 MB, and you can switch variants in settings. Smaller models transcribe faster and make more mistakes; larger ones are more accurate and slower to load and run. small.en is the usual next step up if accuracy matters more to you than latency.

Models are downloaded once and cached locally. After that, dictation works with no network connection at all.

Personal dictionary

Speech models mishear proper nouns consistently — the same name wrong the same way every time. The dictionary fixes that by mapping a term to the strings it gets transcribed as. You can edit it in the app, or from the terminal:

spoken dictionary add "Spok3n" "spoken"

Teach Spok3n a term and one or more ways it gets misheard. Adding an existing term merges the new aliases rather than duplicating it.

spoken dictionary list

Print every dictionary entry. Add --json for machine-readable output.

spoken dictionary remove "Spok3n"

Soft-delete an entry. Removed entries stay recoverable for 30 days.

The CLI exists mainly so AI coding tools can manage the dictionary for you — notice a mistranscription in a conversation, add the correction, and it is fixed from then on. Spok3n also registers a URL scheme (spoken://dictionary/add) for Shortcuts and Automator.

Troubleshooting

The emoji picker opens when I press the Globe key

macOS claims the Globe key by default. Open System Settings → Keyboard, find "Press 🌐 to", and set it to "Do Nothing". Spok3n can then use the key without the picker appearing.

The menu bar shows a warning icon instead of the waveform

That means Accessibility permission has not been granted. Spok3n needs it to insert transcribed text into the app you are typing in. Open System Settings → Privacy & Security → Accessibility and enable Spok3n.

Dictation does nothing and no text appears

Check Microphone permission in System Settings → Privacy & Security → Microphone. If the overlay bar never turns red, the app is not receiving audio.

Transcription never finishes after a fresh install

The speech model download may have been interrupted. A complete model folder contains three compiled files — MelSpectrogram, AudioEncoder, and TextDecoder. Re-downloading the model from settings replaces a partial folder.

Transcription is slow

Spok3n runs best on Apple silicon, where the Neural Engine does most of the work. It runs on Intel Macs but noticeably slower. Switching to a smaller model variant such as tiny.en also helps.

The same name is transcribed wrong every time

Add it to your personal dictionary with the term and the way it gets misheard. That is what the dictionary is for — repeated, predictable errors on names and jargon.

Still stuck?

Email support@spok3n.com. The support page lists the rest of the contact addresses, and the changelog shows what shipped recently in case your problem was a bug we already fixed.