How to Set Up Superwhisper for Dictation on Mac
Superwhisper is a macOS dictation app that runs a speech model locally and pastes the transcript into whatever field your cursor is in. You install it, grant two permissions, download a model, and then hold one key to talk. That is the whole loop.
If you want a repeatable setup — not just “it installed” — the steps below cover model choice, the permissions people forget, and the two settings that decide whether your dictation feels fast or laggy.
What do you need before installing Superwhisper?
A Mac running macOS 13 Ventura or newer. Superwhisper ships as a 30 MB download, and the app itself is free; what you pay for is model access. The free tier gives you a small model and a capped number of weekly words, while the paid plans open up the larger Whisper models and cloud processing.
Storage matters more than people expect. Whisper large-v3 is roughly 1.5 GB on disk and wants about 2 GB of spare RAM while it transcribes. On an 8 GB M1 Air, running large-v3 alongside Chrome and Slack gets uncomfortable. On 16 GB it is fine.
List of things to have ready:
- macOS 13 or later, Apple Silicon preferred
- About 3 GB free disk space if you plan on two models
- Microphone access for the app
- Accessibility permission so it can paste into other apps
How do you install and grant permissions?
Download the DMG from superwhisper.com, drag the app into Applications, and launch it once. macOS will block the paste step until you approve two things.
- Open System Settings → Privacy & Security → Microphone and switch Superwhisper on.
- Same menu → Accessibility, add Superwhisper. Without this, dictation transcribes but never lands in your text field.
- Optionally enable Input Monitoring if you want the hold-to-talk key to work while another app is fullscreen.
The Accessibility step is the one that trips people up. If you dictate a sentence and the app window shows the transcript but the text box stays empty, that permission is the cause nine times out of ten. Toggling it off and on again fixes the stale TCC entry.
Which model should you pick?
Model choice is a speed-versus-accuracy trade, and the answer depends on your machine.
| Mac | Suggested model | Rough latency | Notes |
|---|---|---|---|
| M1 / M2, 8 GB | small or base | Under 1s for 10s of speech | Accuracy drops on names and jargon |
| M1 Pro / M2 Pro, 16 GB | medium | ~1.5s | Good middle ground |
| M3 Pro and up, 18 GB+ | large-v3-turbo | 1-2s | Best accuracy, still fast |
| Intel Mac | cloud mode | Network dependent | Local inference is too slow |
Start with the model one step below what you think you need. It is easier to upgrade after a week of real use than to fix a setup that drops words because the model can’t keep up.
How do you get dictation to sound like your writing?
Out of the box, dictated text reads like speech: filler words, no punctuation control, and line breaks where you paused to think. Two settings fix most of that.
Language and punctuation. Set the output language explicitly rather than leaving it on auto. Auto-detect flips mid-sentence when you say a product name in another language and the rest of the paragraph gets transcribed with the wrong punctuation conventions.
Custom vocabulary. Every dictation tool lets you add words it should know. Put your company name, product names, acronyms, and the surnames of people you email often into that list. This is the single highest-return five minutes in the setup, especially for Superwhisper Dictation users who dictate in Slack with lots of internal jargon.
AI post-processing. Superwhisper can pass the raw transcript through a language model with a custom prompt — for example, “remove filler words, keep my sentence structure, do not add anything.” Turn it on only after the base transcript is clean. Post-processing a bad transcript produces confident-sounding nonsense.
Who is this setup right for?
Anyone on a modern Mac who writes more than 800 words a day by hand, especially people writing on a laptop away from a desk. The local models mean no internet round trip, so latency stays flat even on hotel wifi.
It is a poor fit if you work primarily on Windows or Linux — Superwhisper is Mac-only. It also does not help if you need transcription of long pre-recorded meetings; this is a live keyboard-replacement tool, not a meeting recorder.
Quick Summary
- Superwhisper is a Mac-only dictation app with a free tier and paid plans for larger models.
- Grant Microphone and Accessibility permissions, or transcripts won’t paste.
- Match model size to RAM: small/base on 8 GB, medium on 16 GB, large-v3-turbo on M3 Pro and up.
- Add a custom vocabulary list before you complain about accuracy.
- Skip AI post-processing until the base transcript is clean.
