Blog
Set up local voice typing on macOS with Whisper
Install SleekVoice, grant the macOS permissions voice typing needs, and use push-to-talk without sending speech to a SleekVoice server.
Published · Updated · 5 min read · Bharat Bhavnasi
On this page
- Install the app
- Let the Mac hear you
- Let the app type
- Use push-to-talk on purpose
- Teach it the words you actually use
- Check that it is actually local
- When the words do not show up
- Questions
- Do I need a GPU or a specific Mac chip?
- Can I point SleekVoice at my own Whisper weights?
- Is this the same as installing OpenAI’s Python package?
Local voice typing means the speech is turned into text on the Mac you are speaking into, rather than uploaded for a server to transcribe. SleekVoice is a free, GPL-3.0 app that does that job with Whisper, including when the Mac is offline, and then types the text into the focused app. This is the setup path: install, permissions, push-to-talk, dictionary, and the failures that look like bugs but are switches.
Whisper, in this post, means the speech recognition model OpenAI released under the MIT license. The model card and code are at github.com/openai/whisper (opens in a new tab). It is a general-purpose recognizer. It is not a SleekVoice feature name we made up. SleekVoice has not published which Whisper size it runs. Setup does not depend on that number. You do not pass a model flag. You install the app.
Install the app
SleekVoice is macOS only. The command the project documents is:
brew install --cask sleekvoice
Install Homebrew (opens in a new tab) first if Terminal does not know brew. If Homebrew reports that the cask cannot be found, stop retrying the same command with extra flags. Download a build from GitHub releases (opens in a new tab) and move the app into Applications. There is no SleekVoice account to create before the first launch.
macOS may ask you to confirm an app downloaded from the internet. That is Gatekeeper. It is not a sign that audio is being sent anywhere.
Let the Mac hear you
An app cannot transcribe what it is not allowed to record. Apple’s path is Apple menu, System Settings, Privacy & Security, Microphone. Turn SleekVoice on. The article is Control access to the microphone on Mac (opens in a new tab).
If SleekVoice is absent from the list, open the app and use the shortcut once so macOS records the request. A denied prompt does not keep asking forever. The same Settings pane is where you undo the denial.
If you use a headset or an external microphone, make sure the Mac’s input device is the one you are talking into. Apple’s Dictation guide describes a microphone-source menu for Apple’s feature. The principle is the same if SleekVoice is listening to a muted input: the model never hears the sentence, so the dictionary and the shortcut are irrelevant until the input is right.
Let the app type
Hearing you is not the same permission as inserting text into Mail, a browser, or a notes app. macOS treats control of other apps as Accessibility access. Apple explains the prompt and the pane in Allow accessibility apps to access your Mac (opens in a new tab). When SleekVoice asks, you can open System Settings from the alert and enable it, or deny it and enable it later from Privacy & Security, Accessibility.
This post does not describe an in-app permission wizard, because the repository has not published one. The system panes are the source of truth. Turn the switches off when you uninstall the app if you do not want them left on.
Use push-to-talk on purpose
Push-to-talk is hold, speak, release. The shortcut is held for the duration of the phrase. Letting go is the end of the take. The text is then typed at the cursor.
Click a field that accepts text before you hold the keys. The Desktop does not. A selected message that you have not clicked into does not. A short test sentence is a better first try than a paragraph, because you can see whether the cursor moved.
The project has not published the default shortcut. Read it in the app. If nothing happens when you hold the keys you remember, either those keys are wrong or another app already uses them. System Settings, Keyboard, Keyboard Shortcuts is where macOS lists system and app shortcuts. Change one side of the conflict. Do not copy a shortcut from Wispr Flow or from Apple Dictation and assume SleekVoice bound the same keys.
Per-app modes are the other half of “it works in any app.” SleekVoice keeps a mode per app. The names of those modes are not published, so set what the app offers for the programs you use, then switch apps and dictate one line in each. The mode should follow focus. If it does not, the second app probably was not the focused one. Click its text field.
Teach it the words you actually use
A personal dictionary is how you stop correcting the same name. Add the spelling you want for people, products, and acronyms. The project describes the dictionary and does not describe an auto-learn toggle, so do not expect Wispr’s “auto-add” behavior. Wispr documents that behavior for Wispr. Add the word yourself in SleekVoice, then say it again in the app where you needed it.
The dictionary stays on the Mac. There is no SleekVoice account that would sync it to a phone, because there is no phone app.
Check that it is actually local
The project’s test is the one it stated: transcription still works when the Mac is offline. Turn Wi-Fi off after the app is installed, click a local note that does not need the network to accept text, hold the shortcut, and speak. If text appears, the take did not require a round trip to a SleekVoice server. SleekVoice does not publish a cloud fallback that would quietly upload audio when the network returns.
If the app asks to finish a download on first launch, do that while you are online, then repeat the offline check. This post will not name a model file the project has not named.
Apple’s own Dictation is a different switch. Apple says Keyboard settings tell you whether general text Dictation is processed on the device or sent to Siri servers. Read that note if you are testing Apple Dictation. It is not a readout of SleekVoice.
When the words do not show up
Go in order.
The cursor is not in a text field. Click one.
The microphone switch is off, or the input device is wrong. Fix that before you touch the dictionary.
Accessibility is off, so the app can hear you and still cannot type. The listening state and the missing text are the usual pair of symptoms.
The shortcut is not the one bound in the app, or another tool ate the keystroke.
Homebrew failed. Use the GitHub release.
You are dictating into an app that replaces the selection or rejects the input method. Try TextEdit or Notes as a control. If text appears there and not in the original app, the original app is the variable.
If it still fails, write to hello@sleekvoice.app or open an issue with the macOS version, the destination app, and which of the checks above you already did. “It doesn’t work” is harder to act on than “microphone is on, Accessibility is on, TextEdit receives text, Mail does not.”
Questions
Do I need a GPU or a specific Mac chip?
SleekVoice has not published hardware minimums beyond macOS. This post will not invent a chip requirement. If the app refuses to launch, the issue is the place to put the Mac model.
Can I point SleekVoice at my own Whisper weights?
Not as a documented setting. The app runs Whisper locally. It does not publish a model picker.
Is this the same as installing OpenAI’s Python package?
No. pip install openai-whisper is OpenAI’s command-line tool. SleekVoice is a Mac app with push-to-talk. You do not need the Python package to use the app.
- Whisper
- macOS
- privacy
Related reading
- Why local speech-to-text matters
What “local” changes about voice typing, what it does not change, and how SleekVoice’s claim differs from Wispr Flow’s cloud transcription.
- Introducing SleekVoice
SleekVoice is free, open-source voice typing for macOS. Whisper runs on your Mac, and the text lands in the app you are already using.