Speech recognition small enough to run on a laptop is now accurate enough for everyday dictation. If you have an Apple silicon Mac, you don’t need to send your voice to a server to get good results.
But “offline” and “private” get used loosely. Here’s how to tell what an app really does, and what else matters once you dictate every day.
What “offline” should mean
- Speech is recognised on your Mac. Not “encrypted in transit” or “deleted after processing”. The audio never leaves the machine.
- AI cleanup is local too. Some apps recognise speech locally but send the text to a cloud language model to add punctuation or rewrite it. That’s still your words leaving your Mac.
- It works with Wi-Fi off. This is the easiest way to check the first two yourself. Download the models, turn off Wi-Fi, and dictate.
- No account and no analytics. If an app needs you to sign in or reports usage, it’s talking to a server.
What matters once you dictate all day
- It types into any app. You want dictation where you already write: email, chat, documents, your browser, your code editor.
- Accuracy on your words. Names, jargon and product terms are where dictation fails most. Look for a custom vocabulary.
- Correcting by voice. If every mistake sends you back to the keyboard, you haven’t saved much.
- Formatting that fits the app. New lines that don’t send a chat message, no stray full stops on bullet points or ticket titles, capitals that follow the sentence you’re in.
- Low resource use. It shouldn’t slow your Mac down or spin the fans while it listens.
- A fair price. An app that runs on your own Mac costs its maker nothing per word you dictate.
How spellsafe answers each question
Where speech is recognised: on your Mac, with NVIDIA’s Parakeet model on the Neural Engine. Audio is never uploaded, and running on the Neural Engine leaves your GPU free.
Where AI cleanup runs: on your Mac too, with a local language model (Qwen3-4B). It removes fillers and self-corrections, and powers command mode: select text, hold Right Command and say “make this more formal” or “turn it into a list”.
Offline: yes, after a one-time model download (about 450 MB for speech, about 2.3 GB for the cleanup model).
Account and analytics: none. There’s nothing to sign up for, and the app sends no usage data or crash reports.
Any app: yes. Hold Right Option, talk, and let go, or tap it to see words as you speak. It types into whatever has focus.
Your words: add names and jargon to Vocabulary, with what they sound like, and they’re spelled your way from then on.
Correcting by voice: “scratch that”, “delete last word”, “replace Tuesday with Wednesday”, “select that”, “rewrite that more formally”.
Formatting: spacing and capitals follow the text already in the field, ticket titles and bullet points don’t get stray full stops, new lines in chat apps don’t send the message, and spoken numbers are typed as digits.
Beyond typing: “press enter”, “open Slack”, “switch to tab Gmail”, “new tab”, “scroll down”.
Seeing what it did: History keeps every dictation with its recording (you can turn recordings off), and Stats shows words dictated and time saved compared with typing.
Price: $9 a year, after a 7-day free trial with every feature.
Where spellsafe isn’t the right fit
- You need a language other than English. It’s English only today.
- You’re on an Intel Mac or older macOS. It needs Apple silicon and macOS 14 or later.
- You want to transcribe recordings like meetings or interviews. spellsafe types what you dictate live.
- You need to control every part of your Mac hands-free, including clicking arbitrary buttons and menus.
Try it
The best test of any offline claim is to try it with the internet off. spellsafe’s 7-day trial includes every feature. Download spellsafe.