Five questions, and the shortlist picks itself

Last updated 11 September 2026

Question 1: Must the audio stay on your machine?

Not would you prefer it — must it, because of what you dictate or who you answer to. Privileged legal material, patient notes, source protection, unannounced commercial information, anything under an NDA.

If yes: every cloud product is out, regardless of how good it is. This is the only question that eliminates on an absolute basis, which is why it goes first.

If no: the whole market is open, and the cloud products are genuinely strong — bigger models, more languages, more platforms.

Question 2: Do you dictate on anything other than a Mac?

A phone in daily use, or a work PC. Be honest — not "occasionally", but whether you would be annoyed if it did not work there.

If yes: most local Mac apps are out, because they are Mac-only. This is the single most common legitimate reason not to buy a local app, and it is worth discovering now.

If yes and question 1 was also yes: your options are narrow. Look at the local apps with more than one platform, and accept that the phone may have to use something else.

Question 3: Do you correct the same words every day?

Keep a note for a week. If the list is short and varied, general accuracy is fine and you have no vocabulary problem. If it is the same fifteen proper nouns, you do.

If no: macOS's built-in dictation is free, already installed, and probably enough. Stop here and keep your money.

If yes: you need custom vocabulary, which the built-in one does not have. This is the question that most often justifies a purchase.

Question 4: Do you work without a connection?

Flights, trains, client sites, basements, hotels with hostile Wi-Fi. Regularly, not once a year.

If yes: cloud products are unusable at exactly the moments you wanted them. Local, and check that the tidy-up step is local too — a lot of apps recognise offline and then fail on the rewrite.

Question 5: Do you want to stop paying at some point?

A preference rather than a requirement, and it narrows things usefully.

If yes: look at one-off licences, which exist only in the local half of the market. Cloud products cannot reasonably offer them because inference costs them money on every dictation.

What the answers produce

Questions that seem important and are not

• "Which is most accurate?" Unanswerable for your voice without a published method, and dominated by your microphone, your room and your vocabulary rather than the app.
• "Which model does it use?" An implementation detail that will have changed by next year. Judge the app on text insertion, vocabulary and memory management.
• "How many features does it have?" Dictation apps fail on reliability, not on feature count.
• "Which is most popular?" Popularity tracks marketing budget, and the biggest budgets in this category belong to cloud products.

Then test properly

1. Install two, with different shortcuts Free tiers make this cost nothing.
2. Teach each one the same fifteen words Untaught, they will all look equally mediocre and you will learn nothing.
3. Dictate the same real paragraph into both From your actual work, with your jargon in it. Not a test sentence.
4. Count what you had to correct That number is the comparison. Everything else is marketing.
5. Run the aeroplane-mode test if privacy was a yes Verify rather than trust a new vendor's marketing page.