Buying guideTwelve things worth checking before you pick one.

A framework rather than a ranking — we haven't tested every alternative, so we're not going to rank them. Each check takes a minute, on any product.

Is there a best AI voice keyboard?

There is no single best AI voice keyboard, because the trade-offs are real ones: cloud accuracy against offline reliability, a polished result against an exact transcript, breadth of app support against the permissions that requires. What there is, is a set of 12 checks that separate a tool which saves you time from one that moves your editing pass somewhere else. Each check below takes about a minute to run on any product. WisprBee's answer to each is stated plainly — including the one where we would point you elsewhere, and the 3 that come with a real concession.

01

Do you get a transcript, or writing you can actually send?

This is the difference that decides whether the tool saves you time. A transcript still needs filler removed, punctuation added and paragraphs made before it can go anywhere — and that editing pass is most of the work you were trying to avoid.

Test it yourself: Dictate a rambling three-sentence thought with an “um” and a change of mind in it. Look at what appears. If you'd have to touch it before sending, it's a transcriber.
WisprBee

Finished writing, at four cleanup levels ranging from an exact transcript to a full rewrite. Filler goes, punctuation appears, and a spoken change of mind resolves to what you meant.

02

Does the output change depending on where it's going?

A Slack reply and a client email are different kinds of writing. A tool that produces one flat register makes you reshape everything by hand, which is the same editing pass in a different disguise.

Test it yourself: Dictate the same sentence into a chat app and then into an email client. If the two outputs are identical, the tool doesn't know where it is.
WisprBee

Eight destination profiles, chosen automatically from the frontmost application and, in a browser, the website. There is no mode to pick — picking one is the cost the feature exists to remove.

03

Can you teach it your vocabulary — and does it learn on its own?

Proper nouns are where general-purpose speech models fail, because they have no way to guess how you spell a colleague's name or an internal product. And a word list you have to maintain by hand tends not to get maintained.

Test it yourself: Dictate five names and terms specific to your work. Count the errors. Then check whether correcting one teaches the tool anything, or whether you'll be correcting it again tomorrow.
WisprBee — with a concession

A Personal Dictionary that corrects by sound rather than spelling, so a term is fixed even when the transcript came back as a similar-sounding word. Fifty terms free, unlimited on paid plans. Learning from your corrections automatically is Pro and above — on the free plan the list is manual.

04

Can you change text by voice, or only insert new text?

Most writing is revision. A tool that can only append leaves you reaching for the keyboard the moment something needs fixing, which breaks the one thing you installed it for.

Test it yourself: Select a sentence you've already written and try to replace it by speaking. Check whether what comes back is a complete rewrite or an abbreviation of what you said.
WisprBee

Command mode: select, hold a second key, speak the replacement. It works in any editable field and returns a full rewrite rather than a summary. Included on every plan.

05

Where does your audio go, and is any of it kept?

You will dictate things you would not post publicly. Whether that audio leaves the machine, who processes it, how long it is held and whether it trains a model are four separate questions, and vendors often answer only the flattering one.

Test it yourself: Find the sub-processor list. If the privacy policy names no specific companies, the answer is unknown rather than reassuring.
WisprBee

Cloud processing over an encrypted connection, through four named sub-processors: Paddle for payments, OpenRouter as the gateway, OpenAI and Google for the models. Audio is not retained after processing and is not used for training without permission. History, dictionary, snippets and keys stay on your Mac unless you enable sync.

06

What happens when you have no connection?

Cloud models are more accurate; on-device models keep working on a plane. Which trade-off is right depends on you, but you should know which one you're buying rather than discovering it at 30,000 feet.

Test it yourself: Turn off Wi-Fi and dictate. Note whether it fails, degrades, or performs identically.
WisprBee — we'd point you elsewhere on this one

It degrades rather than failing: an on-device model takes over transcription. Expect lower accuracy, and plainer output, because the polish step is a cloud step. If you work offline most of the time, a fully on-device tool will serve you better.

07

How many languages, and can the output differ from the input?

Two different features get sold as one. Dictating in your own language is common; speaking one language and writing another is rarer and is the thing that matters if you work across them.

Test it yourself: Check the count, then check whether the input and output languages are configured separately. If there's one setting, there's no translation.
WisprBee

75 spoken base languages, auto-detected or set manually. The output language is configured separately, English by default — so speaking Spanish and sending English is one step, not two.

08

Does it work in every app, or only in supported ones?

A tool that lives in a browser extension or a handful of integrations will fail in the app you need it in most, usually a terminal, a native client or something niche to your industry.

Test it yourself: Try it somewhere unglamorous: a rename field in Finder, a search box, a terminal.
WisprBee — with a concession

It runs at the OS level, so it works in any editable field on macOS with no plugins or per-app setup. The cost is that macOS requires the Accessibility permission before any app can type into another — that grant is unavoidable for anything working this way.

09

Does it handle meetings — and does something join the call?

Dictation and meeting capture are usually two subscriptions. Where one tool does both, the question becomes how it captures: a bot in the participant list is visible to everyone, needs admitting, and only supports the platforms it has integrated with.

Test it yourself: Start a call and begin capture. Look at the participant list. Then try it on a platform the vendor doesn't advertise.
WisprBee — with a concession

The built-in Notetaker captures microphone and system audio from the Mac, so every voice on the call is transcribed and nothing appears in the attendee list. It has no platform integrations and is therefore limited by none. Included on every plan. Whether it produces summaries or speaker labels isn't documented yet.

10

If it gets something wrong, can you get your words back?

Speech is a one-shot. A tool that discards the original transcript after processing means any bad cleanup costs you the thought, and you get to say it all again.

Test it yourself: Dictate something long, let it mangle it, then look for the original. If there's no history and no undo, the answer is no.
WisprBee

The raw transcript is written to local history before any cleanup runs, which is what makes undo, re-transcribe and a diff view possible. Thirty days of history free; kept indefinitely on paid plans.

11

What does it do to your clipboard?

Small, and infuriating. Many tools paste by taking over the clipboard and leaving their output in it, so whatever you had copied is gone — usually noticed at the worst moment.

Test it yourself: Copy something you care about. Dictate. Paste. Check what comes out.
WisprBee

The clipboard is borrowed to paste and then restored to exactly what was in it.

12

What does it cost at the volume you'll actually use?

Free tiers are sized to be pleasant in week one. The number that matters is the one at your real usage, and whether the features you were sold are on the tier you can afford.

Test it yourself: Estimate your words per month honestly, then find which tier that lands on and what it withholds. Check the refund terms before paying, not after.
WisprBee

Free is 2,000 words a month. Pro is $12/month for unlimited words; Pro Plus is $20 per user per month for teams. Only two things are actually gated — the word allowance and team ownership — so the free tier is the whole product at a smaller volume, apart from the Adaptive Dictionary. Paddle handles billing, with 14 days to change your mind.

When WisprBee is the wrong answer

6 situations where we'd tell you to buy something else. A guide whose every criterion happens to favour the author isn't a guide — which is why 4 of the 12 checks above are marked against us too.

  • Your audio may never leave the machine. Transcription is a cloud step by default, and that is where the quality comes from — a fully on-device tool is the honest answer here.
  • You work offline most of the time. The fallback keeps you going; it isn't what you'd choose to rely on.
  • You're on Windows, Linux, or an Intel Mac. There's no build. Apple Silicon and macOS 11.0 or later only.
  • You need a compliance certificate today. SOC 2 is planned and HIPAA isn't available, so regulated work should wait.
  • You need a notetaker to attend meetings you're not on. Capture happens on your Mac, so you have to be in the room.
  • You want a transcript, exactly as spoken, and nothing else. Set the cleanup level to exact and it will do that — but you'd be paying for a polish step you're switching off.

Checking WisprBee against this list

Or just run the twelve checks on it. The free plan needs no card, and every test above works on it.

Download free

Put the keyboard down.

Hold a key, say what you mean, and let WisprBee write it right — for wherever it's going. Your fastest, cleanest writing is one keystroke away.

Free to start · macOS · up to 4× faster than typing