Speech to text on Mac: built-in dictation vs. the alternatives

Your Mac already does speech to text. It’s free, it’s built in, and it works on-device. Most people who find this page either don’t know that or have tried it and want something better.

So this covers both: exactly how to turn on and use macOS dictation, and then — honestly — the four things it does badly and what to use instead when those things matter.

How to turn on speech to text on Mac

1. Open the Apple menu → System Settings
2. Click Keyboard in the sidebar (scroll if you don’t see it)
3. Find the Dictation section and toggle it on
4. Click Enable when prompted
5. Choose whether to share audio recordings with Apple — you can decline, and dictation still works

That’s it. No download, no account, no subscription.

How to use dictation on your MacBook

1. Put your cursor where you want the text to appear — any app, any text field
2. Press the Microphone key if your keyboard has one, use your keyboard shortcut, or choose Edit → Start Dictation
3. Wait for the cursor to highlight and pulse, then start talking
4. Press Escape or the Microphone key again to stop

The cursor cue matters. Start talking before it pulses and you’ll lose the first few words, which is the single most common complaint about Mac dictation and is usually just timing.

Setting a keyboard shortcut you’ll actually use

The default shortcut options live in the Dictation shortcut pop-up menu in the same settings pane. If none of them suit you, choose Customize and press whatever combination you want — Option-Z, for instance. Pick something you can hit one-handed without looking, because you’ll use it dozens of times a day or not at all.

Dictating punctuation and formatting

macOS dictation won’t guess at punctuation reliably, so you speak it:

• “period”, “comma”, “question mark”, “exclamation mark”
• “new line”, “new paragraph”
• “open quote” / “close quote”
• “cap” before a word to capitalize it

This is the part that breaks the flow for most people. You end up composing two streams at once — the sentence and its punctuation — which is a different cognitive task from talking.

Changing the dictation language

In the Dictation settings, click Edit next to Languages and pick your language and region. You can add several and switch mid-session by clicking the language abbreviation or pressing the Globe key.

Note the mechanism: you switch. You choose a language and dictate in that language until you choose another one. Hold that thought, because it’s one of the four limits below.

Is Mac dictation processed on-device?

Mostly, yes — Apple states that voice inputs and transcripts for general text dictation are processed on your device and not sent to Siri servers, depending on your settings. Apple’s “About Ask Siri, Dictation & Privacy” page has the specifics for your configuration.

This is genuinely the strongest argument for the built-in option. If your constraint is that audio cannot leave the machine, macOS dictation clears that bar at zero cost, and most third-party apps don’t.

Four things macOS dictation does badly

It’s a competent transcriber. It is not a writing tool, and the difference shows up in four specific places.

1. No cleanup pass

macOS dictation is verbatim by design. Say “um so can we uh move the the meeting to thursday i think friday works better actually” and that is very close to what you’ll get — filler, stutter, and a self-correction you have to resolve by hand.

What you meant was: “Can we move the meeting to Friday? That works better.”

Getting from the first to the second is editing work, and for short messages it frequently costs more time than typing would have. This is the gap that every serious alternative is built to close, and it’s a bigger deal in daily use than any accuracy percentage.

2. No custom vocabulary

Your colleagues’ names, your product names, your internal jargon, your industry’s acronyms — macOS dictation has no mechanism for teaching it any of them. It will guess, and it will guess the same thing wrong every time.

If your work involves a lot of proper nouns, this alone is disqualifying. Modern streaming speech-to-text models accept key terms up front and bias spelling toward them, which turns a recurring irritation into a one-time setup step.

3. Weaker on accents and noisy rooms

On-device models are constrained by what fits and runs on a laptop. That’s a real engineering tradeoff and Apple has done well within it, but the ceiling is lower than a hosted model’s. If you speak with an accent the model saw less of in training, or you dictate in a café, or you’re on a bad connection to a headset, you’ll notice.

The gap is measurable. On short-form English — which is what dictation actually is, a sentence or two at a time — the current generation of hosted models is well under 4% word error rate. Universal-3.5 Pro Realtime posts a 3.87% mean across Common Voice and both LibriSpeech test sets, ahead of Speechmatics at 4.69% and Deepgram Flux at 6.53%.

4. No mid-sentence language switching

You can switch languages in macOS dictation, but you switch between utterances. If you’re one of the many people who move between languages inside a single sentence — Spanglish, Hinglish, or just dropping a technical term in English into an otherwise German sentence — the built-in tool has no answer.

Models with native code-switching handle this without configuration. It’s a small feature with an outsized effect on whether dictation feels usable if you’re bilingual.

What to use instead

If you want finished text, not a transcript: Blurt

Blurt is a free, MIT-licensed macOS dictation app. Hold the right ⌘ key, talk, and finished text lands in whatever app has focus — filler removed, punctuation applied, self-corrections resolved to what you actually landed on.

It closes all four gaps above. Transcription and an LLM cleanup pass run in a single call to the Dictation API, around 134ms p50 on transcription with cleanup typically inside a second. Key terms in Settings bias spelling toward names and jargon. Nineteen languages with mid-sentence code-switching, so the Spanglish problem goes away. And on short-form English it’s running the model that posts that 3.87% figure.

The tradeoffs, stated plainly. Audio goes to the cloud — if macOS dictation’s on-device processing is precisely why you’re using it, this is a step backward and you should use one of the local options below. It’s Apple Silicon only. And while the app is free and open source, it bills through your own AssemblyAI API key at roughly $0.62 per hour of audio, with free credits to start.

Because it’s MIT-licensed, the privacy story is checkable rather than promised: API key in the macOS Keychain, no audio stored, no transcripts stored, no telemetry, and it explicitly refuses to read password fields or send your app name, window title, or selection.

If audio must stay on your machine: Handy or FluidVoice

Handy is MIT-licensed, free, and fully offline. It runs Whisper and Parakeet locally on macOS, Windows, and Linux, with a configurable global hotkey and push-to-talk. Better models than Apple’s on-device engine, same privacy guarantee, no cost.

FluidVoice is GPLv3, free, and local-first, with on-device formatting and capitalization plus a Command Mode for controlling the Mac by voice. macOS 15 and later.

Neither gives you a full LLM cleanup pass. You’re trading finished text for locality, which is the correct trade for a lot of people.

If you want one app that does everything: Superwhisper or MacWhisper

Superwhisper runs local Whisper models offline on Apple Silicon plus cloud options, with LLM post-processing and predefined Modes for email, messages, and formal writing. Mac, Windows, iOS, and Android.

MacWhisper is aimed more at transcribing files and meetings than at system-wide dictation, though it does both. Free tier available; Pro is €64 as a one-time payment with lifetime updates.

If you’re building this into your own product

Every app above is a hotkey, a paste mechanism, and a model call. If you’re a developer adding voice input to your own software, skip the app layer — the Dictation API is one HTTP POST that returns cleaned-up text and the verbatim transcript together, at $0.62 per hour with keyterm prompting and output instructions included.

Quick comparison

How to decide in thirty seconds

Keep macOS dictation if you dictate occasionally, your audio can’t leave the machine, and you don’t mind speaking punctuation. It’s free and it’s fine.

Switch to a local app — Handy or FluidVoice — if you like the privacy but want a better model and a real hotkey.

Switch to a cloud app — Blurt or Superwhisper — if you dictate all day, you’re tired of editing your own filler words, and you’d rather have finished text than a transcript.

Build on an API if you’re putting voice input inside a product rather than choosing something to install.

The reason built-in dictation never got better

macOS dictation hasn’t fundamentally changed in years, and it’s tempting to read that as neglect. It isn’t. It’s a constraint.

Apple committed to on-device processing for general text dictation, which is a real privacy guarantee and one worth having. But it also caps what’s possible — the model has to fit on a laptop, run without draining the battery, and work for every Mac Apple still supports. Under those constraints, verbatim transcription is close to the right answer. A cleanup pass means running a language model too, and that’s a much larger thing to ask of a laptop that’s also running Xcode.

So the split in this category isn’t really Apple versus everyone else. It’s a question about where the work happens, and every option here is just a different answer to it. The local projects are betting that on-device models get good enough to close the gap. The cloud ones are betting that most people, for most writing, will trade locality for text they don’t have to edit.

Both bets look reasonable right now. Which is why you should pick based on your own constraint rather than anyone’s recommendation — including this one.

Frequently asked questions

How do I turn on speech-to-text on Mac?

Open the Apple menu, go to System Settings, click Keyboard in the sidebar, find the Dictation section, toggle it on, and click Enable. You’ll be asked whether to share audio recordings with Apple, and you can decline without losing the feature. Once enabled, press the Microphone key or your keyboard shortcut anywhere you can type.

How do I dictate text on my MacBook?

Place your cursor in any text field, press the Microphone key or choose Edit → Start Dictation, wait for the cursor to highlight and pulse, then speak. Say punctuation aloud — “period,” “comma,” “new paragraph” — because macOS dictation won’t add it reliably on its own. Press Escape or the Microphone key again to stop.

Does Mac dictation work offline?

Yes, for general text dictation. Apple states that voice inputs and transcripts for general text dictation are processed on your device and not sent to Siri servers, depending on your settings. That on-device processing is the built-in feature’s strongest advantage over most third-party dictation apps, which send audio to a server.

What is the best free dictation app for Mac?

Handy is the strongest fully free option — MIT-licensed, runs Whisper and Parakeet entirely on your Mac, and costs nothing to install or run. Blurt is free and open source but bills through your own API key at roughly $0.62 per hour of audio. Both use better speech models than macOS dictation’s on-device engine.

Why is Mac dictation so inaccurate with names and technical terms?

Because it has no custom vocabulary feature — there’s no way to tell it how your colleagues’ names or your product names are spelled, so it guesses, and it guesses the same thing wrong consistently. Speech-to-text models that accept key terms up front bias their output toward those spellings and largely eliminate the problem. This is usually the single biggest accuracy complaint from professional users.

Can I use voice typing on Mac in more than one language at once?

Not with the built-in feature. macOS dictation lets you add several languages and switch between them with the Globe key, but it handles one language at a time and can’t follow a switch inside a single sentence. Models with native code-switching — Blurt covers 19 languages this way — transcribe each word in the language it was spoken without any switching step.