#Use the one already installed first
Every Mac ships with keyboard dictation, and a surprising number of people shopping for an app have never switched it on. Open System Settings, go to Keyboard, then Dictation, and turn it on. Apple's default shortcut is pressing the Fn key twice, the key marked with a globe on current Mac keyboards. The same pop-up menu lets you pick a different shortcut, and Customize lets you invent one. Press Escape or the shortcut again to stop.
Two things make it more capable than its reputation. It types into any text field in any app, because it feeds the same input path as your keyboard, so Mail, Word, a browser form and a terminal all behave the same. And on Apple silicon you can keep typing while it listens, which means you can fix a word by hand mid-sentence without ending the session.
What it does not do is equally clear, and it is the whole reason the paid category exists. There is no transcript window, so the text goes straight into the document with no chance to review it first. It will not touch an audio file that already exists on disk. It has no vocabulary list for names your team says fifty times a day. And it gives you no choice of model, no speaker labels, no timestamps and no export.
#Where does your audio go?
That single question sorts the whole category, and it explains most of the price differences you see. There are three answers.
A model running on your Mac. The app downloads speech model weights once, usually a Whisper variant or NVIDIA's Parakeet, and runs them locally. VoiceInk is the clearest example: it is open source under the GPL, and its documentation describes whisper.cpp plus Parakeet as the local path, with Parakeet recommended for the speed and accuracy balance. Superwhisper does the same with its offline modes, which its site describes as downloading once and then running locally. Nothing goes over the network, there is no per-minute cost, and accuracy scales with model size, as does the wait after you stop talking.
A model on somebody's server. The app uploads your audio and gets text back. Wispr Flow works this way and prices accordingly: a free tier with weekly limits, then a subscription. Cloud recognition tends to hold up better on hard audio, noisy rooms, accents, crosstalk, and it does not care how old your Mac is. It also means the recording leaves your machine, and it stops working on a plane.
The system API. macOS 26 ships SpeechAnalyzer, Apple's on-device speech framework, so a third-party app can transcribe without bundling any model weights of its own. Nothing to download, the same on-device processing as Apple's own dictation, and no server. The catch is the floor: it exists only on macOS 26, so an app built on it will not run on Sequoia, and the machine has to be on the Tahoe supported list in the first place.
| App | Where audio is processed | Pricing model | Best at |
|---|---|---|---|
| Apple Dictation | On device (Apple silicon) | Included with macOS | Typing by voice into any app, zero setup |
| VoiceInk | Local Whisper or Parakeet models | One-time license, from $29 on its pricing page | Local dictation with open source code (GPL v3) |
| Superwhisper | Local models or cloud, your choice per mode | Free tier, subscription or a one-time license | Switching between offline and cloud accuracy |
| MacWhisper | Local, Whisper and Parakeet | Free download, Pro unlock (one-time or subscription) | Transcribing files, meetings and system audio |
| Wispr Flow | Cloud | Free tier with weekly limits, then Pro | Accuracy on messy audio, syncing across devices |
Each row is taken from that app's own site or documentation, checked September 2026. Prices and tiers in this category change often, so treat the model as the durable part and the number as a starting point.
The practical consequence is a privacy one. If you dictate patient notes, legal drafts, unreleased code or anything under an agreement you signed, the local and system-API routes are the only ones where the audio never leaves the Mac. If you dictate Slack messages, that constraint costs you accuracy for nothing.
#Dictation apps and transcription apps are different products
This is the mistake that wastes the most money in this category. Both are labelled speech to text, and they solve unrelated problems.
Dictation is live. You hold a key, talk, and the words appear at the cursor in the app you were already using. Latency is the only metric that matters, because a two second gap between finishing a sentence and seeing it breaks the writing rhythm entirely.
Transcription is batch. You already have a recording, an interview, a lecture, a meeting, and you want text out of it. MacWhisper's own site frames it this way: audio, video, meetings and system audio, transcribed locally. Here latency barely matters, and what you actually want is speaker labels, timestamps, a text editor synced to playback, and subtitle export.
Some apps do both, most do one well. Ask which of these two jobs is yours before looking at a single price: writing an email by talking, or turning a two hour recording into text before a deadline. Buying a file transcriber for the first job leaves you copying text out of a window, which is exactly the friction you were paying to remove.
#What decides whether it feels good
Feature lists rarely predict which app you keep. Five details do.
- How the text lands. An app either synthesises keystrokes or pastes. Pasting is instant even for a long paragraph, but it goes through the system clipboard, so whatever you had copied is gone unless the app restores it or you keep a clipboard history. Either method needs the Accessibility permission to write into another app.
- The key you hold. Push to talk, hold a key while speaking, cannot strand you recording. A toggle can. If you dictate in short bursts between typing, hold-to-talk is the difference between using an app daily and abandoning it.
- Model size against your patience. Local models trade accuracy for delay. A large Whisper model on a busy Mac can take longer to return a paragraph than it took to say it, and the fix is usually a smaller or faster model rather than a different app.
- Custom vocabulary. Proper nouns are where general models fail: colleague names, product names, internal acronyms. An app that lets you feed it a word list will beat a more accurate model that has never seen your jargon.
- The cleanup pass. Several apps run a language model over the raw transcript to strip filler, add punctuation and reformat. It reads better. It is also a second model, often the part that goes to a server in an otherwise local app, so check that separately if privacy is the reason you chose local.
One more that costs nothing: the microphone. A model can only work with what it is given, and a room with an echo or a fan will degrade any of these apps faster than the difference between two model sizes.
#Picking one, in order
- Turn on Apple Dictation and use it for a week. It is free and already there. Note the specific moments it fails you, because that list is your requirements document.
- Sort your job into live or batch. Dictating into apps, or transcribing files. The two shopping lists barely overlap.
- Decide whether the audio may leave the Mac. If the answer is no, you are choosing between local models and apps built on the system API, and the field is smaller and mostly one-time pricing.
- Check the macOS floor. Anything built on SpeechAnalyzer needs macOS 26. Anything running large local models wants Apple silicon and disk space for the weights.
- Trial it on your own worst audio. Not the marketing demo. Your accent, your room, your jargon, the meeting recording with three people talking over each other.
- Prefer one-time pricing unless the cloud is the point. Local models keep working with no account and no billing relationship. A subscription is defensible when servers are doing the work; it is harder to justify for a model running on hardware you already bought.
#Frequently asked questions
Does the Mac have built-in text to speech?
Yes, and it is the opposite direction from everything above. System Settings, Accessibility, Spoken Content turns on a key that reads selected text aloud, with a large catalogue of downloadable voices. Terminal users get the same engine from the say command, which can also write the speech to an audio file.
Is there a free text to speech app for Mac?
The built-in Spoken Content is the free one, and for most purposes it is also the best value, because the voices are system-wide and work offline. Paid text to speech apps mainly add exports, cloud voices and long-document handling; if all you need is a page read aloud, Apple's is already installed.
#What this means for the notch
Dictation is a feature with no home on a Mac. It starts with a keypress and reports back somewhere you have to go and look. The notch is one place to put it: NotchBay runs on-device dictation and transcription through SpeechAnalyzer, the macOS 26 API described above, so audio is processed on the Mac, and the transcript can be copied or inserted at the cursor. That is one feature of an island app rather than a dedicated speech to text tool, and it carries the same requirements: macOS 26 and a MacBook with a notch, with a floating pill at the top of the screen on Macs without one. $19 one-time.
