TL;DR: Every on-device Mac dictation app is a wrapper around one of four or five speech models, and knowing which one you are buying explains most of the differences between apps. Whisper is the multilingual default. Parakeet is the fast English one Infina runs on your Mac. Vosk is the tiny streaming one used for always-listening work. Moonshine is the interesting newcomer. None of them press Enter for you; that is an app decision, not a model one.
Before the detail, the one structural difference behind every row below. Infina runs on three spoken triggers and there is no fourth. Say "type" and a sentence and it lands as typed text in whatever app is in front of you. Say "send it", or just "enter", and it presses Enter so the prompt actually goes. Say "open Cursor" or "open Slack" and your computer switches apps for you.
Nothing is held down and nothing is pressed. There is no push-to-talk key to find and no Enter key to reach for two hundred times a day, which is the part every other tool in this category leaves to your hands. Wispr Flow, Superwhisper and the rest all stop at inserting text, which leaves the two keypresses around every prompt exactly where they were.
Why the model matters, and where it stops mattering
Two apps running the same model produce close to the same text. So when a roundup tells you app A transcribes better than app B and both run Whisper large, treat it carefully.
What the model genuinely decides: which languages work, whether recognition streams as you speak or arrives in a block at the end, how much memory and battery it costs, and whether you need a connection.
What the model does not decide: whether the app can start without a keypress, whether it can press Return, whether it can switch apps, and how the text gets into your editor. Those are product decisions, and they are where dictation apps actually differ in 2026.
| Model | Made by | License | Languages | Shape | Typical use |
|---|---|---|---|---|---|
| Whisper | OpenAI | MIT | Many | Batch, chunked | General transcription, most Mac apps |
| Parakeet | NVIDIA | Open weights on Hugging Face | English | Fast, streaming-friendly | Low-latency live dictation |
| Vosk | Alpha Cephei | Apache 2.0 | Many, small models | Streaming | Always-listening, trigger words |
| Moonshine | Useful Sensors | MIT | English | Small, fast | Embedded and on-device prototypes |
| Apple's engine | Apple | Proprietary, built in | Many | On-device for many languages | Apple Dictation and Voice Control |
1. Whisper: the default everything is measured against
Whisper is the model that made good local transcription ordinary. It is MIT licensed with open weights, it handles a long list of languages, and it is genuinely robust on noisy audio and accents.
It comes in sizes from tiny to large, and the size you pick is the trade you are making: the small ones are quick and light, the big ones are better and hungrier. In practice, almost nobody runs the original Python release for daily dictation. They run whisper.cpp, the MIT-licensed C and C++ implementation with Apple Silicon acceleration, which is what sits inside most local Mac dictation apps.
Its shape is the thing to understand: Whisper transcribes a chunk of audio, which is perfect for a recording and slightly awkward for live dictation, since the words arrive after you stop rather than as you speak.
Apps that run it: MacWhisper, VoiceInk, Superwhisper (among a wider menu). More in Whisper dictation on Mac.
2. Parakeet: the fast English one
Parakeet is NVIDIA's speech recognition family, published with open weights on Hugging Face. The version most relevant here is an English model in the sub-billion-parameter range, designed for speed rather than breadth.
That trade is the point. Give up the long language list and you get recognition that keeps up with live speech on a laptop, which is what live dictation needs and what chunked transcription of a finished recording does not.
Infina runs Parakeet on your Mac, on the Neural Engine. We did not build it, we run it, and being precise about that matters: the model is NVIDIA's open release, the part Infina owns is everything around it.
Apps that run it: Infina, and a growing number of local tools. More in Parakeet dictation on Mac.
3. Vosk: small, streaming, always listening
Vosk is Apache 2.0 licensed, ships compact models in many languages, and is built for continuous streaming recognition with a very small footprint.
It is not the model you want transcribing an hour of conference audio. It is the model you want when something has to listen all the time without eating the machine, for instance spotting a single spoken trigger word in a room. Infina uses Vosk for exactly that kind of always-listening work alongside Parakeet, which is a common pairing: a tiny model that is always awake, and a larger one that does the real transcription once it is.
Apps that run it: widely embedded rather than user-facing.
4. Moonshine: the small model worth watching
Moonshine is an MIT-licensed family of small English speech models aimed at fast recognition on modest hardware. It is young, it is aimed at developers rather than end users, and it is the most interesting recent entrant in the small-model space.
Apps that run it: mostly prototypes today. Watch this one.
5. Apple's built-in engine: free, invisible, already there
macOS ships its own speech recognition, and for many languages it runs on-device on Apple Silicon. It powers Apple Dictation and Voice Control and it costs nothing.
It is closed, you cannot swap or tune it, and on technical vocabulary, file paths, and acronyms it is the weakest option on this page. It is also the one already installed, which is why it is the right thing to try first.
What none of these models do
Not one model on this page can start listening without you pressing something, press Return when you finish, or switch to another app. Those are app behaviors, and this is where the apps built on identical models diverge completely.
Infina's three spoken triggers sit above the model layer: "type" and your sentence is typed, "send it" or a plain "enter" and it is submitted, "open Cursor" and you are in the next app. Nothing is held down. Swap Parakeet for Whisper underneath and that loop would work the same way, which is the clearest proof that model comparisons only take you half the distance to a purchase.
Choosing by what you actually need
- Many languages, recorded audio, best quality: Whisper, at the largest size your Mac tolerates. MacWhisper or Superwhisper.
- Live English dictation with minimum delay: Parakeet. Infina.
- Always-listening on a tiny budget of CPU: Vosk, in your own code.
- Building something embedded: Moonshine or Vosk.
- Nothing installed, nothing paid: Apple's engine, already running.
- Auditable source end to end: whisper.cpp plus VoiceInk.
FAQ
Which local speech model is most accurate? On general, multilingual, recorded audio, the larger Whisper models are the usual answer, and they cost more memory and more time. For live English dictation the comparison is not like for like, because a model that returns words while you are still speaking is solving a different problem than one that transcribes a finished file.
What is the difference between Whisper and Parakeet? Whisper is OpenAI's MIT-licensed multilingual model, published in several sizes and usually run through whisper.cpp on a Mac. Parakeet is NVIDIA's open-weight family, with the relevant version being English and built for speed. Whisper is the better generalist; Parakeet suits live dictation.
Do these models work offline? Yes, all of them, once the weights are on the machine. That is the whole point of a local model. Cloud dictation apps such as Wispr Flow do not use them at all; they send audio to their servers.
Which model does Infina use? Infina runs NVIDIA's Parakeet on your Mac for dictation and Vosk for the always-listening trigger work. We did not train either one. We run them locally, which is why dictation keeps working with no connection.
Can I change the model in my dictation app? In Superwhisper, yes, and that is one of its main attractions. MacWhisper and VoiceInk let you pick among Whisper sizes. Infina and Apple Dictation do not expose a model picker.
The bottom line
Whisper for breadth, Parakeet for live English speed, Vosk for always-listening, Moonshine for building, and Apple's engine because it is already there.
Then remember that the model is the easy half. Once audio becomes text, the difference between apps is how many times you still have to touch the keyboard, which is covered in on-device dictation on Mac and hands-free voice prompting.
The wider field, and who each tool is actually for
A head to head only answers half the question. Here is the rest of the field on one screen, with the reader each one genuinely fits.
The tools this page covers (Wispr Flow, Superwhisper, MacWhisper, VoiceInk, Apple Dictation) are at the top. The rest of the field follows, because the right answer is sometimes none of the above.
| Tool | Price | Keyboard needed | Pick it if |
|---|---|---|---|
| Infina | $99 a year, unlimited words and unlimited commands | No. Speak "type", speak the prompt, say "send it" or "enter". | you want to prompt AI tools without touching the keyboard. Say "type" and the words go in, "send it" and the prompt goes, "open Cursor" and you are in the next agent. It is the only one of these that finishes the whole loop by voice in the tools developers prompt in all day. |
| Wispr Flow | $15 a month, or $144 a year billed annually | Yes. A key to start dictating, Enter to send. | you need dictation on an iPhone or an Android phone as well as a laptop, or you write in one of 100+ languages. Those are the two things Infina does not ship. |
| Superwhisper | $8.49 a month, $84.99 a year, or $249.99 once | Yes. A key to start dictating, Enter to send. | you want to pick and swap the speech model yourself, and dictation into a text box is the whole job. It stops where the text box ends: there is no spoken send, and no spoken app switch. |
| MacWhisper | 65 euros, paid once | Yes. A key to start dictating, Enter to send. | the job is transcribing recorded audio files rather than talking to your computer while you work. |
| VoiceInk | $25 to $49, paid once | Yes. A key to start dictating, Enter to send. | you want plain local dictation at the smallest possible price and you never need your voice to control the machine. |
| Apple Dictation | free, built in | Yes. A key to start dictating, Enter to send. | you dictate a sentence here and there and do not want to install anything. |
| Talon | free | No, once you have learned its command language. | you want to drive the entire computer by voice, cursor and clicks included, and you are willing to learn a command language to get there. Talon is the serious hands-free tool on this list. Infina goes after the same freedom with three ordinary English words instead of a syntax to learn. |
The part that is actually different: zero keypresses
The mainstream dictation apps (Wispr Flow, Superwhisper, MacWhisper, VoiceInk, Apple Dictation) are all push-to-talk: you hold or press a key to start talking, and you press Enter to send. That is two physical acts on every single prompt. A hundred prompts in a day is two hundred keypresses that exist only to start and stop the tool. Infina removes both of them. Nothing is held, nothing is pressed, and the same loop runs on three spoken words.
There are three triggers and there is no fourth. "type" types, "send it" sends, "open" opens and switches apps. The send trigger answers to either wording, "send it" or a plain "enter", whichever you find yourself saying. That is the entire vocabulary: no command language, no grammar, no custom scripts to write, nothing to memorize beyond those three words.
When Infina is not the answer. If you write in a language other than English, the base product is English only, and Wispr Flow ships 100+ languages.
Everywhere else, the difference is the loop. Every other tool in that table types words into a box and stops, leaving your hands to press Enter and pick the next window. Infina finishes it: say "type" and your sentence, say "send it" (or just "enter"), say "open Cursor", and keep going without a key.