The things Sonari won't do, and why
5 min read
Sonari is a dictation app for the Mac. You press a shortcut, talk, and it pastes clean text into whatever app you were already in. All of it runs on your machine.
Before I tell you what it does, here’s what it won’t do.
English only. Apple Silicon only. macOS 14 and up, and the AI cleanup needs Apple Intelligence, which not every supported Mac has.
Every one of those is a no I chose, and they all point the same way. I want Sonari to do one thing, dictation on your Mac, and be really good at it. Everything I said no to is something that would have split the app’s attention.
Most product pages bury the limits three scrolls down, under the features, because a limit reads like an apology. I put them at the top. Each one has a real reason behind it, and the reasons are honestly the most interesting part of how the app is built.
English only
I wanted one model that’s genuinely good at English, and fast.
Multilingual is doable, and I’m not pretending it isn’t. Parakeet, the engine Sonari leans on, has a multilingual model. Sotto ships it and runs the whole thing locally, superwhisper does it in the cloud, and if you dictate in more than one language, go use one of them. They’ll serve you well and I’ll happily point you their way.
I went the other direction on purpose. Parakeet is state of the art for English, and I wanted Sonari to be really, really good at that one language. A model kept to English stays smaller and quicker and sharper on the words you actually say, and that accuracy is the whole reason to reach for it over the dictation already built into your Mac.
I’m not ruling multilingual out down the road. I honestly haven’t decided. For launch I wanted one job done really well before I ask the app to take on more.
The cost is simple. If you dictate in something other than English right now, this isn’t your app, and you should know that going in.
Apple Silicon only
This one isn’t a build flag I forgot to turn on.
The transcription runs on the Neural Engine, the piece of the M-series chips Apple built for exactly this kind of work. FluidAudio, the library doing the Parakeet part, targets the Neural Engine. Intel Macs can in principle fall back to slower CPU and GPU paths, but that path is untested, and honestly I’m not sure it would be fast enough to not embarrass me.
Speed is pretty much the whole feel of dictation. You stop talking, and the text should already be waiting in the app, not spinning while you sit there. On Apple Silicon that gap is short enough to forget it’s there. Any slower and you feel it, and a dictation tool you can feel is one you quietly stop opening. I’m not going to ship an Intel build that maybe works and feels sluggish, and then spend every support email explaining why.
So, Apple Silicon only. If you’re on an Intel Mac, I’m sorry, this isn’t for you yet.
macOS 14 and up, and the AI part needs Apple Intelligence
Two limits in one, and the second is the softest thing on this list.
The floor is macOS 14. That’s where the APIs I depend on are stable, and holding to it keeps me from writing fragile workarounds for three older versions nobody should still be running for a brand new paid app.
The AI transforms are the other half. Those are the bit that turns a raw transcript into a clean email, or a tidy list, or a message that reads like you meant to write it. They run on Apple’s on-device Foundation Models, which means they need Apple Intelligence, which means a newer OS and a capable chip. Not every Mac that can run Sonari can run those.
Here’s the part I actually care about. When Apple Intelligence isn’t there, Sonari pastes your transcript exactly as it heard it. It does not phone home to make up the difference. The cleanup sits on top of dictation that already works on its own. Without it you get a nicer paragraph a little less often. Your text still lands, and it still stays on your Mac.
Why the noes come first
I’ve shipped enough software to know the drawback in your own work is the thing you’re tempted to leave off the page. This is me leaving it on.
Every limit up there is a focus decision. English so the model stays fast and sharp. Apple Silicon so the text lands the instant you stop talking. A recent OS so I’m not patching around three old versions of macOS. No cloud, ever, so your voice has nowhere to go but the app on your desk. Four noes, and every one of them is what lets a single person build something that’s actually good at the one job in front of it.
None of this is the full feature list. I’ll show you that part when there’s a button to press.
Sonari isn’t for sale yet, but it’s close. I’m about to put the first builds in a handful of hands, and the folks on the list are first in line for that and for launch right behind it. If a local, English, Apple-Silicon dictation app that keeps your voice on your own machine sounds like your kind of thing, drop your email at sonari.audio. I’ll be in touch soon.
TJ