Building private audio applications with VoxRT

Your voice is one of the most private forms of data you have, so why would you want your audio to ever leave your device?

• The idea

It’s not complicated. I hit a shortcut, talk and when I’m done the transcript gets dropped into a plain text file next to the original audio. That’s the whole loop. The text is searchable, the audio’s there if I ever want to hear the original tone of something, and none of it leaves the machine.

The part that made this click for me is that the transcription is genuinely local. VoxRT is an on-device speech runtime — the model runs on my hardware, not on a server I’m renting time from. No connection required, nothing uploaded “to improve the service.” For voice notes specifically that matters, because the whole point of a voice note is that it’s the unfiltered version of a thought. That’s the stuff I least want sitting on someone else’s disk.

• Why local, honestly

Two reasons, roughly in this order.

The first is privacy, but not in a paranoid way — more that I just don’t want to think about it. If the audio physically never leaves the laptop, I don’t have to read a privacy policy, I don’t have to trust a retention promise, and I don’t have to keep a mental list of “things I shouldn’t say near the mic.” The architecture draws the line for me.

The second, which surprised me, is that local turned out to be nicer to use, not just safer. There’s no round-trip. I talk, the text appears, and there’s no little spinner while a server somewhere thinks about it. It also works with the wifi off, which means it works on a train, on a plane, in the basement — anywhere I actually get ideas, which is rarely at my desk with perfect signal.

• What’s actually doing the work

Under the hood, VoxRT is a Rust runtime built specifically for running speech models efficiently on whatever platform you’re on. Instead of leaning on a general framework or a cloud endpoint, it runs the model directly through optimized execution paths — the kind of low-level stuff (ARMv8 NEON on ARM chips, WebAssembly SIMD128 in the browser) that’s the difference between “technically runs locally” and “runs locally fast enough that you forget it’s local.”

The model I use is their streaming FastConformer one. Two things about it made it a keeper for notes:

• It’s built for real-time recognition with an 80 ms lookahead, so the text keeps pace with me while I talk instead of arriving in a lump at the end.
• It does punctuation and capitalization on its own. This sounds minor until you’ve read back a wall of lowercase transcript with no sentence breaks. Having it come out as actual readable prose is most of what makes the notes worth keeping.

For a sense of how light it is: VoxRT’s own benchmarks put the FastConformer streaming model at a real-time factor of around 0.30 on file replay and 0.35 on live mic input on a Snapdragon 662 — a mid-range phone chip. Real-time factor under 1.0 just means it transcribes faster than the audio plays. If it clears that bar comfortably on a modest phone, a laptop has plenty of room, which matches my experience: I’ve never once felt like I was waiting on it.

• Would I recommend it?

If you’re the kind of person who thinks by talking and wants to keep those thoughts as text without feeding them to a cloud, then yeah — this is a good way to do it. It’s not a polished consumer app you download and it just works; you’re wiring up your own little pipeline. But that’s also the appeal, because you decide exactly where the audio goes (in my case: nowhere).

The projects and SDKs are on GitHub if you want to poke at it yourself — all of it’s free, and I’m not paying VoxRT anything for what I’m doing. If you get deeper into it and hit integration questions or want custom models, the VoxRT team is at voxrt.com or help@voxrt.com.

For me it’s just become the boring but reliable thing in the background — I talk, the words show up, they stay on my machine. That’s all I really need.