Local or cloud? What really happens to your text when you dictate
Updated
Nearly every dictation app markets itself as “private” and “secure” today. Yet they work in fundamentally different ways: some send your voice to servers and get text back, others compute everything on your device. Both are legitimate — but you should know which model you are using before you dictate your contracts, emails or client notes.
How cloud dictation works
In the cloud model the app records your voice, ships it (or an audio stream) to data centres, a large model transcribes and polishes the text there, and the result comes back into your text field. That has real advantages: server models can be huge, often support 100+ languages, and run smoothly even on weak hardware.
The price: your spoken content leaves your device — with every single dictation. Watch the wording here: “stored on your device” is not the same as “computed on your device”. Some apps sound on-device on their homepage while their privacy policy lists cloud and AI-infrastructure vendors as processors — storage is local, computation is cloud. That is not a scandal, but it is a difference worth knowing. And “zero data retention” means “we don’t keep it” — not “it was never there”.
What “100% local” technically means
Local dictation used to be a compromise: small models, middling quality. Apple Silicon changed that. Modern Macs have enough GPU and Neural Engine power to run serious speech-recognition and language models directly on the device — in real time.
Concretely, in Twirly: the speech recognition (Parakeet v3), the language model for clean wording and rewriting (Qwen3-8B) and the read-aloud voice (Supertonic 3) live as files on your Mac and compute on the Apple Silicon GPU. There is no server anything could be sent to — the app touches the internet only for license activation and the optional update check. You can verify this yourself: dictation works in airplane mode.
The honest trade-off
Cloud wins on: language breadth (100+ languages), device breadth (Windows, Android, iPhone), and it runs on old hardware. If you work in teams that need shared dictionaries and admin features, there is hardly a way around cloud services.
Local wins on: confidentiality (the text only ever exists with you — relevant for lawyers, doctors, journalists, HR and anyone under NDA), availability (train, plane, bad Wi-Fi), latency (no network round-trip), and long-term price: local tools usually cost once, cloud services monthly — which adds up over the years.
Compliance matters too: for some professional-secrecy duties (legal, medical, banking) the question is not whether the vendor is trustworthy, but whether content may flow to third parties at all. On-device processing answers that question structurally instead of contractually.
How to tell what an app really does
Three checks that survive any marketing page:
1. The airplane-mode test — turn off the internet and dictate. If text appears, the app computes locally. If you get an error, the cloud was involved.
2. Search the privacy policy for “processors” or “sub-processors”. If cloud or AI-infrastructure vendors are listed, that is where the computing happens — whatever the homepage says.
3. Account requirement as a signal: an app that works without an account and without a login structurally has fewer ways to tie data to you.
Deciding quickly
You occasionally dictate short messages? Built-in macOS dictation is enough, and it costs nothing.
You need Windows, Android or more than 25 languages? Then a cloud service is hard to avoid — with the knowledge above.
You write on your Mac daily and want text to stay with you — without a subscription? That is exactly what we built Twirly for: dictate, rewrite, read aloud, 100% on your Mac, CHF 30 once.