Offline Dictation on a Mac: How On-Device Speech-to-Text Works in 2026
Your Mac's chip can now transcribe speech faster than you talk — no cloud, no internet, no audio leaving the machine. Here's how on-device dictation works, why it's faster than typing, and what it means for privacy.
For a decade, good dictation meant one thing: your voice was recorded, uploaded, transcribed on a server farm, and the text sent back. It worked, but it required an internet connection, added latency, and meant that everything you said out loud passed through someone else’s computer.
That era is over. An Apple Silicon Mac — any M1 or newer — can now run speech recognition locally that rivals what only the cloud could do a few years ago. Here’s what changed, how it works, and why it matters for the way you write.
Why dictation is worth caring about at all
The math is simple and a little embarrassing for keyboards. Most people type 40–60 words per minute. Conversational speech runs at around 150 words per minute. If a meaningful part of your day is producing text — emails, messages, notes, documentation, code review comments — speaking is the single biggest input upgrade available to you.
What kept most people from dictating wasn’t speed. It was that raw transcripts read like, um, well, you know, like this — and cleaning them up by hand ate the time you saved. Modern dictation fixes that with a second step: after speech becomes text, a language model removes fillers and repairs punctuation. The output reads like something you typed, at the speed of something you said.
How VoiceDict’s on-device speech recognition works
VoiceDict has three stages. The first two run on your Mac; the third is optional:
1. Capture
The app listens only while you hold or toggle a shortcut (in VoiceDict, a global hotkey you choose). Audio goes into memory, not into a file on a server.
2. Speech recognition
A neural speech model transcribes the audio. VoiceDict uses NVIDIA’s Parakeet v3, a model that covers 25 European languages and runs comfortably in real time on Apple Silicon. This is the part that used to require a data center: the combination of efficient model architectures and the Mac’s neural-friendly chips closed the gap.
3. Cleanup
The raw transcript — “um so basically can we uh move the demo to thursday” — can optionally go to Luna through the OpenAI API using your own key. It removes filler words, restores punctuation and capitalization, and follows the conventions of the language you spoke. VoiceDict sends transcript text, never audio. Then the text is inserted at your cursor, in whatever app you were using.
Without formatting, transcription works in airplane mode. Luna formatting requires an internet connection.
What “offline” buys you
Privacy that doesn’t depend on a policy. Cloud services can promise not to store your audio; an offline app can’t store it anywhere else in the first place. Dictation audio is uniquely sensitive — it’s your unfiltered thinking, spoken aloud, often including names, numbers and things you’d never put in a sent message. The strongest privacy guarantee is architectural, not contractual.
No connectivity tax for transcription. Train Wi-Fi, flights, flaky hotel networks, a VPN that blocks half the internet — none of it prevents local speech recognition. Optional formatting waits for the OpenAI round trip.
No word caps or metering. Cloud transcription costs the provider money per minute, which is why subscription apps meter free tiers (2,000 words a week is a common cap). Your own chip doesn’t send invoices. Dictate a novel if you like.
Predictable app cost. Local speech recognition is why VoiceDict can remain a €9.99 one-time purchase. Optional Luna formatting is metered separately by OpenAI through your own API account. A typical dictation subscription is $144 a year, every year.
What you need
- An Apple Silicon Mac — M1 or newer. The models lean on the unified memory and the GPU/Neural Engine; Intel Macs aren’t up to it.
- A recent macOS — VoiceDict needs macOS 15+.
- A one-time model download — the speech and cleanup models are fetched once (this is the one moment you need internet), then live on your disk.
- Some memory while dictating — VoiceDict’s cleanup model uses roughly 1.5 GB of RAM while you’re actively dictating and releases it about ten seconds after you stop.
Try it against your keyboard
The only benchmark that matters is your own next email. Start a free 7-day VoiceDict trial — no credit card — press your shortcut, and say the email instead of typing it. Watch the fillers disappear and the punctuation snap into place, then check your Wi-Fi icon: it never mattered.
Also read: Wispr Flow alternatives without the subscription · VoiceInk vs VoiceDict