Parla
How it works
No step between your voiceand your text you cannot see.
Every dictation tool calls itself private. The more useful question is which bytes cross the network, and what happens to your words on the way. Here is the whole of it.
The pipeline
Four stages, all on your Mac
Streaming recognition on the Neural Engine
NVIDIA's Nemotron 3.5 multilingual speech model — 600 million parameters, converted to run on Apple silicon — transcribes while you speak, with partial results as you go. It switches between English and Spanish on its own.
Remove only what is certainly filler
"Um", "uh", "eh", doubled words. Words that are filler in one sentence and meaning in the next — English "like", Spanish "este" — are left alone.
Your vocabulary, matched by sound
A recogniser does not misspell "Reckitt"; it hears "racket". So Parla compares how words sound, not how they are spelled, and uses a second, stricter pass for words no dictionary knows.
A language model, on a short leash
The model built into macOS fixes punctuation and grammar, and can lay out paragraphs and lists. Its output is only used if it passes the checks below.
Privacy
The honest ledger of what leaves your devices
Including the entries a marketing page might prefer to leave out.
Stays with you
- Your voice. Transcribed and discarded. No recording is written to disk, on the Mac or the phone.
- Every transcript. Recognised on your Mac's Neural Engine — or, as a fallback, on the phone itself.
- The cleanup. Done by the language model already inside macOS.
- Your vocabulary. Colleagues, clients and product names never travel to anyone else.
- Your history. Text only, on your Mac, readable only by you, easy to switch off.
Crosses a network
- Phone audio, to your own Mac. Over home Wi-Fi or your Tailscale network, sealed with a key the Mac made for that phone and handed over only inside the pairing code.
- Tailscale's coordination. If you use it away from home, Tailscale knows your devices exist. The traffic between them is end-to-end encrypted.
- The speech model. Once. About 600 MB, on the Mac's first launch.
- Nothing else. No telemetry, no analytics, no crash reports, no accounts.
Guards
A tool that quietly rewrites you is worse than no tool
-
Late output is discarded
If the polished version does not arrive in time, you get the cleaned transcript instead. Waiting is never the price of polish.
-
Length is a lie detector
A result more than 40% shorter or 60% longer than what you said is rejected. So is one that looks like the model replied to you rather than corrected you.
-
Registered names are protected
If you said a name on your list and the polished text lost it, the polish is thrown away. Getting a company name wrong is far worse than an unpolished sentence.
-
You always know which you got
Whenever a model's edit is used, Parla says so. Only the explicit draft mode is allowed to write words you did not say.
Structured style
Dictate an email, receive an email
Prose, then a list, then more prose — which is what a dictated message really looks like. The same works in Spanish, with primero … segundo … por último.
hello peter I hope you are okay I just wanted to agree the list of actions for our next meeting one book taxis two book hotels looking forward to our workshop in lisbon tomorrow thank you
Hello Peter, I hope you are okay. I just wanted to agree on the list of actions for our next meeting: 1. Book taxis 2. Book hotels Looking forward to our workshop in Lisbon tomorrow. Thank you.
Offline
A tunnel is not an outage
On the Mac, the speech model lives on disk from the first launch, so dictation behaves the same on a train or a plane as it does at your desk.
On the phone, the Mac is the better ear. When it cannot be reached, Apple's on-device recogniser steps in, followed by the same vocabulary correction and polish — and the result tells you which engine produced it.
Measured
Numbers from an actual machine
Recorded on an Apple M4 Mac mini with 16 GB of memory. Not projections.
| Figure | What it measures |
|---|---|
| 12–17× | Faster than real time, transcribing while you speak |
| ~0.7 s | Polishing a dictated sentence |
| 1.6–4.1 s | Structured output — a 110-word email corrected and split into paragraphs at the top end |
| ~4 s | Reloading the speech model after it was unloaded to save memory |
| 30 min | Idle time before the model is unloaded, keeping the Mac light |
| 0 | Bytes of audio sent to anyone but your own Mac |
Known limits
What it does not do well, yet
-
Lists need numbers
"Book the flights, sort out the hotel, get insurance" stays as a sentence. Without ordinals there is no reliable way to tell a list from prose, and bulleting things you did not mean to list would be worse.
-
Grammar is corrected gently
Agreement, tense and apostrophes are fixed; register is left alone. That is deliberate — preservation is weighted heavily.
-
Spanish accents are mostly, not always, restored
The cleanup model adds missing accents most of the time. Not every time.
-
Sound-alike names have a cost
With "Reckitt" on your list, "rocket" can become Reckitt too. Only context separates them, and Parla does not guess at context.
Questions about the internals?
Happy to talk about how it is built
The design decisions — and the dead ends — are written up in detail. If you work on something similar, get in touch.