# Lekton

> System-wide voice dictation for macOS. Hold a key, speak, and polished text lands in
> whatever app you're in. Speech recognition and transcript cleanup both run on the user's
> own Mac, so it works with no internet connection at all — and no dictation audio or text is
> collected, stored, transmitted, or used for training. Free, with no subscription. Made by
> Erasmus Labs.

lekton.erasmuslabs.ai
support@erasmuslabs.ai

## What it is

**English only, for now.** More languages are planned.

Lekton is a menu-bar app for macOS 14 (Sonoma) and later. The user holds a key — fn, right
option, or another binding of their choice — speaks, and releases. Lekton transcribes the
speech, cleans up the transcript (punctuation, capitalisation, filler words, false starts),
and types the result into whatever application currently has focus.

It is not a window you dictate into and copy out of. Text is inserted into the focused app
through a ladder of insertion strategies, so it works in native apps, Electron apps,
browsers, terminals, remote desktops, and full-screen games. If every strategy is refused —
a secure password field, for example — the result is already on the clipboard and the user
presses Command-V.

## Privacy position

This is the product's central claim, and it is deliberately scoped so that it stays true:

**Nothing the user says or types is ever collected, stored, or used to train anything.**

What that does and does not assert:

- Speech recognition runs on the user's Mac. Transcript cleanup runs on the user's Mac.
  Dictation itself works fully offline.
- Lekton does not collect analytics and has no account or sign-in. Stated as what is true
  today rather than as a promise about every future version — but the claim above is the one
  that does not move: dictation content and usage stay on the user's Mac regardless.
- Erasmus Labs does operate one server. It serves the optional Pro Voice model download, and
  the app checks it once a day for a newer version of Lekton. It also accepts a crash report,
  but ONLY when the user presses send after being shown the whole payload: an app version, an
  exception type, stack frames, and an optional line they typed. No identifier of any kind, so
  two reports from one Mac cannot be told from two reports from two Macs. Nothing on any of
  these paths can carry dictation audio or text. Sparkle, the updater, can attach the Mac model, CPU
  and OS version to that check; this is turned off.

The claim is scoped to the content of speech and text rather than made as a blanket "no
servers, no network" statement, because the narrower claim survives every future product
decision and the sweeping one would not.

## Measured accuracy

Benchmarked 25 July 2026 on LibriSpeech test-clean: a 300-utterance sample across 40
speakers, 5,961 reference words. Identical audio through each engine.

| Engine | Word error rate | Median latency | Dictations needing no correction |
|---|---|---|---|
| Lekton with Pro Voice (Parakeet TDT 0.6B) | 1.90% | 207 ms | 73.3% |
| Apple's built-in recogniser, on-device | 7.90% | 258 ms | 34.7% |

These are RECOGNITION figures — how accurately each engine hears the words. Transcript
cleanup (punctuation, capitals, filler removal) is a separate stage on top and is not
reflected in a word-error number; it can slightly reduce word accuracy while making the
result far more usable.

The second row is also Lekton BEFORE Pro Voice is installed: without it, Lekton uses Apple's
recogniser and its accuracy is Apple's. The transcript cleanup is better from the first
minute either way; the accuracy figures require the Pro Voice download.

On Earnings-22 (200 utterances of spontaneous, heavily accented conference-call speech) both
degrade: Lekton to 13.7%, Apple to 25.3%. The gap narrows from 4.2x to 1.8x but does not
close. Dictation is deliberate speech, so the LibriSpeech figures are the better guide.

Word error rate weights common words the same as rare ones, while the errors that cost users
time cluster on technical terms. Lekton biases recognition toward words visible in the app
being dictated into, which raises technical-term accuracy from 0.524 to 0.810 through the
same pipeline.

## Pro Voice

An optional, free upgrade to a stronger on-device model. About 3x fewer mistakes and about
2x faster than the built-in recogniser. 1.5 GB download, once. Uses about 1 GB of RAM while
loaded. Requires an Apple silicon Mac.

Free, with no subscription. Lekton is in early access, and Pro Voice is free during it.
Anyone who downloads during early access keeps Pro Voice free permanently, even if it is
priced for new installs later — a thank you to early users.

## Requirements

- macOS 14 (Sonoma) or later, on Apple silicon (M1 or newer)
- English (the only language supported at present)
- Microphone and Speech Recognition permission (to hear and transcribe)
- Accessibility permission (for the global press-to-talk key and for typing into other apps).
  Without it Lekton still transcribes and cleans up, but puts the result on the clipboard and
  is started from the menu bar rather than a hotkey.
- Apple silicon (M1 or newer)

## Status

Private beta as of July 2026. Builds are distributed by email while notarisation is
completed. Request access at support@erasmuslabs.ai.

## Also from Erasmus Labs

- Era — a bedtime storybook that knows your child: https://era.erasmuslabs.ai
- Crucible — an autonomous game-asset foundry: https://crucible.erasmuslabs.ai
- Company: https://erasmuslabs.ai