One key to dictate anywhere, and nothing leaves the Mac.

Hold a key, speak, let go: the text writes itself wherever the cursor is, in a message, a document or an email. Transcription runs on the Mac's own processor, offline, with no account and no subscription.

From the first file to the shared tool: 1 day
What's left of it today: 94 544 words transcribed, none of them sent to a server
First file: 15 June 2026

Open VoixFlash Free, open source (MIT) · written for myself first, then shared

lun. 12:41

Sans titre

micro plein : prêt

chargement → prêt → enregistre → transcrit → écrit

Hold the button, or the right Option key, then let go.

What the sentence meant

Dictating a message, transcribing a meeting, turning a recording into text: all of it existed already. Provided you paid a subscription per tool, around €20 a month (the going rate on the market, checked in July 2026), opened an account every time, and accepted that the audio would travel to a server. And provided you lived with tools that work every other day.

On 15 June 2026, at the end of the day, I stopped waiting for one of them to get good. The repository was born at 19:25; the first run is from the same evening. At 11:06 the next morning it was public on GitHub, installer and instructions included; the first shared version came out that same evening.

What was built

A tool for myself first. A menu bar application, held in a single file of 2 723 lines, sitting on the native components of macOS. Transcription runs on the Mac's processor; the model downloads once at install time, and everything after that is offline.

Two paths for the same sentence

On the left, what an online service takes for granted. On the right, what VoixFlash does.

The gesture, seen through the icon

The icon has no colour: its shape is what says the state. One dictation goes through four of them.

The text is pasted in one block, never letter by letter: that is what keeps documents intact.

The key is chosen from the menu: right Option by default, or any other, captured on the fly. A very short press dictates nothing.

Meetings and files

LiveStart a meeting
A menu item, for as long as you like, 3 h at most as a safeguard. The text appears in a window of the app, editable, copyable, exportable to a dated file. The clipboard is left alone.
A fileImport an audio file…
mp3, m4a, wav, flac, ogg, aiff, or the audio track of an mp4 or mov video. Progress shows in the menu, and can be cancelled without losing the text already obtained. Files of 2 h and more.
Who is speakingSeparate the speakers
For live meetings: “Locuteur 1”, “Locuteur 2”, from 2 to 6 people expected, or automatic detection. A local module of about 50 MB, downloaded once, with no account.

Installing, in one double-click

You double-click the installer. A Terminal window opens, prepares the application's own corner, downloads the model once, and registers it to start with the session. Nothing is placed in the system: it all lives in one folder of the user's library, and a second double-click removes the lot.

Ce que Terminal affiche

install.command

===================================================================
   Installation de VoixFlash
===================================================================

• Code copié dans : ~/Library/Application Support/VoixFlash
• Création de l'environnement isolé (téléchargé si nécessaire)…
• Installation des bibliothèques (transcription, audio, clavier)…
• Téléchargement du modèle de transcription (une seule fois).
   [/] Téléchargement en cours… (plusieurs minutes, c'est normal)
  ✓ Modèle prêt : la transcription marche hors-ligne.
• Configuration du démarrage automatique…
• Vérification du démarrage de l'application…

===================================================================
   Installation terminée : l'application tourne.
===================================================================

“Everything is local. After installation, no audio and no text leaves the Mac. No account, no telemetry.”

What VoixFlash does not do

It only runs on an Apple Silicon Mac, and it is not signed: macOS warns on first launch.
It does not use the graphics card: everything goes through the processor. At small quality, an hour of audio takes about fifteen minutes; medium is markedly slower.
A dictation and a meeting each have their maximum length, and only one transcription runs at a time.
Speaker separation degrades when several people talk at the same time.
Protected files (DRM) are refused.

The result

“I'm done with tools that ask you to pay” became: 94 544 words transcribed on this Mac since 26 June 2026, dictations and meetings together, without a single one leaving the machine. An 8 GB M2 Mac, zero subscriptions.