Record
One tile per way you talk.
Voice Note, Meeting, System Audio, Translate — that’s the starter board. Every tile is yours to change: its source, its model, its colour, its start sound, its hotkey, and what happens to the text when it’s done.
WhisperMLX turns speech into text on your Mac’s own silicon — live while you talk, offline if you like, and delivered straight into the app you were working in.
macOS · Apple silicon · iPhone companion
Fast, private speech-to-text on your Mac.
Live preview
Okay — three things before the review tomorrow. First, the export bug: it only happens when the folder name has a slash in it. Second, ask design about the empty state on the library. Third, and this is the important one: ship the model manager as it is. It’s better than what we said we needed.
Record
Voice Note, Meeting, System Audio, Translate — that’s the starter board. Every tile is yours to change: its source, its model, its colour, its start sound, its hotkey, and what happens to the text when it’s done.
Live preview
A quick model drafts a rolling preview while you’re still speaking, so you can see it’s getting things right. When you stop, the final pass runs on the model you actually chose.
Delivered
Stop — or press the same hotkey again — and the transcript is typed at your cursor in the app you started from. Or copied. Or sent to an app you name. You choose per tile, once.
Dictation tools end with you hunting for the text. Here the text comes to you — four ways, chosen per tile.
The classic. The transcript lands on the clipboard the moment it’s done — with or without timestamps, your call.
Sentence by sentence into the app you were in when you pressed record. A long silence ends a sentence, so it reads like you meant it.
…and that’s the whole idea.▍
The whole transcript at once, for apps that don’t like being typed into. Same cursor, one keystroke’s worth of waiting.
A tile can always deliver to one app — your notes, your editor. It’s brought to the front when the transcript is ready, opened if it has to be.
Typing and pasting into other apps use one Accessibility permission, granted once. Password fields are always left alone.
0
bytes of audio leave your Mac.
Whisper runs on your Mac’s own silicon — weights in unified memory, compute on the GPU, through Apple’s MLX. There is no server to send your voice to, so aeroplane mode changes nothing.
Six models, from a 75 MB scratchpad to the full Large v3. Download the ones you want; keep one warm in memory so recording starts instantly.
The only network requests the app makes are the one-time model downloads themselves.
One track from the microphone — you — and one from the Mac’s own audio: everyone else on the call. The transcript reads like a conversation, plays back mixed, and either track can be isolated when you need to hear one side again.
Capturing the Mac’s audio uses macOS’s screen-recording permission, granted once.
Start a Translate session from the menu bar and a small floating window captions whatever it hears — in the language being spoken, out in English. For the meeting you only half-follow, the lecture, the film without subtitles.
…das Modell läuft komplett auf dem Gerät.
The model runs entirely on the device.
Nothing is sent anywhere — that’s the point.
Record, stop, quick-transcribe, copy the last transcript — without ever opening a window. The panel comes in nine styles, from system-standard to a mission-control console.
WHISPERMLX
Recent — click to copy
Transcription is usually a service wearing an app as a disguise. This one is just an app.
No account, no key, no server. The engine lives on your Mac, so privacy isn’t a policy you have to trust — it’s a network request that never happens.
Keep recordings for re-transcribing, encrypted at rest with AES-GCM, the key held in your Mac’s Keychain. Or keep nothing — audio retention is a choice, not a default you have to find.
Screen-sharing while you record? The menu bar icon can hide itself for the length of the session, so nothing on screen betrays that a recording is running.
The things that don’t need a headline.
Each tile can have a system-wide shortcut. Press it to start, press it again to stop and transcribe — whatever app you’re in.
The companion app records from the Action Button, your Mac transcribes, and the text lands on both clipboards. Away from the Mac, recordings queue until you’re back.
Keep the audio and run a bigger model on it later — the quick draft you took in the hallway becomes a clean transcript at your desk.
Segments are editable in place and deletable line by line. Every change is versioned, and any version is restorable.
Nested folders, drag and drop, search within the folder you’re in. The whole library is mirrored to plain Markdown on disk — yours to grep.
Names, jargon, product words — keep a list and the decoder is steered toward them, so “MLX” stops coming out as “emelex”.
Two-pass mode transcribes with the quick model immediately, then re-runs the big one and swaps the text in when it’s ready.
Each tile can start with its own sound and record in its own colour — you can tell recordings apart by ear, and by the menu bar icon.
WhisperMLX is part of AlexApps. The subscription covers all of them — including whatever comes next.
What the subscription covers
Or $29.99 a year — two months back.
Coming soon