Everything said at this desk becomes text.

VonFlow is the voice layer of the studio's own desk. Hold a key and speak, record a call and read it back afterwards, or drop in a voice note. It all runs on the Mac it is installed on, and nothing it hears ever leaves that machine.

Untitled

Transcribing…

Inserted

One dictation. Drawn from the application’s own interface rather than filmed: hold the key, speak, let go, and the text is placed. The bars are the mark. At rest they settle into the V; a voice pushes them out of it and silence puts them back.

What it is

Dictation on a Mac usually means one of two things: the one the operating system hands you, or a subscription that ships your voice to somebody else's servers to be written up. VonFlow is the third option. It was built here in a single day on 3 July 2026, it has been how this desk writes ever since, and every word it hears is transcribed, cleaned and inserted on the machine itself. There is no account to make, no upload, and no third party holding a copy of anything it hears. The only copy is on this Mac.

Execution

What it does

Hold a key, speak, let go

The text lands where the cursor already was, in whatever was already open. From releasing the key to the words appearing takes about a second and a half at the median, spoken at a median of 123 words a minute; the transcription inside that runs at about 13 milliseconds per second of audio, roughly 76 times faster than real time. 506 dictations have gone through it so far: 70,218 words across 10.3 hours of speech on 40 days of work.

It hears two languages without being told

The language is read per dictation rather than set once in a preference: 439 English and 67 Dutch so far, with nothing to flip between them. Switching halfway through the morning costs nothing, because there was never a switch.

Record a call, read it afterwards

A call is captured as two separate tracks and written up with the speakers labelled, so the transcript reads as a conversation rather than a wall. 8 calls, 1.6 hours and 905 labelled segments since 10 July 2026. The one accuracy soak on record, on 31 July 2026, put word error at roughly 2 to 3 percent, clustered in proper nouns the model had never met.

Drop in a voice note

A voice note, or any recording the Mac can open, goes down the same path and comes back as text with no export step in the middle. What somebody left as audio becomes something you can read at your own pace.

It tidies up what was actually said

Fillers come out, 2,008 of them so far and about four per dictation. Punctuation goes in. A correction spoken mid-sentence is applied rather than transcribed. Names it has never heard are fixed against a dictionary of 39 entries, every one of them fed in by hand.

Nothing leaves the Mac

No server in the path, no account, no upload. The longest single dictation so far ran 1,517 words in one breath of 10 minutes and 43 seconds, and not a syllable of it touched the network.

Who runs it

One person built it and one person keeps it running. The first commit is dated 3 July 2026 and a working app landed the same day; there have been 30 commits since. It stopped being a demonstration immediately: 437 of those 506 dictations went straight into the editor where this studio’s work gets written. Every figure on this page was read from the application’s own database, on the maker’s Mac, on 15 September 2026.

We build tooling like this for other people's desks too: See what the studio automates

The studio’s character

This is Vonzie.

His mouth draws the same V the waveform settles into, which is the whole reason he looks the way he does. He belongs to the studio, not to this application: VonFlow turns speech into text, and that is the entire job. He is here because a desk you can talk at is worth a picture.

The studio behind it

Built here, run here.

We build our own software for the same reason we build yours properly: because the thing that has to keep working after launch is the thing worth getting right.