yak.ad/diane
Diane
Dictation of the highest quality, into any application on any device, that stays confidential.
builtmeasuredin daily use on Windows; Android companion at concept stage
Not yet distributed — ask for access.
Windows. A microphone at the edge of the screen is the whole interface while you speak. Today the text arrives on the clipboard and is pasted where your cursor was; typing it straight into the field is in development.
Diane · dictationstoday
09:41 · 0:19 · mail2.8 s
The short version: keep the current plan, move the review to Thursday, and ask for the updated numbers before we send anything out.
09:37 · 0:28 · Slack2.1 s · 2 contested
Read the reports, check the patches, and try to figure out what might still be valuable from that, and put what you think yourself might still be useful later as a number of files in the same folder under shared, so we can pick it up later.
09:02 · 6:40 · document4.6 s
Help me write a specification for the future work on this direction. First, the current status: what we have is a good first prototype, and it proved useful to track a certain initiative live, being able to use it for further work on it. It also had very limited adoption, so we should try to put it to use in a few new ways. Let us talk through these ways one after another…
08:55 · 0:06 · terminal1.4 s
Rebuild the image and rerun the failing test on the staging host.
You correct the text where it landed, not here. Every dictation is still kept in full on your machine, contested words marked, so a dictation from a day ago can be found — and, in development, so that your corrections are picked up automatically and used to improve future transcriptions.
09:37▂▄▆ · 82
Can you send the notes from this morning?
Sure, one minute
The short version: keep the current plan, move the review to Thursday, and ask for the updated numbers before we
Android. The same microphone, floating over whatever app you are in; the text lands in the field you were typing into. Today it uses an ordinary cloud transcription service — see where this stands.
What it does
Press a key or tap the microphone, speak — a sentence or ten minutes — and the text is ready to paste where your cursor was: an email, a chat, a terminal, a document. Three to five seconds after you stop, measured, whatever the length — transcription runs while you are still talking.
Why the text is better
A single speech model is confidently wrong in ways you cannot see: a name becomes a word that exists, byte becomes bite, fluently and without a marker. Diane runs three models trained on different data, so they fail differently, and has a language model reconcile the three transcripts against your own vocabulary. Where the models agree, the text is almost always right; where they disagree, you are shown exactly where. In daily side-by-side use this produces cleaner text than any single-model tool or hosted transcription service compared against it — including the best of them — without giving up either latency or confidentiality.
What leaves your machine
Your audio never does. The three speech models run on your own GPU; the recording and every transcript stay on your disk, readable in full and deletable. How much of the text processing stays local is a deployment choice, and the choice widens with the hardware:
8 GB GPUAudio stays local. Reconciling the three transcripts happens in the cloud — today through Gemini under your own key, soon through confidential GPU compute, where the provider cannot read the text either.in daily use
12 GB GPUEverything local. A reconciler model runs beside the speech models; nothing leaves the machine, not even text.available
no GPUAudio and text are both processed in confidential compute — attested hardware the operator cannot look inside. The same quality and the same guarantee on any laptop, and on the phone.planned · the phone concept runs on an ordinary cloud service today
Measured on the reference machine: three to five seconds from the moment you stop speaking to text in the field, for a dictation of any length. Windows 11; Android 10 or later.
Where this stands
Windows app — three local engines, reconciliation, transcription while you speak, result to the clipboardin daily use
Typing directly into the focused field of any applicationin development
Picking up your corrections automatically and keeping them locally to improve future transcriptionsin development
Reconciliation — Gemini under your own key, or fully local on a 12 GB GPUin daily use
Android — the floating mic, typing into other apps, an ordinary (non-confidential) cloud transcription serviceconcept · in daily use
Reconciliation inside a confidential-compute APInext
Speech models in confidential compute — no GPU required, phone and laptop alikeplanned
