Local Dictation
Voice and screenshots, collected in one flow. A local Mac prototype for explaining work without losing your train of thought.
Keep explaining while you show it.
A bug report often needs both words and a screenshot. Local Dictation is designed to collect both during one thought, then leave the final message for you to review.
It uses Apple’s on-device speech engine and local, rule-based corrections. It never presses Send.
Spoken input · tested English fixture
I want to add two into this idea. Oh no, sorry, I want to add three, not two.Corrected text
I want to add three into this idea.
Speak. Capture. Review.
Hold Right Option
Start dictating into the field you selected.
Press P to capture
Capture the display under your pointer while continuing to speak.
Press I to annotate
Mark the captured image. Supported spoken corrections clean up the transcript.
Release and review
The final text follows the ordered screenshots. You review and send it yourself.
Implemented in the prototype
- Apple on-device transcription, once a compatible speech asset is installed.
- Screenshot ordering, annotation and recovery when insertion is uncertain.
- Conservative corrections for supported phrases, without inventing image descriptions.
Planned exploration
- Optional local language-model cleanup for more flexible editing.
- Optional cloud API processing with explicit choice and clear data handling.
Providers are not selected. These options are not implemented, and no adaptive learning or reinforcement learning is claimed.
Current limits and readiness
This is a personal-test prototype for Apple Silicon and macOS 26 or newer. Complete live dictation and cross-app screenshot acceptance are still pending. The latest local check found stale macOS permission grants; signing, installation and physical workflow checks must be completed before a public release. Generic image insertion remains application-dependent.
Where your speech and screenshots go
The current transcription path runs on the Mac after Apple’s speech asset is installed. Screenshots are taken only by the capture shortcut and are not interpreted by an AI model. Output is inserted into the selected application; that application’s own data handling still applies. Recovery material may remain stored locally.
Have a workflow this could help?
Tell us which Mac apps you work in and what you need to explain. We will discuss the use case and what would need testing.