Withy Voice

Push-to-talk dictation for macOS that runs entirely on your machine.
Hold a key, speak, let go — the text is typed into whatever app has focus.
Nothing is uploaded, nothing is transcribed in a datacentre,
and there is no per-seat licence.

withy  /ˈwɪði/  — rhymes with smithy
_{a willow branch cut for weaving}

_{The weeping willows on the water, and withies cut and sorted for weaving — Almere Buiten}

---

Install

git clone https://github.com/diraneyya/withy.git
cd withy
./install.sh

Then grant Hammerspoon Accessibility and Input Monitoring in System Settings → Privacy & Security, hold right Option, and speak. Requires macOS with Homebrew; no Python environment to manage.

Or hand the whole job to a coding assistant — paste this at it:

Install Withy Voice on my Mac by following https://raw.githubusercontent.com/diraneyya/withy/master/LLM.md

Updating: withy update, or Update Withy Voice… in the menu. Every choice you have made is kept.

---

Why you might want it

It is free, in the sense that matters. Not a trial, not a seat, not a plan you lose access to your own words without. If you use a hosted model to tidy the text you pay for those tokens and nothing else — no margin on top, and you can change provider, run a local model instead, or switch polishing off entirely and still have a working tool.

It is private if you want it to be. Speech recognition is always local. The tidying step is the only part that can leave your machine, it is your choice whether it does, and there is an offline switch that guarantees it doesn't.

Your words cannot be rewritten. Whatever model tidies the text, its output is checked against what you actually said. It may punctuate, paragraph, quote and list, and it may drop filler and a phrase you visibly replaced — but a word you did not say cannot reach the screen. The formatting is negotiable; the content is not.

Your vocabulary is a text file. Names, jargon, product names — the words speech recognition mangles — live in a list you can read, diff, and copy to another machine. Not a database, not a settings pane.

You can read the instructions it sends, and change them. The prompt is a file. Want British spelling, no em-dashes, shorter paragraphs? Say so, in your own words. And withy prompt test runs your version against a suite of real cases so you can tell whether you improved it or broke it — the same suite that gates changes to the shipped one.

Editing it doesn't freeze it. Withy remembers the instructions it last gave you, so an update can tell your edits from its own old defaults: untouched, it updates silently; edited elsewhere, both changes survive; edited in the same place, it asks you and shows what each version does.

You can keep talking while it catches up. Release the key and press again — dictations queue, and the banner stack shows you what is recording, what is waiting and what is being worked on. Results are typed in the order you spoke them, never the order they happen to finish.

Nothing is hidden. Which model is in use, how fast each one is on your machine (withy benchmark), what your vocabulary contains and where each term came from, what the polishing step was told, and every dictation you have made — all inspectable, all yours.

---

What it is honest about

It is a prototype, and it is built on Hammerspoon. That is why the permissions you grant are to Hammerspoon rather than to an app called Withy, and why there is no signed bundle in the menu bar of its own. It was written quickly, with an LLM, by someone who wanted it to exist — and it works, every day, as the tool its author dictates with.

It is slower than the commercial ones. Transcription runs on your CPU, so there is a pause after you let go that a paid tool with a datacentre behind it does not have. The queue makes that pause livable rather than absent. Pointing it at a transcription server you control is the obvious next step and is not built yet.

The polishing is good, not perfect. It is a prompt and a safety check, both of which are visible to you and both of which you can change — which is rather the point.

It is macOS only, and there is no Windows or Linux port planned.

---

Using it

Hold the record key, speak, release. That is the whole interface. The willow in your menu bar and a small banner in the corner say what is happening.

Everything else lives in that menu:

Recent dictations	click to copy; Type again, Copy unformatted, Redo from audio
Record key	fn, right ⌥, right ⌘, right ⇧, right ⌃, left ⌥
Polishing LLM	off, a local model, a local CLI, or a remote API
Speech model	whichever you have; more downloadable from the menu
Edit vocabulary	words a speech model would otherwise get wrong
Polishing instructions	how it should write — edit, test, or check for updates

The default record key is right Option, not fn — fn is the nicer key to hold, and it is also the one macOS and most commercial dictation tools take, so Withy stays out of its way until you choose otherwise. Install it on a different key from whatever you use today and compare them on the same sentences.

For organisations: terms everyone needs — internal tools, product names — can be shipped with the code in withy/vocabulary.d/ rather than typed into every config file. A fork adds its own file there and keeps merging cleanly from upstream. withy vocab always shows the user every term and where it came from.

Everything else — troubleshooting, how the pieces fit together, the design decisions and why each rule exists — is in LLM.md and CLAUDE.md, written to be read by a person or handed to an assistant.

---

The name

A withy is a willow branch cut for weaving.

Withy Voice was written in a garden in Almere Buiten, beside a heap of willow branches a neighbour had cut down the day before. The tree itself is gone — the one whose long strands used to move against the sky above a summer fire. What remains is a pile of rods and the question of what to do with them.

The traditional answer is to weave them. Withies become baskets, fences and screens; the tree goes on being useful in another shape. It is a fair description of a tool that takes speech — loose, disordered, and gone the moment it is made — and weaves it into something that holds.

It is also, plainly, a free local alternative to the dictation tools people pay per seat for. But it is named after the tree, not after them.

---

Licence

Apache-2.0. Chosen over MIT for the explicit patent grant, which is the question that comes up when someone installs a tool on a work laptop.