Sayso — Git skill for Claude Code

Sayso

Local voice dictation for Windows. Press a key, speak, press it again, and what you said is typed into whatever you had focused — your editor, your browser, your chat window, anything.

It runs entirely on your own machine. No account, no subscription, no API key, and no audio ever leaves the computer.


What makes it different

**It types what you actually said.** There is no "clean it up with an LLM" pass, because a reworded instruction is a wrong instruction when you are dictating into a terminal or a code editor. Whisper supplies the punctuation and the capitals; the words stay yours.

**It learns your words, not the internet's.** Whisper already knows English. What it fumbles is proper nouns: your clients, your street, your product names. Put them in the Dictionary and they are fed to the model as hints *before* it decides, which fixes the problem at the source instead of find-and-replacing the output.

Requirements

• Windows 10 or 11
• An NVIDIA GPU is strongly recommended. large-v3 needs about 3.1 GB of VRAM and transcribes 6 seconds of speech in under a second on an RTX 3060 Ti. Without a GPU it falls back to CPU automatically and is slow but works.
• About 3 GB of disk for the model, downloaded once on first run.
• uv for the Python environment.

Set it up with Claude Code or Codex

Paste this into Claude Code or Codex from the folder you want it in:

Set up Sayso, a local Windows dictation app, from
https://github.com/YourFutureSiteDev/sayso

1. Clone it, then run `uv sync` in the repo root.
2. Run `uv run python -m sayso --devices` and show me the microphone list so I
   can pick mine.
3. Run `uv run python -m sayso --self-test`. It downloads the Whisper model
   (about 3 GB, once) and checks the microphone, the GPU, transcription
   accuracy and the typing path. Show me the word error rate it reports.
4. If that passes, run `build.cmd` to produce dist\Sayso\S