Septa

Talk.
It comes out clean.

Hold a hotkey, speak, paste. A 0.6B model on this machine strips the ums, keeps the correction, and leaves your words alone. No cloud. No API key.

Download for Mac Download for Windows

Apple Silicon Mac · ~424 MB cleaner on first run · unzip, then double-click Install.

Raw ASR
so um send it to rahul no wait rohan, the invoice is twenty thousand rupees due march third
Septa
Send it to Rohan. The invoice is ₹20,000 due March 3.
Septa cleaner, raw transcript next to cleaned text
Casual wording kept after cleanup Dictated list turned into markdown bullets

Mac is the full app.

Septa is Hex with the cleaner wired in. Global hotkey, on-device Whisper or Parakeet, then the local model before paste. Quit Hex once you switch.

Hold to talkSame press-and-hold / double-tap as Hex. Text lands where you were typing.
Fails openIf the cleaner is asleep, you still get the raw transcript. Dictation is never dropped.
Stays yoursgonna stays gonna. It will not invent Priya’s email. Read the PRESERVE contract in the repo.

Windows gets the same cleaner.

Hex itself is macOS-only, so the Windows download is the sidecar: same 0.6B GGUF, same POST /v1/clean, same browser UI. Paste a messy transcript, get the cleaned one. First run fetches llama.cpp and the model.

Double-click installUnzip Septa-windows.zip and run Install Septa.bat. Python 3.10+ needed.
Local onlyNothing leaves the machine. The UI is at http://127.0.0.1:8742
Same fixtures30/30 on the held-out set as the Mac Q5 build.