Duper Whisper: private speech-to-text for Windows
My own project
Press a hotkey, speak, and the text appears in whatever app you're using. It runs on a local GPU, so nothing leaves the machine and there's no subscription.
The problem
I wanted dictation that works in any app, including a code editor or terminal, without sending my voice to a cloud service.
- 33
- user stories in the written specification
- 6
- planned phases, one commit each
What I built
- A Windows tray app with a global hotkey, microphone capture and two ways of entering text: paste, or simulated typing for apps that block paste
- A FastAPI server in WSL2 that keeps a Whisper model loaded on the GPU, and can swap models without restarting
- A nine-step clean-up of the text: punctuation, numbers, filler words, spoken commands like "new line", and custom word replacements
- A settings window with live changes, start-up at login and clear messages when something goes wrong
About 1,600 lines of Python, with a README covering setup, GPU checks and troubleshooting.
Tools: Python, FastAPI, faster-whisper, CUDA, WSL2, customtkinter
Diagrams and screenshots
