Hyperwisper

Hyperwisper is a voice dictation app for Windows that runs entirely offline. You press a hotkey, speak, and the transcribed text is pasted straight into the active window, whatever the application. No account, no subscription, no data leaving the machine: it is a free alternative to paid tools such as SuperWhisper.

Transcription runs on whisper.cpp, with six Whisper models available from 32 MB (Tiny) up to 1 GB (Large-v3). The default model ships with the installer, so the app works offline right away, and the others download on demand. Voice activity detection filters out silent recordings and automatically discards anything under 250 ms.

The real challenge was latency: dictation is only usable if the text appears almost instantly. Thanks to GPU acceleration through Vulkan, less than a second passes between releasing the key and the pasted text (measured on a Ryzen 9 5900X with an RTX 3060). A CPU-only build stays 2 to 3 times slower but remains practical day to day.

Under the hood, the core is written in Rust with Tauri 2, and the interface in React and TypeScript. The audio chain captures the microphone through CPAL and WASAPI, resamples to 16 kHz with Rubato, then feeds whisper.cpp via the whisper-rs binding. System integration handles a customisable global hotkey, synthetic keyboard input and the clipboard, restoring previous content and falling back to clipboard-only for protected windows.

Around all this I focused on real-world use: a six-step setup wizard (mic test, model choice, hotkey), a draggable floating overlay showing the waveform and a timer in two sizes, toggle or push-to-talk modes, local history, light and dark themes. It ships as a self-installing portable executable requiring no administrator rights, with a clean uninstall. The code is on GitHub under the MIT license, free to use commercially.