What is whimprflow?

WhimprFlow is a voice dictation app: you hold down a key, speak, and clean text appears wherever your cursor is, whether that's a chat box, an email, or a document. Speech recognition happens on your own device using a tool called Whisper, and a local AI model then cleans up the raw transcript by removing filler words, fixing spoken self corrections like changing your mind mid sentence, adding punctuation, and formatting lists and paragraphs. There is also an optional cloud option if you would rather send the cleanup step to OpenAI or Anthropic. The project is explicitly a proof of concept, built quickly, and the author is upfront that it is rough around the edges and not production ready. It works as a starting point rather than a polished tool. It runs on macOS, where it has been built and tested on Apple's own chips, and on Windows, where it also builds and runs, including push to talk with the Right Ctrl key, speech recognition, and cloud based cleanup. A couple of Windows only gaps remain: the on-device cleanup AI runs on the processor rather than the graphics card, and automatic learning of new words from your corrections only works on macOS so far. Both platforms currently require building the app from source, since there is no ready-made installer yet. Other features include a small floating status bar showing whether the app is idle, recording, or processing, a personal dictionary that learns names and terms you use often, and local usage statistics like words dictated, typing speed, and daily streaks. Under the hood it is built with Tauri, pairing a Rust backend with a React and TypeScript interface. The code is split into separate pieces for speech recognition, audio capture, cloud cleanup, and the local cleanup AI, so each part can be worked on independently. The app is not affiliated with any existing commercial dictation product, it is an independent recreation of that kind of workflow built from scratch. Privacy is a stated priority: speech recognition and default cleanup both happen on your device, and if you choose the cloud option, only the text transcript is sent, never the audio, with API keys stored in your operating system's secure keychain rather than in a plain file. It is written mostly in Rust and released under the MIT license.