01The problem
Voice dictation tools are cloud-bound, app-bound, or both. I wanted Wispr-Flow-style dictation that works in any window on my Linux machine, with nothing leaving the machine.
02What I built
Hold right-Alt → mic capture → faster-whisper on CUDA → an Ollama LLM tidies English, while Hindi goes through a deterministic Devanagari→Hinglish transliterator I wrote (rule-based, so words are never mutated) → text is typed into the focused window. Ships with a systemd service.
03The outcome
I dictate into terminals, browsers, editors — anything. Still a work in progress and honest about its limits, but 'मैं ठीक हूँ' coming out as 'main theek hoon', offline, in any app, is already something I use.