- Stars
- 1
- on GitHub
- Category
- rust
- Developer
- justin06lee
- Language
- Rust
- Runs on
- desktop
A self-hosted Wispr Flow replacement that learns your voice. Hold a hotkey anywhere, speak, release — clean text lands in the focused app.
Everything runs on your own hardware: Whisper (via faster-whisper) for transcription and a local LLM (via Ollama) that strips filler words, punctuates, and adapts tone to the app you're dictating into. And the longer you use it, the better it gets: every dictation is recorded, your corrections become training pairs, and once enough accumulate, una LoRA-fine-tunes Whisper on your own voice, evaluates the result against the current model, and hot-swaps it in — only if it's actually better.
┌─ client (Tauri: macOS + arm64 Linux) ────┐ ┌─ server (your NVIDIA Linux box) ─────────┐│ tray + floating HUD + settings │ LAN │ FastAPI ││ global hotkey → record 16kHz │─────▶│ faster-whisper (CUDA) ││ POST wav → receive text → paste-inject │◀─────│ → Ollama cleanup (falls back to raw) ││ mDNS auto-discovery of _una._tcp │ │ SQLite + WAV history │└──────────────────────────────────────────┘ │ web dashboard: home · review · ││ dictionary · training · insights ││ fine-tune: LoRA → WER gate → hot-swap │└──────────────────────────────────────────┘
Information
- Developer
- justin06lee
- Category
- rust, desktop app, self-hosted, machine learning
- Compatibility
- desktop
- Language
- Rust
- Updated
- sep 1, 2026
- Year
- 2026
- Source Code
- github.com/justin06lee/una
