Flowr
A local, on-device AI dictation app for Windows: hold a hotkey, speak, release, and the transcription is inserted into the focused app. Local Whisper / Parakeet / Moonshine speech recognition with Silero VAD endpointing, local-LLM cleanup via Ollama with per-app style context, custom vocabulary and voice snippets, and a fine-tuned speech-correction model — no account, cloud, or telemetry.
- Rust
- Tauri 2
- React
- ONNX
- Whisper
- Ollama
The pipeline
Hold a hotkey, speak, release. The transcription lands in whatever app has focus, entirely on-device, with no account, no cloud, and no telemetry.
Local speech and language processing
- Local speech-to-text via Whisper (GGUF) and Parakeet / Moonshine (ONNX) families, with Silero VAD endpointing and per-device mic selection.
- An optional local-LLM pass through Ollama removes fillers and fixes grammar before insertion, at a None / Light / Medium level, with per-app Style context.
- A dictionary handles term replacement and cleanup-prompt biasing; snippets let spoken triggers expand to literal text or an LLM-generated block mid-dictation; transforms rewrite the current selection in any app through a prompt.
- A fine-tuned speech-correction model keeps latency low without a round trip to a larger cleanup model for common corrections.
Shipped as a real product
Day-grouped transcript history with full-text search (SQLite FTS5), usage insights, global hotkeys, tray integration, autostart, and installer packaging (NSIS + MSI), built for daily use rather than a demo. Flowr started as a fork of Handy, rebranded and relicensed to Apache-2.0, extended toward feature parity with commercial dictation tools.