WezaFlow

WezaFlow is a local-first desktop dictation pipeline inspired by Wispr Flow. The implementation is organized as a Tauri desktop shell plus Python services for audio capture, speech recognition, context detection, LLM formatting, and text injection.

The first implementation slice focuses on stable module boundaries:

• services/asr: microphone capture, VAD, and Whisper transcription
• services/context: active app and browser context detection
• services/llm: local Ollama-based text formatting
• services/injection: clipboard and keyboard text insertion
• services/pipeline: shared contracts and orchestration
• db: SQLite schema for vocabulary, snippets, profiles, corrections, and history
• scripts: platform setup, startup, and test helpers

Windows

Run the scaffold tests with:

py -3 -m unittest discover -s tests -v

Build the Windows Tauri release with:

npm run tauri -- build

macOS test build

The macOS path is intended for developer testing from a GitHub checkout. It uses the same Tauri shell and Python runtime, with macOS-specific Command+C/Command+V text injection through pyautogui.

From a Mac:

bash scripts/setup_macos.sh
npm run tauri -- dev

For a local macOS release build:

npm run tauri -- build

On first use, macOS may prompt for Microphone and Accessibility permissions. Grant Accessibility to the terminal during dev runs, or to WezaFlow when running the built app. The GitHub workflow in .github/workflows/build-macos.yml can produce a downloadable unsigned macOS test artifact from GitHub Actions. See docs/macos-test-build.md for current limits.