Zuban
Offline Speech-to-Text (Handy-based repo)
Problem
The current Zuban repository is not a chat tutor stack; it is an offline speech-to-text app codebase focused on private transcription across desktop platforms.
It follows the Handy architecture: local recording, VAD-based silence filtering, and model inference without sending audio to cloud APIs. This keeps privacy guarantees strong and costs predictable.
Architecture
Key Decision
Kept the transcription path local-first: Tauri desktop shell, Rust backend for system/audio integration, and selectable model engines (Whisper and Parakeet).Global shortcut
└─→ Start/stop local recording
└─→ VAD trims non-speech segments
└─→ Whisper or Parakeet inference
└─→ Transcribed text
└─→ Paste/insert into active text field
Impact / Learnings
The recent commit stream is strongly focused on transcription benchmarking and CPU efficiency, including faster-whisper evaluation paths and compute-type tuning.
This page intentionally reflects repository reality to avoid resume drift. If a separate tutor product exists, it should be linked as a distinct repo with its own architecture and metrics.