AI · Building
Tarjuman
Self-hosted, fully offline live speech transcription, translation and speaker diarization — in a single Rust binary.
What it does
Jul 2026
Tarjumān (ترجمان) means "interpreter" in Urdu, Arabic and Persian. It listens through the browser mic, applies voice-activity detection, transcribes in the original language — built with Hindi, Urdu and Arabic in mind, though any Whisper language works — translates each utterance to English live, and labels who is speaking.
It runs entirely on your own hardware. No cloud APIs, no telemetry, no audio ever leaves the machine, and no audio is stored at all: only text persists, to a local SQLite database with a browsable history view.
Results stream over WebSockets so partial transcriptions appear as you speak. Metal acceleration on macOS, an ARM build small enough for a Raspberry Pi, and one binary that ships the HTTP and WebSocket API together.
Build
Personal project
- Deployment
- Single binary
- Cloud APIs
- None
- Smallest target
- Raspberry Pi
Stack
- Rust
- TypeScript
- Whisper
- SQLite
- WebSockets
- Docker