How it works

Aleph has three moving parts:

Satellites

The devices you talk to: a desktop overlay, an Android kiosk on a wall tablet, your phone, or an ESP32 board with a microphone and speaker. Satellites stream audio to the server over the Alabama protocol.

The server

The aleph binary. It routes audio, manages satellites, hosts the Web UI, and supervises the ML pipeline.

The ML pipeline (sidecar)

A subprocess the server launches and manages for you. It runs speech recognition, the language model, and speech synthesis on your GPU. You never interact with it directly.

A voice turn, end to end

1. A satellite hears the wake word and starts streaming your voice.
2. The server forwards audio to the pipeline; speech-to-text produces a live transcript.
3. A turn detector decides you've finished speaking.
4. The LLM answers — possibly calling tools along the way.
5. Speech synthesis streams the spoken reply back to the satellite.