Voice pipeline

What happens between you speaking and Aleph answering.

The stages

1. Wake word — the satellite (or server) listens for the wake phrase.
2. Voice activity detection — separates your speech from silence.
3. Speech-to-text — a streaming transcript appears as you talk.
4. Turn detection — a small model decides when you've finished your thought, so you don't have to pause awkwardly.
5. The LLM — generates the answer, calling tools if needed.
6. Text-to-speech — the reply streams back as audio while it's still being generated.

Why it feels fast

TODO: streaming at every stage, speculative starts.

Interrupting

TODO: barge-in behaviour.