25/08/2026
Following up on the AI phone demo I posted last week β here's how it actually works under the hood.
Instead of sending calls to a cloud AI service (slower, costs per call, sends audio outside your network), I built a self-hosted pipeline: Asterisk handles the call, then a local AI worker (Whisper for speech-to-text, Ollama for the AI response, Piper for text-to-speech) handles the conversation β all on my own infrastructure.
The same setup also supports multiple phone lines, each with its own AI "personality," from one deployment.
Still very much a work in progress, but wanted to share the architecture for anyone curious how these AI phone systems are actually built.