Frierenclaw/fern ? reverse-engineered prompt

Build the core backend for a real time voice conversation system. I want this service to sit between a desktop client and the rest of the AI stack, handle the live audio connection, receive microphone audio, run voice activity detection and speech to text, send the text or tokens onward to the inference service, then stream back synthesized speech plus lip sync or viseme data for the client to animate.

It should feel strict and fast, so if the user starts talking while the assistant is speaking, it should cut off the current buffers immediately and switch over cleanly. Use async code throughout and make it reliable enough for continuous streaming. Include support for storing anything needed in Postgres, Redis, and S3, and wire up the config so it can run locally with Docker. If you need to check current docs for the audio or streaming pieces, look them up online first.