fix: update Dockerfile dependencies, remove audio processing, and replace models

Update model stack, add audio stream extraction, and harden inference wiring

Initial commit