A Blue RnD R&D project: talk to a 3D puppet by voice or text and it answers out loud while it moves. A state machine runs record, speech-to-text (whisper.cpp), chat (llama.cpp) and text-to-speech (Piper). The LLM emits emotion and pose tags that drive 13 skeleton poses, with amplitude-based lip-sync.
Providers sit behind one interface with three backends: local servers, cloud APIs, and on-device ONNX through Unity Sentis. I also built a Piper voice-cloning pipeline and shipped a custom voice, multiple puppets with their own voices, and a scripted-scene player.