A system for producing conversation-driven character animation includes a computing platform having processing hardware and a system memory storing software code, the software code including multiple trained machine learning (ML) models. The processing hardware executes the software code to obtain a conversation understanding feature set describing a present state of a conversation between a digital character and a system user, and to generate an inference, using at least a first trained ML model of the multiple trained ML models and the conversation understanding feature set, the inference including labels describing a predicted next state of a scene within the conversation. The processing hardware further executes the software code to produce, using at least a second trained ML model of the multiple trained ML models and the labels, an animation stream of the digital character participating in the predicted next state of the scene within the conversation.