This HeyGen realtime avatar workflow is what I used to build a live AI version of myself that can respond with low latency like a Zoom call. The key is the training footage structure (active listening, speaking, idle) plus a strong persona/context setup so the avatar doesn’t feel generic.