If you read their GitHub README and the code, it's possible to separate and cache the voice cloning step, and they have some streaming variant with low first chunk latency. I only played with it via Hugging Face. It is astonishingly good with certain rare accents, and it responds well to longer input clips