Comment by 100ms

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