How to specify the vocals about karasinger

[jerrygood0703]

Hello, we've experimented with two methods:

1. Use one-hot singer embedding as the start token, and the vocal identity can be controlled.
2. Use all-zero vector as the start token, the output singer will be randomly selected from the training data.

Both methods generate vocal identity within the training data.
Referencing by a sample voice is something we'd like to try but haven't try yet, thank you for your question!