The advantage of this method is that you can train on random speech datasets, and then using only a minute or two, find the "voice embedding" of a person, and g

The advantage of this method is that you can train on random speech datasets, and then using only a minute or two, find the "voice embedding" of a person, and generate anything with his/her voice. This voice embedding is much like the word2vec word vectors, and have similar arithmetic properties. Responding to multiple sibling comments.