Why speaker encoder can encode the information of speaker?

[liyihao17]

I observed that the speaker encoder are not trained with the speaker index (label), so why the speaker encoder has the ability to encode speaker information?