If you had enough recordings, you could (I think) build weights based _solely_ on whale speech. Humans wouldn't be able to understand the weights, and the word

If you had enough recordings, you could (I think) build weights based _solely_ on whale speech. Humans wouldn't be able to understand the weights, and the word vectors in that model wouldn't match the word vectors in an English model, but I suppose there's a chance that vectors might be similar? I don't know. I think you'd have to be very good at both linguistics and also AI to know.