Theoretically this is very possible using their techniques. They tokenize the audio and learn next tokens autoregressively. Instead of text tokens -> audio toke

Theoretically this is very possible using their techniques. They tokenize the audio and learn next tokens autoregressively. Instead of text tokens -> audio tokens as input, just tokenize a prior song and continue it.