How does an NN generate audio without generating samples? The outputs are somehow interpreted as a continuous function / polynomial with tons of terms? And nobody ever once tried to not do that?