It seems that the discriminator tests if the output of the generator obeys the pattern of a language. Isn't it still possible that after training the generator outputs the text wrong for the input? On the other words, the output is understandable but actually not what the sound indicate?