Model Inversion Involves Pronunciation Sound