On Using Backpropagation for Speech Texture Generation and Voice Conversion