Deep Learning Based Assessment of Synthetic Speech Naturalness