Semi-Supervised learning is of great interest in a wide variety of research areas, including natural language processing, speech synthesizing, image classification, genomics etc. Semi-Supervised Generative Model is one Semi-Supervised learning approach that learns labeled data and unlabeled data simultaneously. A drawback of current Semi-Supervised Generative Models is that latent encoding learnt by generative models is concatenated directly with predicted label, which may result in degradation in representation learning. In this paper we present a new Semi-Supervised Generative Models that removes the direct dependency of data generation on label, hence overcomes this drawback. We show experiments that verifies this approach, together with comparison with existing works.