Background Matting is a computer vision problem where the foreground of an image or a video is separated from the background. We emphasize the real-time performance of the model for Matting without using any input besides the original captured image, such as a tri-map and a background image. We also prioritize the use case for this work on video conference. To solve this problem, we use a model with an Encoder-Decoder architecture. Our main contribution to this work consists of proposing and experimenting with a new loss function for training the model of the matting problem, replacing the Normalization layer, and creating a composite dataset for video conferences named Typical Conference Backgrounds (TCB). Promising empirical experimental results have been achieved by our method on the public AIM and Distinction-646 datasets.