Abstract: A system and method for providing video conferencing are described herein. Original video streams from one or more cameras are sampled to form downscaled versions of the video stream. The downscaled version is analyzed to find regions of interest and generate metadata describing aspects of the regions of interest, including a region that includes a preferred speaker and the best view of each participant present together in the video conference. Cropping instructions are generated from the metadata, and portions of the original video stream are removed to form a set of presentation views for each participant, including the preferred speaker. The set of presentation views is stitched together to form a composite display for each participant viewing the stream remotely.