A method for video conferencing includes receiving, at one or more computers from at least some remote clients from a plurality of remote clients, information representing a plurality of media frames. The method also includes receiving, at the one or more computers from at least some of the plurality of remote clients, a plurality of video-based speech activity signals each associated with a respective media frame from the plurality of media frames. The method further includes selecting, at the one or more computers, at least some media frames from the plurality of media frames based on the video-based speech activity signals, decoding the selected media frames, generating a mixed media stream by combining the decoded media frames, transmitting, from the one or more computers to at least some remote clients from the plurality of remote clients, the mixed media stream.