Systems and methods for setting up live online conversations between people, and selecting participants for such conversations in a manner that optimizes the enjoyment of the participants. Conversations are analyzed and attributes of participants and conversations are extracted from the audio of the conversations. A quality level of the conversations is determined from the audio, and subjective quality assessments may be received from participants. Participants and conversations are divided into classes and clusters, where the classes and clusters represent groups of entities which have certain attributes in common. Based on the attributes, classes, clusters, and quality levels, compatibilities are determined between arbitrary pairs of participants and for arbitrary groups of participants. Conversations are generated, and participants are selected for the conversations based on the determined compatibilities.