Abstract: A method of large-scale fault-tolerant audio conferencing in an audio conferencing system using a purely packet-switched network. An endpoint places a call to a conference gatekeeper indicating an audio conference. The conference gatekeeper determines whether the call contains sufficient information to establish the audio conference. If there is insufficient information, the endpoint is connected to an IVR server that obtains sufficient information from the endpoint. Either way, a CACS selects an MCU hosting or that will host the audio conference. The CACS then responds to the endpoint with routing instructions indicating the selected MCU and the endpoint connects or transfers to the selected MCU. The MCU mixes input from all endpoints in the audio conference to form a voice stream, which is then returned to each endpoint in the audio conference. Audio conference participants can dial-out from the MCU to bring additional participants into the audio conference.