The variation of packet-arrival intervals is one of the problems to be solved in realizing real-time voice communications on asynchronous networks such as the Internet. Although the variation can be absorbed by a receiving buffer, the end-to-end delay increases in proportion to the amount of buffered data. Therefore, the guideline for the receiving buffer's design that considers the trade-off between voice quality and delay is needed. In this paper, the design of the receiving buffer by means of the voice-speed-conversion technique is discussed. According to our scenario, the buffer size was expected to be equivalent to the time length in which the probability distribution function of the packet-arrival intervals was about 95% to 99%. However, it is shown that synchronized voice reproduction can be achieved within a preset delay, which is shorter than the corresponding time to the buffer size. This is achieved by keeping the rate of voice speed conversion within the range of 50% to 150% depending on the variation of the intervals. The proposed method was implemented in the conventional UNIX workstations (WSs) connected to an Ethernet segment. The average voice delay between WSs was 150 to 180 ms and the synchronized voice reproduction was achieved with a standard deviation of 13 to 25 ms. The proposed method is applied to VSAT-satellite-channel gateways to realize the real-time voice communication between the WSs through a VSAT channel.