In view of the trend that the transmission control protocol (TCP) is increasingly used instead of the user datagram protocol (UDP) for multimedia information transmission in the next generation networks, this paper studies the performance of the real-time multimedia stream transmitting over TCP. Based on the overall consideration of the characteristics of TCP including fast retransmission, fast recovery and changing regularity of the congestion window, the impacts of TCP buffer on the playback quality are analyzed from a micro perspective. Based on a further study made on the process of assembling audio frames and disassembling video frames, the conclusion that it is unfeasible to transmit real-time audio on the Internet using TCP is proved. Also, it is presented that to transmit real-time video via TCP with the acceptable playing performance, the sending-delay of a video frame should be less than the difference between the play buffer delay and half of the round trip time (RTT). The conclusions are evaluated to be significant and referable for developing the streaming media delivery technology through the extensive simulations using the Ns2 simulator.