| With recent changes in the social and physical characteristics of the Internet, streaming media applications have acquired great popularity. In this environment, it is important to develop efficient techniques for reliable transport of streaming media, since general-purpose techniques, such as those used by TCP, do not satisfy the real-time requirements of streaming applications, and new types of media provide new challenges. In this thesis we develop three new techniques to combat loss in streaming media, in all parts of the spectrum of possible methods: concealment, recovery and avoidance of loss.; First, we provide a technique for loss concealment in streaming multichannel immersive audio, a promising future application. Our technique is based on the insight that this environment has inherent inter-channel redundancy. We use this insight to develop a method for concealing losses in a channel by using samples from other channels to substitute for the losses. We subjectively evaluate the performance of our method in comparison to the traditional intra-channel concealment method, and find that our method significantly reduces the number of audible artifacts.; Second, we provide a technique for loss recovery that performs efficient retransmissions by avoiding timers and network delay estimation. This technique is based on the insight that a self-clocking mechanism can be used for repeat retransmissions of the same packet, allowing automatic adaptation to varying network delay. We implement this technique over UDP both for unicast and multicast streaming, and demonstrate its performance gains in comparison to a similar method based on timers.; Finally, we propose a technique for loss avoidance, applicable when streams are delivered over multiple paths using multiple senders or overlay networks. We systematically develop and study the foundation of this proposal, a method for path fingerprinting based on the distribution of packet pair dispersions over the path. We establish the properties of our new fingerprinting method using simulation, analysis and measurements, demonstrate that these properties form a basis for detection of shared links, and propose that such detection can be used for loss avoidance. |
