EE Systems Seminar

Communication systems for interactive voice and video applications must operate under strict latency constraints and yet provide high reliability. We will study an information theoretic model for streaming communication where source packets (messages) arrive at the encoder in a sequential fashion, and must be reconstructed at the decoder by the associated deadlines. We will focus on a variety of channel models.

Motivated by VoIP applications we first consider the packet-erasure model. We will show that the structure of loss patterns (e.g., burst vs random losses) have significant effect. We will introduce distance metrics associated with these loss patterns and propose a family of codes that are near optimal with respect to these metrics. We will also present some experimental results over real-world traces.

We next consider the discrete memoryless model. We assume that a stream of messages arrives at the encoder, one per block of say n channel uses, and each message must be decoded after a fixed delay of T blocks. We study the channel dispersion and moderate-deviations constant for this setup and show how each improves over block coding schemes. If time permits we will also discuss some results on fading channels.