Transmission of Speech in Packet-Switched Networks

Abstract

Packet switching speech not only allows straightforward multiplexing of speech and data, but also enables efficient use to be made of transmission capacity. This thesis shows how the transmission efficiency of a packet-switched speech network can be optimized. To do this, an analytical model of the packet queueing process is developed, and subjective tests on the effects of packet losses are described. Together these provide a powerful tool for the performance assessment of a packet-switched speech link. The analytical model is based on a two-dimensional Markov chain, and models the queue as a continuous variable. This enables the delay distribution and amount of packet loss to be found by solving a first-order matrix differential equation. The accuracy of the analysis is investigated through the use of simulations, and is found to be good for the range of parameters of interest. The subjective tests clarify inconsistencies in the literature, and give reliable indications of the effects of packet losses. A preliminary test shows that 8ms packets with gaps filled by repeating the previous packet give the best speech quality. A listening test shows that speech with packet losses of up to 20% can be understood with only moderate effort. Finally, a conversation test shows that 2.5% loss is acceptable to almost everyone, 10% loss is acceptable to 70% of people and 25% loss is acceptable to 33% of people. Some factors that affect the performance of a packet-switched speech link are investigated, namely the way packet loss and delay are used, packet length and receiver buffering. Ways to ensure adequate data performance, and the trade-off between TASI advantage, delay and packet loss are also investigated. From these investigations, principles for optimizing the transmission efficiency of a packet-switched speech network are established.