This paper describes an audio delivery system model called Kendra. We focus
on Kendra's adaptive delivery processes and the experimentation carried ou
t to analyse it. The system trades off quality of sound against periods of
silence; therefore, adaptation changes the format of the audio data to save
bandwidth when bandwidth becomes more scarce. Typically adaptive systems a
re composed of a monitoring component and a switching component. The Kendra
monitoring component has predictive capabilities, which advise the switchi
ng component to adapt either to a better quality or to a more bandwidth fri
endly format when it detects lower bandwidth. The prediction element can va
ry the amount of data it uses in its prediction, trading-off accuracy with
the system's impedance on the performance. Such systems exhibit many other
tradeoffs that determine how sensitive and/or optimistic the system is to i
ts environment, and this is explored in our experimentation. We conclude ou
r work by discussing these network conditions and tradeoffs and how they af
fect the performance of the system. (C) 2000 Elsevier Science Inc. All righ
ts reserved.