Musical sound can be generated from numerical solutions obtained from
simple physical models of wind-instruments. We call such synthesizers
'virtual instruments'. Crude caricatures capture the global oscillatio
n behaviour of original instruments, providing the musician with a mea
ns of expression comparable to that obtained when playing real instrum
ents. Music is, however, produced by details of the flow which corresp
ond to much smaller temporal and spatial scales than the control of th
e global oscillation. This suggests that one has to split the physical
model into two distinct parts: a simple numerical oscillator followed
by a complex sound production module, driven by the output of the fir
st module. We illustrate the background of this idea by a discussion o
f the clarinet, the human voice, and the trombone.