Historically, we appear to be approaching parity between real and virtual acoustic instruments in the context of recorded music playback. That is, we are approaching the time when many virtual instruments can be considered interchangeable with their real-world counterparts for recording purposes. Already, sampling synthesis gives us full interchangeability for the case of a single played note--the note which was sampled. Model-based techniques, however, are beginning to provide parity over a wider variety of performance expression, and they require far less memory (though more computational power) to achieve this.
The continuum between sampling and modeling is analogous to the extremes of motion photography versus computer-generated animation in film making. Just as computer-generated graphics is finding increasing use in films, model-based musical instruments are likely to grow in importance over time as the quality/cost ratio associated with their use increases. Note that ``cost'' should include overall ease of use as well computational and hardware costs.
Often students will ask why we bother simulating traditional instruments when the computer is capable of generating any possible sound. Why don't we focus mainly on exciting new instruments that are light-years beyond preexisting instruments? One short answer to this question is that artificially computed sounds tend to sound artificial. In other words, we simply don't know very many ways to generate deeply communicative sounds from scratch. Another answer is that traditional musical instruments are important because they are recognizable. Communication via sound waves is generally symbolic, referring to a shared library of common experience. Traditional musical instruments provide the important vocabulary necessary to articulate musical statements in terms of the prior repertoire. Also, attempting parity with traditional musical instruments provides an excellent test of our ability to construct efficient models, since we can test by direct comparison how well the model performs.
Once the natural behavior of traditional instruments has been conveniently captured in the form of computational models, the evolution of musical instruments can transfer from ``corporeal form'' in terms of wood and metal, etc., to the virtual world of computational models. At this point, they can begin an accelerated evolution. For example, virtual physical instruments may be distorted in ways that would be impossible in the real world, yet because they are model-based, distortions tend to remain recognizable to the listener as ``morphs'' of the basic instrument. Anyone who has tweaked ``parameter fifteen'' of a complex FM patch knows that this ``morphability'' property is not typical. Often, small changes in an instrument parameter will change the sound so drastically that there is no obvious connection between the ``before'' and ``after'' sounds. It is straightforward to provide recognizable morphs using sampled instruments, but in that case it's still difficult to obtain nonlinear phenomena such as saxophone growl, or the overblowing of a flute. Again, it all comes down to shared experience: A physical model captures in concise form a wide variety of sounds, all recognizable as coming from a particular kind of instrument by a wide variety of people. It is a valuable resource for composers to have instrument models capable of creating such a rich collection of vivid illusions in the mind of the listener, with intuitive controls.
Some instruments have already begun their evolution in the virtual world. For example, in the attempt to imitate the piano, various ``standard'' electric pianos have been created, such as the ``Rhodes'' and ``Wurlitzer'' electric piano sounds. Nowadays, it is relatively rare to see a real Rhodes or Wurlitzer keyboard; instead, they have become a family of presets on various synthesizers.
It does not necessarily follow that human performers will be replaced by computational models. The best way to test a virtual acoustic instrument is with a real player connected via physical controllers to the parameter dimensions of the model. On the other hand, the development of virtual performers is also greatly facilitated by virtual instruments, and such instruments will be upward compatible with arbitrary levels of sophistication in the automated enforcement of accuracy, style, and idiom.