Playing with Carolina's idea of an expanded theremin.
There are two major questions: inputs and outputs.
Minority-Report-style lights on hands in front of a webcam?
Kinect body scanning in real time?
Arduinos with accelerometers?
Power-glove-style strain gauges?
Theremins use capacitive sensing. This is clearly a possibility, but it is harder with high-speed electronics around, since the capacitance signals you're looking for are very small. An actual theremin circuit might be the best way to measure proximity.
Ultrasound signaling is another traditional approach for VR equipment.
Some of this is tricky because the input channels we're considering are possibly pretty high latency: webcam plus image processing algorithms, say. So it might make some sense for the musician to control the tempo and type of rhythm rather than the exact timing.
(Some prototyping will reveal how big a problem this is in practice. You could imagine that a QVGA 320x240 image wouldn't really take that long to process, and might be adequate.)
The possibility of the music coming out is, fortunately or unfortunately, many-dimensional. Even a single note can be loud or soft, high or low, bent up or down, of many different timbres, and of many different envelopes, and at any time. An entire melody has many notes, and the relationships they bear to one another are potentially very complicated.
So one crucial question is how many low-latency high-bandwidth dimensions of input we can achieve: two with each hand? Another with angle? Another with finger separation? Other joint orientations?
One possible mapping is with radius for volume or timbre, and angle for pitch. Going around the circle more than once would go up or down a whole octave.
Another possibility is moving virtual objects around a space, maybe a time-frequency representation where time is streaming past you.
Another kind of exploration would adjust parameters of different state machines with your movements: a melody generator, a rhythm generator, a timbre generator or two. Maybe, at its simplest, you could have one direction to move in to seek change, and another direction to move in to stay the same; or you could have a sort of pie menu thing where you're always selecting one of the parameters to vary, and your distance from the center controls how fast it varies. It's crucial for the feedback here to be within a few hundred milliseconds in order for the interface to be discoverable.