Quote:
Originally Posted by jamwerks  Question: do convolution reverbs using IR's achieve this "Billion" or close? Do they give you the "real" room? What are the limitations of IR? |
Convolution reverbs don't achieve the billions of resonances. Technically, a convolution impulse is an FIR filter, which has NO resonances (i.e. poles) - just anti-resonances (i.e. zeros). In practice, this makes no real difference. Convolution can have a high enough perceived resonance density to accurately simulate a non-time varying room.
However, most rooms are time-varying, due to temperature variations in the room which cause small variations in the speed of sound. These small variations are enough to cause largish errors in the late frequency response of the convolution impulse, and the small pitch changes that the real room would have cannot be captured by convolution. A real acoustic space will demonstrate the same types of randomization as found in the best algorithmic reverbs, although a real acoustic space will have a far greater number of modulation sources, with far more subtle "settings" of each modulator.
Having said this, in a fairly small acoustic space (a "room," not a "hall"), the speed of sound differences will be fairly subtle. I think that using convolution to model smaller spaces is a good idea. This also works well with how the cost of convolution scales with the length of the impulse - long halls will take more CPU than short rooms. Compare this to a typical algorithmic reverb, where the CPU is the same for any decay length.
A counter to the argument I just gave: People love the EMT250 for short ambiences. This algorithm is TOTALLY time varying.
Getting back to the billions of resonances: The billions of resonances are from a physical perspective. From a perceptual perspective, it is highly unlikely that we can perceive those billions of resonances accurately, or discriminate them from several hundred thousand resonances with the right type of time variation applied. I haven't seen many studies on this, but one paper suggested that a few thousand resonances can do the trick, if time variation is properly applied to each resonance separately. This approach (a few thousand parallel 2nd order filters, each with its own randomizer) is still too expensive for modern computers, although it would lend itself well to GPUs. I would want to hear the results before commenting on the quality of this approach. Many academic papers have made claims about how a given reverb approach is indistinguishable from an acoustic space, but very few of those claims hold up over time. The demands of the marketplace put higher demands on sonic quality than what seems to be going on in acadamia.