Convolution is mathematically exact for any system that does not change over time, which is precisely why it cannot follow a source you are moving.
An impulse response is a recording of one space, measured from one source position to one microphone position, on one day. Everything about it is frozen. That is not a defect, and it is the reason convolution captured from a real space still sounds better than anything modeled when the source stays put.
Geometry-based simulation starts from the opposite end. Feed it room dimensions and surface materials, and it works out where reflections would arrive and how loud they would be. Nothing is fixed, so the source can travel. That is the appeal of an engine that recalculates reflections while sources move, and it is why both approaches now sit inside the same plugins.
You can disguise this. Shorten the pre-delay, ride the wet level, filter the return as the source retreats. What you cannot do is change where that measurement was taken. A cathedral captured with the source twenty meters from the microphone will always be a source twenty meters from the microphone, however you automate around it.
The payoff is everything a model leaves out. Real halls scatter sound off seats, pillars, and surfaces no simulation bothers to describe, and they couple with adjoining volumes in ways that show up in the tail. That detail is why film and orchestral work still leans on captured spaces for anything static.
Late reflections get handled differently, usually with ray tracing, because that copes with scattering and builds a plausible diffuse tail. The two methods together produce something that behaves like a room. The decay time you end up with is still an estimate, and it can drift noticeably from what the textbook formula predicts for the same dimensions.
That formula is worth knowing, since every modeled room leans on it. Reverberation time comes out as 0.161 times the volume divided by the total absorption, in metric units. It lands within roughly 10 to 15 percent of measurement when average absorption stays below 0.25, and it starts overestimating once absorption passes 0.30.
Which means your material choice moves the result far more than your room shape does. Swapping concrete for heavy curtain changes the answer. Switching from a rectangle to a hexagon of the same volume barely moves it at all.
Below that point, discrete standing waves take over, and the frequency response can swing 10 to 20 dB when a listener moves a single meter. Geometric modeling is only valid above the transition, because it assumes energy spreads evenly in all directions, which is exactly what stops being true down there.
A captured impulse response does contain the real modal behavior of the space it came from, but only for the one position the microphone occupied. Accurate for a seat nobody is sitting in.
So treat the low end as something you place by hand rather than something either engine solves. Keep bass and kick close, centered, and mostly dry; let the modeled or captured space work from a few hundred hertz upward, and check what the room is doing to your sub content before you commit to a decay time that flatters the mids.
An impulse response is a recording of one space, measured from one source position to one microphone position, on one day. Everything about it is frozen. That is not a defect, and it is the reason convolution captured from a real space still sounds better than anything modeled when the source stays put.
Geometry-based simulation starts from the opposite end. Feed it room dimensions and surface materials, and it works out where reflections would arrive and how loud they would be. Nothing is fixed, so the source can travel. That is the appeal of an engine that recalculates reflections while sources move, and it is why both approaches now sit inside the same plugins.
An impulse response freezes one seat in one room
Convolution is the optimal representation of a system that is linear and unchanging. Rooms qualify, right up until you start moving things around inside them. The moment a source travels, the system becomes time-variant, and the math that made convolution exact no longer applies.You can disguise this. Shorten the pre-delay, ride the wet level, filter the return as the source retreats. What you cannot do is change where that measurement was taken. A cathedral captured with the source twenty meters from the microphone will always be a source twenty meters from the microphone, however you automate around it.
The payoff is everything a model leaves out. Real halls scatter sound off seats, pillars, and surfaces no simulation bothers to describe, and they couple with adjoining volumes in ways that show up in the tail. That detail is why film and orchestral work still leans on captured spaces for anything static.
Simulated geometry buys movement, not exactness
The common approach computes early reflections with the image source method, which treats every wall as a perfect mirror and builds a set of virtual sources behind it. It is fast and reasonably accurate for the first few bounces. Cost climbs steeply with each additional order of reflection, which is why plugins cap the depth rather than letting you run it forever.Late reflections get handled differently, usually with ray tracing, because that copes with scattering and builds a plausible diffuse tail. The two methods together produce something that behaves like a room. The decay time you end up with is still an estimate, and it can drift noticeably from what the textbook formula predicts for the same dimensions.
That formula is worth knowing, since every modeled room leans on it. Reverberation time comes out as 0.161 times the volume divided by the total absorption, in metric units. It lands within roughly 10 to 15 percent of measurement when average absorption stays below 0.25, and it starts overestimating once absorption passes 0.30.
Which means your material choice moves the result far more than your room shape does. Swapping concrete for heavy curtain changes the answer. Switching from a rectangle to a hexagon of the same volume barely moves it at all.
Both methods stop being honest below one frequency
Every room has a transition point where its behavior changes character, found by multiplying the square root of reverberation time divided by volume by 2000. A modest 100 cubic meter studio with a 0.6 second decay transitions around 155 Hz. A 3000 cubic meter hall with a two-second decay transitions near 52 Hz.Below that point, discrete standing waves take over, and the frequency response can swing 10 to 20 dB when a listener moves a single meter. Geometric modeling is only valid above the transition, because it assumes energy spreads evenly in all directions, which is exactly what stops being true down there.
A captured impulse response does contain the real modal behavior of the space it came from, but only for the one position the microphone occupied. Accurate for a seat nobody is sitting in.
So treat the low end as something you place by hand rather than something either engine solves. Keep bass and kick close, centered, and mostly dry; let the modeled or captured space work from a few hundred hertz upward, and check what the room is doing to your sub content before you commit to a decay time that flatters the mids.
