Something Cool I've Been Working On

I'm new to IR's and would like to understand what I am looking at. Is the blue the original IR and the red is the ultra-res replication of the blue?
 
voes said:
Interesting comment of Jay Mitchell on the Gear Page on this subject

"Cliff is mistaking mathematical artifacts for acoustic ones. It's a very common mistake but a mistake nonetheless."

Here's the link:

The Gear Page - View Single Post - Future IR Technology from Fractalaudio?
Click to expand...

It's not a mathematical mistake and nothing to do with windowing.

That peak in the bass response might be the speaker, it might be the room. Looking at the actual IR it appears to be the speaker as there are no discernible early reflections.

The frequency resolution of an IR is the sample rate divided by the number of samples in the IR. The window function has nothing to do with frequency resolution (except for making it even less). So a 1K IR at 48 kHz sample rate has a frequency resolution of roughly 48 Hz. If a speaker has a resonance (formant) at, say 80 Hz with a Q of, say, 3.0, then 48 Hz is insufficient to capture that resonance accurately. You need a frequency resolution of several Hz to accurately recreate that resonance. I chose 80 Hz and a Q of 3 because that's what that response looks like. The Q could even be higher than that.

It doesn't take much mental energy to realize that if you have a narrow formant at a low frequency then you need fine frequency resolution to reproduce that. An 80 Hz formant with a Q of 3 only spans about 25 Hz. Obviously a frequency resolution of 48 Hz is not going to be able to reproduce that.

Windowing only smooths the response even more. This is basic FFT theory. The less time-domain information you have, the less frequency domain information you have and vice-versa. This is the uncertainty principle. I always window IRs with a Hann window.

EDIT: I broke out my impedance measurements for that Vox cabinet and the speaker resonance is 80 Hz.
 
Another way to look at it is to think in terms of formants. That particular speaker has a pronounced 80 Hz formant. It takes well over 100 ms for the energy of that formant to decay to the point of imperceptibility. Obviously a 20 ms IR can't reproduce an event that occurs for over 100 ms.

Here is a zoom of the original non-minimum-phase IR (IOW raw time response):
vox_ir_zoom.jpg


You can clearly see the 80 Hz formant. There are some room reflections but they are very small. The 80 Hz formant starts well before any reflections. It's obviously a high-Q resonance as it rings for quite a while. The higher the Q, the longer it takes to decay.
 
FractalAudio said:
The problem with conventional IRs is that they are too short to capture the detail in the low frequencies. There are those that maintain 20 ms is the maximum length you need to fully replicate the speaker. This would be about 1000 samples at 48 kHz.

I disagree with this as I have many IRs here that exhibit significant energy beyond 20 ms. I believe the room has some influence as the low-frequency modes of the room will impact the resulting sound. The amount of this impact depends on the room, the mics, distance, etc., etc. Or perhaps certain speakers have particularly high Qs in the low frequencies. Regardless, it is my opinion that you need IRs much longer than 20 ms to fully capture the "mic'd amp in the studio" sound.

My tests show that IRs of 8000 samples are required to fully capture the low-frequency detail. Unfortunately to process an 8K IR in real-time require copious processing power...

Fortunately I have developed "UltraRes (TM)" cabinet modeling. UltraRes cabinet modeling provides the frequency detail of a very long IR with little or no added processing power requirements.

The following image depicts the response of UltraRes cabinet IR processing:
The blue trace is the frequency response of the original (long) IR. It is a Vox AC-30 cab.
The green trace is the frequency response of the same IR truncated to 1K samples. This is "normal resolution" in the Axe-Fx and 2-4 times longer than what other products use.
The red trace is the frequency response using UltraRes processing.

I’ve shifted each trace by 0.5 dB to make comparison easier.

ultra_res.jpg
Click to expand...
You lost me after "The" :lol ​ AWESOME Cliff !!!!!!!
 
FractalAudio said:
It's not a mathematical mistake and nothing to do with windowing.

That peak in the bass response might be the speaker, it might be the room. Looking at the actual IR it appears to be the speaker as there are no discernible early reflections.

The frequency resolution of an IR is the sample rate divided by the number of samples in the IR. The window function has nothing to do with frequency resolution (except for making it even less). So a 1K IR at 48 kHz sample rate has a frequency resolution of roughly 48 Hz. If a speaker has a resonance (formant) at, say 80 Hz with a Q of, say, 3.0, then 48 Hz is insufficient to capture that resonance accurately. You need a frequency resolution of several Hz to accurately recreate that resonance. I chose 80 Hz and a Q of 3 because that's what that response looks like. The Q could even be higher than that.

It doesn't take much mental energy to realize that if you have a narrow formant at a low frequency then you need fine frequency resolution to reproduce that. An 80 Hz formant with a Q of 3 only spans about 25 Hz. Obviously a frequency resolution of 48 Hz is not going to be able to reproduce that.

Windowing only smooths the response even more. This is basic FFT theory. The less time-domain information you have, the less frequency domain information you have and vice-versa. This is the uncertainty principle. I always window IRs with a Hann window.
Click to expand...



Took the words out of my mouth :)
 
Here's another example. This is one of the new OwnHammer IRs. The IR is OwnHammer_412_MAR-CB_D-120_SS_RBN-121. These IRs are 100 ms long (4800 samples). I windowed the original IR to 4K to prove a point.

The blue trace is the original IR (windowed to 4K samples).
The green trace is the "typical" 20 ms IR (windowed to 1K samples).
The red trace is the UltraRes version.

ultra_res_2.jpg


As you can see there is significant error at 60 Hz between the original and 20 ms version. Not coincidentally this is the resonant frequency of the speaker itself. Again looking at the original raw IR there is a very pronounced formant that is still ringing at 100 ms. The error is approximately 6 dB! Anyone who believes this isn't audible is kidding themselves.
 
Here is the raw time series for that D120 IR:

d120_ir_zoom.jpg


Again one can see a pronounced formant at the speaker resonance.
 
  • Like
Reactions: sam
FractalAudio said:
So a 1K IR at 48 kHz sample rate has a frequency resolution of roughly 48 Hz. If a speaker has a resonance (formant) at, say 80 Hz with a Q of, say, 3.0, then 48 Hz is insufficient to capture that resonance accurately. You need a frequency resolution of several Hz to accurately recreate that resonance. I chose 80 Hz and a Q of 3 because that's what that response looks like. The Q could even be higher than that.

It doesn't take much mental energy to realize that if you have a narrow formant at a low frequency then you need fine frequency resolution to reproduce that. An 80 Hz formant with a Q of 3 only spans about 25 Hz. Obviously a frequency resolution of 48 Hz is not going to be able to reproduce that.
Click to expand...

An 8k file @ 48kHz has a frequency resolution of 5.8Hz - I'm not an expert (maybe I should keep my mouth shut, but I can't :lol ), but is it possible to use a fine frequency resolution on lower frequencies and "avarage" (good) resolution on highs? Mmh.....I guess I better keep my mouth shut.....:nightmare:
 
Ringleader said:
Interesting that this "more sliders in the EQ" concept appears to finally being realized. And don't let Jay get you down about it. Not everything he says is law. Curious to hear it for sure!

http://forum.fractalaudio.com/amps-cabs/71451-clr-guys.html#post876004
Click to expand...

Yes, that's exactly the way to look at it. The problem is that human perception is logarithmic and IRs are a linear process. 48 Hz resolution is way more than necessary at, say, a few kHz but not nearly enough at low frequencies. The brute force solution is to use very long IRs, 8K or more. UltraRes solves this in a novel way that uses little to no extra processing power and no additional latency.
 
voes said:
Interesting comment of Jay Mitchell on the Gear Page on this subject

"Cliff is mistaking mathematical artifacts for acoustic ones. It's a very common mistake but a mistake nonetheless."

Here's the link:

The Gear Page - View Single Post - Future IR Technology from Fractalaudio?
Click to expand...
hmm... it's been also said many times before by some guys who produce very good IRs, that the longer, uncut .wav IRs sound better because they include more of what's initially been captured and which just can't 100% be reproduced by the shortened IRs which we for example use.
will have to let our ears decide who's right or wrong....
 
You must log in or register to reply here.
Back
Top Bottom