1/N-Octave smoothing <=> Frequency dependent window
New version:
Version 1.4.1.4 (2009-10-15)
Bugfixes:
* Factor of 2 when calculating 1/N octave filters
* Show legends on butmaps can be unchecked
* Invert amplitude causes normalization to be invalid
New User Guide with Mathematics v0.0.6:
http://holmacoustics.com/holmimpulse.php
Or
HOLMImpulse > Help > 'User Guide (Checks for update)'
on page 15 to 22 I derive why:
1/N-Octave smoothing <=> Frequency dependent window
New version:
Version 1.4.1.4 (2009-10-15)
Bugfixes:
* Factor of 2 when calculating 1/N octave filters
* Show legends on butmaps can be unchecked
* Invert amplitude causes normalization to be invalid
New User Guide with Mathematics v0.0.6:
http://holmacoustics.com/holmimpulse.php
Or
HOLMImpulse > Help > 'User Guide (Checks for update)'
on page 15 to 22 I derive why:
1/N-Octave smoothing <=> Frequency dependent window
Worked for me. I read the section in question and remain unconvinced. In fact his own example shows how the smoothing rejects the delayed noise. The point here is that in this example thats what you want, but what if there is some delayed diffraction effects - they shouldn't be rejected by the smoothing, but they will be with this technique.
The math only shows how a filter in the frequency domain has an analog as a window in the time domain. It does not prove that the two things will yield the same results for a spectral smoothing done as suggested in this discussion. I think that its clear that they won't.
Clearly the time domain windowing is going to make a speaker look better. Its advantage is that its easy.
Take for example a mouth reflection from a waveguide. These are delayed in time and will therfor be cut-off at some point by the shortened time window at higher frequencies when smoothing is applied per the technique shown here.
Last edited:
Well I have found the smoothing to be closer to what I actually am hearing. Couldn't this be useful for at least getting an approximation of the way we actually hear things in a room? That is what I have been using it for personally so I don't see the smoothing as useless exactly. Maybe not totally accurate because as you point out you are probably throwing away some audible artifacts along with some of the reflections that I don't normally hear. But I guess I just don't find the reflections that a mic picks up exactly useful to what I am testing for.
My data says the exact opposite. Group delayed signals are highly audible, increasing in audibility at higher SPLs. This effect would be ignored with the time domain windowed technique, but not with a true frequency averaging (albeit this kind of thing is never highly significant in the frequency domain). "Smoothing" makes a lot of sense, but not if it rejects signal components that we know cause audible artifacts.
How do you smooth ?
How do you smooth?
Please reveal your formulas, so I can implement your method
How do you smooth?
Please reveal your formulas, so I can implement your method
How do you smooth?
Please reveal your formulas, so I can implement your method
I average the magnitudes of the frequency bins as a weighting filter is swept through the frequency response data. Its not very efficient and I'd like a better method, but it does work.
I plan to try yours as well.
Last edited:
Isn't this what is called stepped sine method?
What when you measure your polar-patterns using impulse responses from HOLMImpulse?
----------
Well, I do not claim to have the correct smoothing formula. I now have 4 methods:
A. No smoothing, no window
B. Time-window
C. Absolute log-smoothing (I use this for digital room-adaption)
D. Complex log-smoothing <=> Frequency dependent window
Personally I don't see the advantage of D for correcting speakers (this is my goal) or making room-adaption. A lot of other measurement software implement D in a not flexible and transparent way, so I find it nice to see what I'm doing in the time-domain.
I need to make a careful comparison with pink noise power response and stepped sine.
Last edited:
No its not stepped sine, thats different.
I take the impulse response and find the FR, normal FFT. Then I go to a set of log spaced frequencies and at each of those find a weighting function that is 1/N wide - I use a gaussian type of shape. The magnitudes at each of the FFT data point is then weighted by this shape and integrated (summed). This procedure has trouble at LF when the frequency spacing of the data is comparable to the smoothing bandwidth, but works good otherwise.
There is no "absolute" correct method - all measurements of loudspeakers are evolving, but its important to understand how and why they all might differ.
Above a few hundred Hz the ear's critical bandwidth is measurable in logarithmic terms such as fractions of an octave, but below this the critical bands approach constant frequency width the lower you go. Karjalainen et al in JAES (Journal of the Audio Engineering Society) Vol 47 No. 1/2 show a filter structure for speaker correction taking this aural characteristic in mind, which shows improved performance in the lower frequencies.
gedlee 1/N smoothing
1. When you sum the RF - is it then a complex sum? or do you sum the amplitudes?
2. When you use your weighting function - do you multiply with 1/f to compensate for your equidistant frequencies?
1. When you sum the RF - is it then a complex sum? or do you sum the amplitudes?
2. When you use your weighting function - do you multiply with 1/f to compensate for your equidistant frequencies?
ARTA smoothing vs HOLMImpulse
Thanks, ARTA "2.3 Frequency Resolution of DFT and Octave-Band Analyzers"
if ARTA window = BLACKMANN3, then this corresponds to HOLMImpulse amplitude smoothing with N = 1, N = 3
The HOLMImpulse FR makes a calculation for all discrete frequencies, not only the ISO-standard-steps. I might implement an ISO-standard-step FR which will show bars and not a curve.
Thanks, ARTA "2.3 Frequency Resolution of DFT and Octave-Band Analyzers"
if ARTA window = BLACKMANN3, then this corresponds to HOLMImpulse amplitude smoothing with N = 1, N = 3
The HOLMImpulse FR makes a calculation for all discrete frequencies, not only the ISO-standard-steps. I might implement an ISO-standard-step FR which will show bars and not a curve.