Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Merlijn van Veen]

Taking it bit further, I took the liberty of measuring a convolution reverb to get some more realistic data to demonstrate the differences between smoothing settings. The first picture shows six instances of the same trace at different settings, no offsets applied. Notice the difference in level.

Unfortunately for me, which kinda sucks, there's no notable difference in slope or roll-off for the high frequencies regardless of the amount of smoothing applied. Apologies for apparently talking BS.

The second picture shows the same traces side by side.

The third picture is an attempt to best fit the 1/12 8ve trace on top of the 1/48 8ve trace. I had to use an offset of +3 dB to get the best fit. Still at 31.5 Hz, 125 Hz, and north of 4 kHz inconsistencies of 2-3 dB's remain.

The last picture is an attempt to best fit the 1/3 8ve trace on top of the 1/48 8ve trace. It took three attempts at +7, +5 and +3 dB offsets to respectively match the lows, mids and highs. 4 dB of variance within one trace is to much IMO.

And now without further ado, back to the OP's question. Sorry for the brief sidetrack.

[Doug Fowler]

Taking it bit further, I took the liberty of measurking a convolution reverb to get some more realistic data to demonstrate the differences between smoothing settings. The first picture shows six instances of the same trace at different settings, no offsets applied. Notice the difference in level.

Unfortunately for me, which kinda sucks, there's no notable difference in slope or roll-off for the high frequencies regardless of the amount of smoothing applied. Apologies for apparently talking BS.

The second picture shows the same traces side by side.

The third picture is an attempt to best fit the 1/12 8ve trace on top of the 1/48 8ve trace. I had to use an offset of +3 dB to get the best fit. Still at 31.5 Hz, 125 Hz, and north of 4 Khz inconsistencies of 2-3 dB's remain.

The last picture is an attempt to best fit the 1/3 8ve trace on top of the 1/48 8ve trace. It took three attempts at +7, +5 and +3 dB offsets to respectively match the lows, mids and highs. 4 dB of variance within one trace is to much IMO.

And no without further ado, back to the OP's question. Sorry for the brief sidetrack.

Merlijn -

Can you apply various windowing parameters to the unsmoothed measurement as a demonstration of its effect?

[Merlijn van Veen]

Merlijn -

Can you apply various windowing parameters to the unsmoothed measurement as a demonstration of its effect?

Do you mean by parameters, window functions? I've used SMAART 7 in MTW mode to create these images. Window functions in this version can only be set for spectrum measurements. I'm unsure if these settings also affect transfer functions. I'll look into it. The window function found in SMAART 5, is AFAIK not featured in the current version.

P.S.: thanks for the typo

[Doug Fowler]

Do you mean by parameters, window functions? I've used SMAART 7 in MTW mode to create these images. Window functions in this version can only be set for spectrum measurements. I'm unsure if these settings also affect transfer functions. I'll look into it. The window function found in SMAART 5, is AFAIK not featured in the current version.

P.S.: thanks for the typo

What is the format of the convolution file?  If I can open it in SysTune I can demonstrate this. 

[Paul Tucci]

Do you mean by parameters, window functions? I've used SMAART 7 in MTW mode to create these images. .... The window function found in SMAART 5, is AFAIK not featured in the current version.

P.S.: thanks for the typo

Although not visible on the immediate GUI, the time window parameters Doug is asking about do in fact still exist in v7. They are underneath, inside of Measurement Configurations. The MTW is but one choice. The other different sized FFT measurements will correspond to differing time windows. The larger FFT sizes will look at data longer, yielding better resolution in the low frequencies AT THE COST of too much info in the higher freqs, thereby wasting CPU power. MTW is a good place to set that parametewr and forget it UNTIL you want to measure semi-anechoically.

PT

[Merlijn van Veen]

Merlijn -

Can you apply various windowing parameters to the unsmoothed measurement as a demonstration of its effect?

Although not visible on the immediate GUI, the time window parameters Doug is asking about do in fact still exist in v7. They are underneath, inside of Measurement Configurations. The MTW is but one choice. The other different sized FFT measurements will correspond to differing time windows. The larger FFT sizes will look at data longer, yielding better resolution in the low frequencies AT THE COST of too much info in the higher freqs, thereby wasting CPU power. MTW is a good place to set that parametewr and forget it UNTIL you want to measure semi-anechoically.

PT

Doug & Paul,

It took some time for me to understand what you meant. But I believe this is what you'd like to see.

SMAART 7 currently has 9 FFT sizes to choose from plus MTW. Ten instances of the transfer function (the convolution reverb from my previous post) in one picture was to cluttered, so I'll show them side by side going from small to large FFT size. I labeled each FFT size with frequency resolution and the corresponding time constant that goes with it.

Take notice of the pattern, it's power of 2. Each doubling of the FFT size, halves the frequency resolution but doubles the time constant. At smaller FFT sizes the traces appear to be incomplete, this is not the case however. Take the first trace for example, FFT size 128, frequency resolution 375 Hz. The first data point is at 375 Hz, the second at 750 Hz the third at 1.125 Hz, etc. etc.. There's no data below that first point, this is the way of FFT or more correctly DFT (Discrete Fourier Transform). FFT (Fast Fourier Transform) are algorithms to reduce the arithmetical steps in order to save processing power without loss of information. In case of FFT size 256, the first data point is at 187,5 Hz or one octave lower in comparison to the former FFT size and this fashion repeats itself at large FFT sizes.

Evidently you need larger FFT sizes, as Paul correctly mentioned, in order to measure LF. By the time you reach FFT size 32k you end up having 1,5 Hz resolution, which suffices for these frequencies. However on the other end of the spectrum, using the same FFT size, you'll have 6.666 data points from 10 kHz to 20 kHz. This is superfluous and makes for cluttered transfer functions.

It's starting to become clear that one FFT size for the entire spectrum won't work, there's no magic bullet. This realization lead to quasi-logarithmic scales, FFT being linear, comprised of more than one FFT size like. The tradeoff is resolution vs. time. The first is quit easy. Most people agree that 48 points per octave (48 PPO) suffices to identify most problems and act accordingly. The latter, time, is more difficult.

This has to to do with the way we hear things, tonal, spatial or echos. Clearly you don't want to equalize echos, so its beneficial to keep them out of the equation and deal with them in a more fitting way. Every time a dual-FFT analyzer measures a data point, the "door" stays open for a certain amount of time, allowing reverberation and late arrivals from other speakers to creep into the measurement. The only way of keeping this from happening, is closing the "door" earlier. Hence shorter time constants allow for less of these influences at the cost of frequency resolution and vice versa.

How many FFT's of what size and time constant is what distincts the various dual-FFT analyzers rom each other. The decision is made by the developer, based on their experience end opinions of psycho-acoustical phenomena.

The final trace in the picture below shows MTW (Multi Time Window). AFAICT, Rational chose to divide the spectrum in 6 parts, each with their own FFT. Striking a balance between frequency resolution and time. I've labeled these frequency resolutions and time constants as far as I'm able to identify them. The actual FFT sizes are part of a proprietary process which is beyond me and has to do with "downsampling" if I'm not mistaken.

The same measurement, over and over again, is greatly influenced by the FFT size chosen as Doug mentioned in a previous reply. Both in terms of level and accuracy.

I stand by Paul's recommendation to stick with the default of your dual-FFT analyzer, unless you know what you're doing.

[Merlijn van Veen]

In conclusion,

Smoothing and FFT size greatly influence the way we are able to "see" what we hear, using dual-FFT analyzers. The first is a post-processing step applied to the latter after the fact, the original data. A visual filter if you will, regardless of how well it's being done.

I've included one final picture, trying to demonstrate the smoothing process in a graphic way, being less abstract to most of us including myself. The top left is two instances of Mona displaced. The top right is Mona and something completely different. Below that is a section of barcode and a gradient ramp.

As smoothing increases things start to blend up to the point where everything is one big blur. Most notable is the barcode where the clear contrast between black and white, dips and peaks, is reduced to grey. More smoothing might make for a trace that's more "readable" but moves you further from the truth until everything is grey.

Doug, is it possible to somehow split this thread in two and replace the OP's question in it's original forum?

It feels like I hijacked it and turned it into a measurement lecture especialy now that it has been moved. The topic is to interesting and might miss out on replies of the general visitors sitting here in the measurement section.

I did feel however, to point out the dangers of smoothing and how this may visually "warp" what we hear. I'm able to tell tonal differences by ear, but I'm not able to translate that information into a slope or roll-off and stick a number to it as was requested by the OP. Most people must have come to a conclusion about these figures on there preferred tuning by using some sort of analyzer. All I intended to do is make sure that this "visual" assumption correlates to what's heared and preferred.

[Merlijn van Veen]

What is the format of the convolution file?  If I can open it in SysTune I can demonstrate this.

Doug,

The reverb is a plug-in. If my last posts don't answer your question, I can give the convolution file a try. All the data in my replies is created from within my MacBook by means of Soundflower, AUlab and SMAART. If you're interested, I can PM you how to do this or open up a new thread explaining the process (if doesn't already exist). I know of similar setups in a Windows environment, maybe it's also possible with Systune.

Sincerly,

Merlijn

[Merlijn van Veen]

I find the console output EQ method much more efficient. I can raise frequencies up to something like 160Hz, while raising the processor output only works up to the 100Hz crossover frequency.

+1

I'd like to add the input EQ of your processor as an alternative or insert a pair of graphics.

This way, the relationship between mains and subs stays unaltered for various reasons that are off topic.

[Arthur Skudra]

The reverb is a plug-in. If my last posts don't answer your question, I can give the convolution file a try. All the data in my replies is created from within my MacBook by means of Soundflower, AUlab and SMAART. If you're interested, I can PM you how to do this or open up a new thread explaining the process (if doesn't already exist). I know of similar setups in a Windows environment, maybe it's also possible with Systune.

Hi Merlijn,

So beautifully illustrated!  Wow!

I would love to have a link to get that convolving reverb plugin myself.  I use Soundflower & AULab with Smaart as well, a great educational tool!  Which version of OSX are you running?

Arthur

[ProSoundWeb Community]