MAnalyzer bug or user error?

[clepsydrae]

Hi -- sorry if this is asked all the time: I couldn't google anything about it.

When I send white noise into MAnalyzer (noise from Melda noise plugin or other plugins, doesn't matter), and look at the "Average (Infinite)" plot, it is not flat. (Yes, slope is at zero. :-) )

Is this some dumb user error or something wrong with the plugin?

I note that with "super resolution" enabled, the issue is pronounced. With it turned off, the issue remains, but after 30 or 60 seconds the averaging brings the plot flatter (though never quite flat.)

See these two images comparing SPAN's results to MAnalyzer. In one of the images I forgot to change the MAnalyzer settings tab to "Analysis" but the settings are the same for both: 0 slope, 0 gain, 0% time res, 0% de-harm, 4x overlap, 65k FFT, Hann windowing.

Image 1 (super res turned on): http://lacinato.com/pub/bugs/melda_superres.png
Image 2 (super res turned off): http://lacinato.com/pub/bugs/melda_nosuperres.png

Thanks for any ideas!

[MeldaProduction]

It's probably due to the log-scale conversion - basically MAnalyzer (and all our analyzers) first convert the data into log scale, which is how you "hear" things, so it's more natural. This conversion is not 100% loseless, so there may be some slight inaccuracies, mostly with noises, because , well, that would take a long time to describe . Here's how it should look:

MAnalyzer00.png

In your case it looks like wrong slope, prefiltering or something.

[clepsydrae]

Thanks for the reply! -- as mentioned, slope was at zero for those images. There is no prefiltering (it is disabled). I see the same strange slope up to 5k, then leveling off, whether using Average (Infinite) or any of the other graphs.

You can see the settings in the second image I posted above (they are the same settings used in both images). No amount of tweaking of the settings I do seems to make the slope-to-leveling-off pattern go away.

Am I missing something?

Thanks again,
-c

[clepsydrae]

Just to rule out some variables, I installed the Melda free plugin pack in a Windows XP 32bit VM and tested there. I instantiated MNoiseGenerator, went through the initial Melda config screen without changing anything, then instantiated MAnalyzer, turned off normalization, enabled super resolution mode, set slope to zero, and changed the graph to infinite average (I don't think I'm forgetting any steps). Same issue:



Incidentally, this is all in Reaper 5.27, which I mention in case it's some kind of host interaction issue.

[MeldaProduction]

Ok, so please send me this simple project to info@meldaproduction.com, I'll open it here, perhaps there's something being overlooked.

[clepsydrae]

Demo project sent. Thanks for looking in to it.

[clepsydrae]

Sorry to resurrect this thread, but was just curious if this ever got fixed? I'm a big fan of all the Melda plugins but was a little frightened to use MAnalyzer after this came up. (I haven't installed/tested a version of MAnalyzer since these posts.) Thanks!

[Alphacodex]

I would be interested in the result as well...
If there was a bug and was fixed I'm ok with that, just would like to know.
For example I used the analyser and noise generator to find (and reduce) inlinearities in the frequency response of my recording room, so a wrong slope would be pretty relevant for me...

[pumafred]

Sorry to resurrect this thread, but was just curious if this ever got fixed? I'm a big fan of all the Melda plugins but was a little frightened to use MAnalyzer after this came up. (I haven't installed/tested a version of MAnalyzer since these posts.) Thanks!

Sorry for barging in, but wouldn’t it be better to rerun the test? I may try to do so when I get a spare moment.

[pumafred]

Could this, from Wikipedia, be of help?

“with a white noise audio signal, the range of frequencies between 40 Hz and 60 Hz contains the same amount of sound power as the range between 400 Hz and 420 Hz, since both intervals are 20 Hz wide. Note that spectra are often plotted with a logarithmic frequency axis rather than a linear one, in which case equal physical widths on the printed or displayed plot do not all have the same bandwidth, with the same physical width covering more Hz at higher frequencies than at lower frequencies. In this case a white noise spectrum that is equally sampled in the logarithm of frequency (i.e., equally sampled on the X axis) will slope upwards at higher frequencies rather than being flat. However it is not unusual in practice for spectra to be calculated using linearly-spaced frequency samples but plotted on a logarithmic frequency axis, potentially leading to misunderstandings and confusion if the distinction between equally spaced linear frequency samples and equally spaced logarithmic frequency samples is not kept in mind.”

[clepsydrae]

Thanks pumafred -- the slope we're seeing above is not (to my knowledge) due to a log/vs/linear issue. A white noise signal should show flat on an FFT with no slope enabled (see the SPAN result in the demo images in the first post). And the log/vs/linear issue wouldn't have the plateau that we're seeing. I will indeed run the test again with the latest version if I get some time to do so; hopefully the dev(s) could chime in if they have a quick answer, though.

The "error" (if it is one) in the graph seems to be present whatever the settings (super resolution on/off, window size, windowing type, averaging type, etc.) But super-resolution does seem to increase the issue. (Or rather, did, in the old 10.04 version this was tested on.)

[MeldaProduction]

Did we discuss this over email? If so, what was the end result?

[clepsydrae]

D'oh, we did indeed! I forgot! Here's the exchange:

MeldaProduction: "I checked and it's the super-resolution mode. It's not such a big surprise for a noise to have such a behaviour with it, this mode is really designed for "normal" signals as opposed to analytical signals, which often require a little different handling."

Me: "Thanks -- turning it off flattens the graph with that FFT size; but if I open that demo project again, turn off super resolution, and change the FFT size to 65536, I still see a slope (slightly different shape -- no kink at 5k) with a vertical difference end-to-end of about 2dB. If I change the FFT to 262144, the slope is almost 5dB... ?"

MeldaProduction: "aaaha, indeed. Try increasing Time resolution. Basically this minor slope is caused by the log-scale transformation, which doesn't assume the densely populated energy in high part of the spectrum. Time resolution fights that by processing smaller blocks hence minimizing the probability of near spectral bins to be populated. Anyway the analyzer is just made with real audio signals in mind, there are plenty of scientific analyzers after all , which are far easier to program, but the nonlinearity of human hearing pose a huge problem in several aspects of further visualisation with scientific methods. Human ears are just not mathematical ."

...so it would sound from this like MAnalyzer maybe shouldn't be used for scientific applications, but has some advantages in real-world use?

[DocSnyder]

So noise is considered an analytical signal here? What are examples for other analytical signals? What about a nice hihat, the slope bug will affect it too, right? And what exactly are the advantages of this "musical analyzer"? Smells fishy.

[MeldaProduction]

clepsydrae, ok, yes, the super-resolution mode is NOT intended for scientific purposes. It's for purely musical signals, since it lets you get the nearly ideal time response - with normal analyzers you can either make them slow and accurate in bass or fast and inaccurate in bass, that's the uglyness of spectral processing. The super resolution mode gets you accurate and fast at the same time. Useful for musical signals, where bass naturally is slow and treble is fast. But there may be inaccuracies caused by the rather overly complex transformation.