"Aliasing" a problem in MAutoDynamicEQ and MDynamicEQ?

[alexis1]

Hehe ok, actually you unintentionally found the worst scenario. Anyway I looked a little closer to your screenshots, so here's your explanation :

You generated test-tone is at about 1kHz, but you placed the peak filter at 3kHz, thus in the area where the only existing signal is some kind of noise. MAutoDynamicEq tried to adapt to the signal in that band, but since it is basically empty, it was "jumping a lot" and the minimum release time in the plugin is 0ms, while TBFix has 10ms, so it is considerably slower. The distortion seems to take place at about 1ms, so well, there you have it.

If you place the band around 1k, no distortion takes place, because there actually IS a signal. Of course we could add some "slow" switch or something, but the test scenario you created almost never exists, since it doesn't make sense - you don't want to eq something that's not there. But we'll see...

Well, this all makes perfect sense, thanks!

I do note that he said when he drops the release time down to 1 msec in the TB plugin the curve is still nearly as clean as it is at 12 msec ... but if I understand the point to be that it is an observation with little to no applicability to real-world audio (i.e., there's no audio need to ever apply a cut to digital silence at the noise floor) ... then (though the post-processing curves may be different) it won't make any difference at all to anybody's project ... someone please correct me if I'm wrong!

I think I'd like to see what the two curves look like when the cut is at the frequency of the sine wave, with similar attack/release times (preferably at the Melda "auto" values)!

Thanks -

[MeldaProduction]

Ok, already implemented . Just 1ms RMS removes the distortion almost entirely. (now I wonder why it hasn't been there all the time )

Anyway these testing signals naturally don't exist and some distortion may actually get some character. The RMS length will serve well to control "how much character you want". Though for natural signals it will most likely be inaudible, who knows for some trained ears .

[mesaone]

You generated test-tone is at about 1kHz, but you placed the peak filter at 3kHz, thus in the area where the only existing signal is some kind of noise. MAutoDynamicEq tried to adapt to the signal in that band, but since it is basically empty, it was "jumping a lot" and the minimum release time in the plugin is 0ms, while TBFix has 10ms, so it is considerably slower. The distortion seems to take place at about 1ms, so well, there you have it.

Due to the way threshold initially behaves in MAutoDynamicEQ and MDynamicEQ, could this be part of what leads to the distortion shown above?

By default, the threshold is silence. As with a conventional compressor, the implication is that dynamic processing is always active. This might seem odd at first, but it reflects how dynamic equalizers are most-often used, to provide a smooth and continuous leveling effect.

Source: http://soundbytesmag.net/meldamdynamiceqpart1/

[alexis1]

I can report that on my system, there is no distortion/artifact added by MAutoDynamicEQ to a 1kHz sine wave. I tested as did "Mads" above with dynamic EQ of -10 dB. I did 3 tests - two of them with attack/release at auto, one of the two with Q = 10, and the second with Q = 20. The third test was with Q = 20, and attack at 0.050 msec. It might be hard to see on the pics, the scale went down to -93 dBFS (0 dBFS at the top).



Each of these curves was identical to each other (superimposed completely on top of each other to the resolution of my computer monitor), and also to the unfiltered 1kHz sine wave.

The images below show that - first picture just to show the unfiltered sine wave, and the 2nd picture to show the superimposed curves of Q= 10 and Q = 20 with auto attack, and Q = 20 with attack = 0.050 msec. (Not shown is that upsampling to 4x didn't change anything either):





So, would this be considered a good test to show that there is no appreciable distortion or aliasing introduced by Melda MAutoDynamicEQ? I'm asking in all honesty, I really don't know ... I'd feel better about drawing that conclusion if there were a good explanation about why the results would it be so different on Mr. "Mads" system, as in the above post ...

Mads, if you redo your curves by putting the filter at the frequency of the tone generator, does it look/sound OK?

If this is a satisfactory test to show clean signal processing by the plugin - I am buying it as soon as it rolls around with the 40-50% off weekly sale!

Would appreciate any thoughts/comments -

Thanks!

PS Where did those 3 hours of my life just go??

[MeldaProduction]

Hehe you just spent 3 hours of your life getting some technical knowledge, you could use in the future, so it's a time well spent .

About the analyses - I can personally confirm the Mads' findings - if you use really create such a pure signal and place a band in a different place, hence drive by extremely silent signal, then it really starts distorting it due to the wast fluctuations. This won't happen in really use, but anyway it is worth "improving", so the next version with have RMS length and Hold features, and Hold will default to a few milliseconds (it seems the true hold algorithm we are using provides much smoother response than RMS and it god damn fast).

mesaone: Yes, you are right. The default threshold at "silence" lets it process the rather nonexistent signals.

[mljung]

So glad the reason was found!

My new tests made today, using a 993Hz sine tone, confirm Alexis1 findings and Vojtechs description: MAutoDynamicEq tried to adapt to the signal in that band, but since it is basically empty, it was "jumping a lot" and that apparently caused the distortion. Seems like this discussion has been very fruitful as new improvements has been made for those rare occasions where fluctuations actually do occur.

My final screenshot is made on the 993 band with 0,001 ms attack time = no distortion, no aliasing and no up-sampling needed. All good - actually very good!

I will return to make full use of the wonderful MAutoDynamicEQ again - thanks for the digging into these issues!

Very happy user here

::
Mads

[MeldaProduction]

Very well . Stay in tune for the major update

[mljung]

I'm ready

::
Mads

[MeldaProduction]

Damn, then you are faster then me

[djscorb]

Just stumbled on this thread. Interesting!

This is another reason I am glad that a threshold control was added. I never use MDynamicEQ with threshold at it's default of -inf dB as I tend to use it so it only acts on rogue peaks and resonances. I can imagine scenarios in practice where this problem might have manifested itself so I'm glad it's been sorted!

One thing I would like is an option for lookahead and hold (perhaps with a little smoothing) as It would make for better control of fast spiky signals.

What do you think Vojtech?

Cheers

Scorb

[MeldaProduction]

Hi Scorb, I'm afraid look-ahead is almost impossible (long story). But hold is already there, plus rms ...

[ericzang]

I've done the "manual lookahead" approach with it by duplicating the track, sliding it ahead, and sending it to the sidechain.

[alexis1]

I've done the "manual lookahead" approach with it by duplicating the track, sliding it ahead, and sending it to the sidechain.

When you slide, is it samples or milliseconds range you are doing?

Thanks much -

[ericzang]

I usually do it in milliseconds, as that relates to the settings of the compressor (attack, release, etc).

[alexis1]

I usually do it in milliseconds, as that relates to the settings of the compressor (attack, release, etc).

Thanks much for the info