subsonics ... and above

This is about the results and effects (or consequences) of subsonic filtering on LP recordings. I have very little background with filters, but what I think I know doesn't seem to explain some of the results I see.

All my LP recordings, and the few from other sources I have been able to look at, have a great deal of very low frequency. This is observable in the Frequency Analysis graph, especially using a large FFT. Purposely unmodulated areas, such as between tracks, show a steadily raising level from the highest to the very lowest frequencies. Even some areas of moderately loud music have a subsonic level that is much higher than any of the music frequencies. I believe I understand enough about where this energy comes from; that is not the object of this inquiry.

Tests were carried out on a 6.5 minute track from an LP recording. There is thus the unprocessed file (fileA), the filtered file (fileB), and the removed material (fileC). FileC is obtained by doing a Mix-Paste Inverted of the filtered file into the unprocessed file -- or vice versa.

(1)  CoolEdit's Scientific Filters has a preset labeled "remove subsonic rumble." The Sonic Foundry NR plugin, Click and Crackle Removal section, has "remove low frequency rumble" turned on by default. CE's preset is an 18th order Butterworth filter with a cutoff of 28Hz. The SF processing is not explicitly specified, except that the Help file says it removes frequencies below 30Hz. The SF filtering can be applied with no click and crackle removal by setting the three control sliders to their minimum positions.

(2) FileB RMS stats, (Analysis/Statistics), are little different than fileA stats, except for Minimum RMS Power. This value was 8dB lower using CE and 4dB lower using SF. The other fileB RMS values were within about 0.1dB of fileA.

(3)  The two fileBs look quite different in the lower frequencies on the Frequency Analysis graph, especially 30Hz and below. The CE processed fileB has a markedly steeper slope below 30Hz than the SF one. My first guess would be that this indicates the SF processing uses a lower order filter, but not knowing a great deal about filters, I wouldn't want to venture too much money on that interpretation.

(4)  FileA and fileB look very much the same to me in Spectral view, but there is a difference in overall appearance in Waveform view that is hard for me to characterize. Some places have higher amplitude, some have lower, sometimes there is a shift in overall weight above or below the minus infinity center line.

(5) The two FileCs, from both CE and SF, are quite different than I expected. They look reasonably like fileA, but at a somewhat  reduced amplitude. Spectral view also looks markedly similar to the original, but sort-of shrunken. There are some greater differences between fileC from CE vs SF than between fileB for the two programs.

(6)  The average RMS power of fileC is reduced, compared to the fileA, by 5.35dB in the SF version, but only by 1.14dB in the CE version. There seems to be too much of fileC to be accounted for by fileB's small decrease in RMS power values from fileA. When I play the two fileCs, they both sound enough like fileA to be easily recognized. The CE version sounds rather nearer the original, while the SF version seems more bass heavy. Both are missing much of the higher frequencies, such as the drum set's cymbals in this jazz track.

(7) For fileC, zooming in on a Spectral View peak, I saw 18kHz in the CE version and 16kHz in the SF version. This peak reached the top of the display in fileA. I conclude that FileC contains much of the higher frequencies of fileA, only at a reduced amplitude.  If I set the Range for Spectral view to 200dB, the higher frequencies look very close to fileA (viewing fileA at the default range of 120dB), but frequencies below 2kHz are brighter in fileC.

Perhaps there are so many unknown factors between these two processes that my first question is meaningless but ... Being that the SF processing effects very low frequencies less severely than the CE subsonic filtering does (see (3)), is it consistent that it also effects the highest frequencies less (see(6) and (7))?

The graph used to display the filter's effect in CE's Scientific Filters Transform window seems to indicate that essentially nothing is effected above about 30 Hz, except for phase. That seems inconsistent with the results of the Mix-Paste Inverted. Clearly some of nearly all frequencies are removed from fileA (see(6) and (7)). Should the reduction in level above the cutoff frequency be linear over frequency? If this is so, then nothing is lost except amplitude, and that loss is restored when the file is normalized. Yes, No?

Should it be possible to get something from fileC very near to fileA by doing some EQ on FileC?

I know that some people claim to hear undesirable effects of subsonic filtering. Either I can't, or I don't know what to listen to. I've used subsonic filtering mainly because of its potential effect on amplifier and speakers. However, considering what I see removed by both CE and SF, as revealed by fileC, I wonder if some result should be audible throughout most of the audio range. If people who claim to be able to notice a difference really do notice a difference, is that likely due only to the very low frequencies below and near the cutoff, or to some effect on the range of frequencies more easily heard by most people?

Changing CE's subsonic rumble removal preset from 18th order to 1st order produces a markedly lower amplitude fileC, but it also does not reduce subsonics very much. FileC from this experiment still sounds passably like the original file, once one turns up the volume significantly.

I've previously acknowledged that I'm a bit slow. About midnight, as I was preparing dinner, it occurred to me that the factor that produced fileC was somewhat, perhaps mainly, a secondary effect of the filter -- the phase changes. The filtering process changes the phase at higher frequencies, so fileB is different than fileA, regardless of what the filter removed or did not remove. Yes? No?

I'm not really going to try to figure out what you did - this sounds ridiculously complicated...

But the higher the order of filter, the greater the phase change at the turnover point - and an 18th order filter is actually a pretty dodgy thing to use in-band. Primarily, the effect of the very rapid rate of change of response at the turnover point is that any frequency adjacent to this is likely to cause the filter to ring, and can easily give rise to an amplitude increase in some signals that you would have thought should have been attenuated. You can easily get the AA notch filter to ring, for instance - we've covered this before.

And digital filtering is by no means perfect at lower frequencies, either - this has caught out several people. It take a while to explain why, but it's all centred around having a lack of material to analyse - and the bottom line is that you are rather better off in response terms by using a rather lower roll-off rate (like no more than 12dB/oct) at about 40Hz and gently rolling off towards zero at this rate rather than using a brick wall and wondering why there are all sorts of anomalies. Also IIR filters are recursive - technically that's what causes the ringing effects - and quite frankly, I'd rather have a FIR filter for this sort of job. But an 18th order filter is crazy - this is going to put 18 poles in the response somewhere, and I really can't see the point of anything attempting to be that steep, especially at LF in a digital filter. This seems like asking for trouble!

My very-occasionally-used turntable rumbles like a cart - I seem to get good results using noise reduction (carefully tweaked to affect only the lowest frequencies) rather than filters.  FWIW.

I believe I understand the recommendation, but what about this alternative? Is there a reason it should be a less desirable solution?

The FFT filter can be set to attenuate with a virtually infinite slope , say cutoff at 30Hz, 30dB attenuation (or maybe 20Hz, to preserve the that deep bass on the rare recording that has any). By the tests I can make, the FFT filter  has virtually zero effect above 30 Hz, while a second order Butterworth, cutoff at 40Hz, has measurable effect over the entire audio range.

My previous questions may have been complicated, or at least messy, but what I did was very simple.
(1) apply filter to file
)2) obtain difference between original file and filtered file with Mix-Paste Inverted

This procedure is my evidence that the FFT filter reduces everything by 30dB below the cutoff but has no effect above the cuttoff frequency, in comparison with the Scientific filter's lesser subsonic reduction (for 2nd order) and considerable effect on higher frequencies (my guess is an effect on phase rather than on amplitude).

My turntable has no audible rumble. Being direct drive, anything it makes is subsonic, and I don't think there is much of that. I've often used NR restricted to lower frequencies, but it effect on subsonics is much less than the filters. However, some experiments just completed, adjusting the NR graph to a straight-line cutoff at 30 Hz, using 100%NR reduced by 40dB, I can get a greater subsonic reduction than with the FFT filter, which itself is greater than the second order butterworth.

The result sounds fine, as do the other approaches, and the Mix-Paste Inverted shows little effect much above the 30Hz cutoff. So, from a theoretical viewpoint, is there any reason to choose one of these three (Butterworth, FFT, NR) over the others?

On the face of it, I'd go for the NR solution. If you get a good sample, then as a subtractive process it will be removing what you actually want to remove from the particular source you've got, and I much prefer this methodology rather than the blanket removal you get with LP filtering, however you do it.

I ssuspect that you are correct about the phase changes making a difference to the Mix-paste inverted results - all of the different filter types have different phase characteristics.

But ultimately, you have to trust your ears - bottom line is that if it sounds fine when you've finished, and looks like a sensible signal with no silly out-of-band noise profiles, etc then quite frankly, does it really matter how you achieved it, as long as you could do it again if you had to? By and large, if you understand the implications of these processes, you are more likely to make a sensible decision about what to do anyway, I think.

It's a wood, not just a collection of trees...

Coming home from a show in Bellevue tonight I saw a license plate that said simply "Andy H".
Thought that might help...

Haha!  Cool.   But is our Andy H from Washington?  

In retrospect, this old thread still is pretty helpful if a bit oversimplified.  But what it does is classify which filters are IIR and which are FIR -- see David Johnston's post fifth from the bottom.

Thread:
http://audiomastersforum.net/synforum/viewtopic.php?t=11844

For rumble I'd probably go for NR also, but if not it would be good old FFT filter.

My standard practice is to use the subsonic rumble filter in Sonic Foundry's 2.0NR declicker plug-in, and a dash of general across-the-range NR from Waves X-Restoration. If it needs more than this then the AA noise reduction filter does a brilliant job, set to target the problem lower frequency range alone.

It does seem to depend on the vinyl quality though - some LPs come through my normal processing as clean as a whistle, with inter-track levels somewhere down around -72dB to -75dB on the AA meters, others never get near this even with the AA NR applied, sitting around -60dB to -65dB.