Skip to main content

Notice

Please note that most of the software linked on this forum is likely to be safe to use. If you are unsure, feel free to ask in the relevant topics, or send a private message to an administrator or moderator. To help curb the problems of false positives, or in the event that you do find actual malware, you can contribute through the article linked here.
Topic: How Much Do High Frequencies Matter In Practice? (Read 3470 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

How Much Do High Frequencies Matter In Practice?

Lossy codecs such as Lame and FhG AAC cut high frequency content (>15-17 KHz) when encoding at moderate bitrates.  I'm 34 and can easily hear pure, near-full-scale tones at 17 KHz and normal listening volumes, so I had always assumed that this filtering was a major source of any audible artifacts at these bitrates.  I decided to test this hypothesis.

To my surprise, I could not ABX the original FLAC of a song with lots of cymbals and distorted guitars against either a FLAC filtered at 15.6 KHz or a 128 kbps FhG AAC, which cuts at about this frequency.  I therefore tried to create an even more-sensitive test to determine whether I might plausibly be able to hear the difference if I listened carefully enough for the right artifacts.

I took the original FLAC and highpass filtered it at 15.6 KHz to hear only the sounds that would normally be filtered out, without masking from lower frequencies.  At normal listening volumes, it sounds like complete silence to me.  I don't think I'd be able to easily ABX it from complete silence without listening at volumes that would be unpleasantly loud on normal music.  My conclusion is that on most real world music at real world listening volumes, a 15.6 KHz lowpass filter is very unlikely to be audible because high-frequency content in real music is both much lower amplitude and much broader spectrum than pure tones.

I've attached short samples of the original song ("Westbound Sign" by Green Day) and the highpass filtered version for your entertainment.

Re: How Much Do High Frequencies Matter In Practice?

Reply #1
The answer is most likely in your conclusion, ie "on most real world music at real world listening volumes, a 15.6 KHz lowpass filter is very unlikely to be audible because high-frequency content in real music is both much lower amplitude and much broader spectrum than pure tones."

It is not only the amplitude of these frequencies in the recording, but also the decreasing sensitivity of human hearing outside the midranges.  While you might hear 17khz + frequencies with test tones, the sound of music at frequencies where our ears are most sensitive masks the content of higher frequencies.

Re: How Much Do High Frequencies Matter In Practice?

Reply #2
If you are speaking of digital equipment -- it is very important to avoid a brickwall filter in the region where there is significant audio energy.  This importance of avoiding the brickwall is independent of hearing up to the point of the frequency sensitivity of the input of subsequent equipment, and dependent of hearing in the range where the effects of the brickwall can be detected.  The reason for avoiding the brickwall filter in the region of significant audio energy is to avoid the negative consequences of Gibbs effect.  (That is, the 'ringing' that happens because of the truncation of the spectrum.  It is a mathematical necessity -- no way other than nonlinear techniques to get rid of that.)

So, basically a 22kHz brickwall is okay if there is little audio signal energy in that range -- because you cannot hear the difference anyway, and subseqent equipment wont be  affected by Gibbs phenomenon.  However, if there is significant signal energy in the 22kHz region, then Gibbs effect WILL happen, and even if your hearing cannot detect it, subsequent equipment might be affected and the distortions resulting from the Gibbs effect might be audible.

If we are talking about a lower frequency brickwall, and there is significant signal energy -- perhaps a frequency that can be heard, it is possible to hear some of the consequences of Gibbs and also subsequent equipment might be affected.

As done in video, the trick is to do a slow rolloff (e.g. Thompson/Bessel filter, or ANALOG filter with Q=0.500 for 2nd order) which maintains enough of the higher frequencies to avoid ringning by the filter alone, and then brickwall (chop) the frequences where the energy is low enough.  (Resonance in the filter is also best avoided -- because that can also avoid the high peaks causing distortion or audible effects -- but that is looking at fitlering in different way, and also part of the intuitive reason for using Q=0.500.)

This is one reason why one might want to use high sample rates -- in case there is significant energy at higher frequencies, and do a slow rolloff and brickwall when the energy is low enough.

Sometimes there is little that one can do about the situation, and in that case it MIGHT be necessary in some instances to have a negative effect on signal in the audible region OR make sure subsequent equipment in the chain is not affected by Gibbs phenomenon.   Nonlinear techniques might help, and sometimes just making sure that the signal doesn't have a lot of energy at 22kHz (for 44.1kHz sampling) will keep the negative effects of a linear phase brickwall from causing trouble.

Also, when brickwalling (i was originally above just thinking about sine wave signals -- if you have a 10kHz square wave -- that comprises 30kHz, 50kHz, etc -- so if you brickwall at 22kHz, then there will be Gibbs from 10kHz square wave also.)  As I mentioned above, the key is a slow rolloff until the signal level is low enough that the Gibbs is small enough that the Gibbs ringing will be small enough not to negatively affect subsequent equipment.  In the real world, a 10kHz audio square wave doesn't happen much -- unless using a signal generator.  Some of the worries about a 22kHz brickwall chopping off audio frequencies dont' manifest because the original signal probably doesn't have a huge amount of energy at 22kHz.  Using mikes with 50kHz bandwidth CAN have negative consequences...

So -- the above is one of the problems when trying to mess around too much with 'high frequencies' in the digital realm.  This whole problem is WELL UNDERSTOOD by people dealing with video design, but audio people sometimes seem to forget about the VERY BASIC troubles when dealing with high frequencies -- EVEN IN THE BEST CASES.

John


Re: How Much Do High Frequencies Matter In Practice?

Reply #4
Examples are always welcome.
I cannot really give examples -- I'll try to explain the math:

The evil thing about Gibbs isn't that there is ringing (if it isn't in the hearing band), it is that it can cause nonlinearity in subsequent audio equipment -- and often you CAN hear that nonlinearity produced when audio equiopment encounters energy outside of the expected band.

First -- note that a square wave is has the following energy: 2/pi * (fund + thirdharm/3 + fifthharm/5 +...)
if you miss any of those harmonics -- the square wave is no longer square.

Basically, think of it like this -- if you have a square wave, and chop off the bandwidth, then you get ringing because of the left over sines that are missing.   Gibbs is basically 'missing constituent sinewaves.'  It takes ALL of the constituent sine waves to make the ripple (ringing) go away.  The effect is called 'Gibbs phenomena'.  If you just have sines to begin with (which I didn't make it super clear), then the brickwall doesn't cause ringing -- but audio very seldom has just a sine wave.


If you start with bandwidth limited material to begin with, and you use the same sample rate and have no energy in the region where the brickwall occurs -- then no trouble.
The big evil is when you have all of those missing sinewaves in a waveform that expects them.
So, if you start with material that is already sampled at 44.1k, then convert to analog (only do linear processing), and then convert back to digital -- there shouldn't be much problem.
The domain where I work, I am doing NONLINEAR processing, which means that I must deal with such truncation of bandwidth.

John

Re: How Much Do High Frequencies Matter In Practice?

Reply #5
This reads like quoted out of marketing brochures from the last 20 years telling us why we need HighBitrate music.
Is troll-adiposity coming from feederism?
With 24bit music you can listen to silence much louder!

Re: How Much Do High Frequencies Matter In Practice?

Reply #6
This reads like quoted out of marketing brochures from the last 20 years telling us why we need HighBitrate music.
The only reason why you need high bitrate to process the audio (linearly) is if there is energy 'up there' to begin with.  No energy up there, then no reason to worry.  Sometimes, you can just nuke that energy also (if it is coherent) -- then also no-biggie.  If something is already brickwalled and you are dealing with linear equipment, then the damage is already done (no energy.)

The big problem is when you chop the bandwidth containing energy, and you cause the Gibbs effect.  If there is no energy, then no Gibbs -- simple as that.  If you start with a theoretical square wave (having infinite bandwidth), then every time you chop you'll get Gibbs.  Real world square waves have limited slew, so the spectrum is limited, and there is some frequency where the Gibbs will diminish to practically zero (if brickwalled at that frequency.)

Do you need the high frequencies about 20k or so to hear EVERYTHING:  no... Period.  If you have material where you chopped of the bandwidth, and there is energy that got chopped off (constituent of part of the signal), then Mr Gibbs can start bugging you.  (Perhaps causing subsequent overload or nonlinearity in equipment.)

It really is simple to understand.  Any idea that chopping off the spectrum somehow naturally is hearable (except certain very limited cases) is wrong.  What happens is that Gibbs 'ringing' can cause problems in subsequent equipment (mostly if analog) that CAN be audible.

In the case of my nonlinear digital processing -- it can also produce a non-typical audio spectrum (old DolbyAs were really bad about that), and such issues need to be dealt with.  In that case, you can start with a spectrum that is limited to 20k, and end up with a spectrum that goes 'way up there'.   That is only due to the nonlinear processing.

John

Re: How Much Do High Frequencies Matter In Practice?

Reply #7
This reads like quoted out of marketing brochures from the last 20 years telling us why we need HighBitrate music.

Ironically, this experiment made me even more skeptical than I previously was of high-res (>CD quality) music by showing that even a 44.1 KHz sample rate is borderline-overkill on real music.  At least in a theoretical world of perfect resampling and perfect brickwall filters, a 32 KHz sample rate is plenty if frequencies >16 KHz are inaudible in real-world music and listening conditions.

Maybe I should try a similar experiment where I listen to the low order bits of 16-bit audio and see how many would be needed for audibility at normal gain levels.  The logic would be the same:  If a signal sounds like silence when not masked by other sounds, it's very unlikely you'll be able to ABX its presence/absence even with extreme training/concentration when it is partially masked by other sounds.  (Obviously the inverse isn't true, though.)

 

Re: How Much Do High Frequencies Matter In Practice?

Reply #8
This reads like quoted out of marketing brochures from the last 20 years telling us why we need HighBitrate music.

Ironically, this experiment made me even more skeptical than I previously was of high-res (>CD quality) music by showing that even a 44.1 KHz sample rate is borderline-overkill on real music.  At least in a theoretical world of perfect resampling and perfect brickwall filters, a 32 KHz sample rate is plenty if frequencies >16 KHz are inaudible in real-world music and listening conditions.

Maybe I should try a similar experiment where I listen to the low order bits of 16-bit audio and see how many would be needed for audibility at normal gain levels.  The logic would be the same:  If a signal sounds like silence when not masked by other sounds, it's very unlikely you'll be able to ABX its presence/absence even with extreme training/concentration when it is partially masked by other sounds.  (Obviously the inverse isn't true, though.)

The only REAL problem most of the time is if the GIbbs effect occurs and causes trouble in subsequent equipment.  A good all digital example (but not practical) is if you have a square wave within about a dB or so of full scale clipping ( forget the exact amoun), do a brickwall where the spectrum is significant, then there can be clipping due ot Gibbs.  But, in an all linear digital world where the Gibbs is handled correctly (there ARE ways to control it), the usual most serious 'electronics' problem would be clipping.  If you are doing dynamics processing -- then Gibbs CAN be a problem (not always, esp RMS processing done correctly.)
The most probable problem in the audio world is mixing digital and analog -- then the spectre of 'brickwalling' significant energy can cause interesting effects.


John

Re: How Much Do High Frequencies Matter In Practice?

Reply #9
The only REAL problem most of the time is if the GIbbs effect occurs and causes trouble in subsequent equipment.  A good all digital example (but not practical) is if you have a square wave within about a dB or so of full scale clipping ( forget the exact amoun), do a brickwall where the spectrum is significant, then there can be clipping due ot Gibbs.  But, in an all linear digital world where the Gibbs is handled correctly (there ARE ways to control it), the usual most serious 'electronics' problem would be clipping.  If you are doing dynamics processing -- then Gibbs CAN be a problem (not always, esp RMS processing done correctly.)
The most probable problem in the audio world is mixing digital and analog -- then the spectre of 'brickwalling' significant energy can cause interesting effects.


John

In theory, brickwall filtering of a full-scale signal can cause clipping, but in theory almost any transformation to a full-scale signal can cause clipping.  This is part of why audio editing is typically done in 24-bit or 32-bit float even if the final result will be exported in 16-bit.  With these extra bits, one can allow a lot of headroom for transformations that increase peak-to-peak amplitude and also keep the noise floor below the -96 dB of the final 16-bit result.  One can similarly solve these problems in the analog domain by allowing sufficient headroom.

The amplitude of the Gibbs phenomenon is 1.089 (+0.38 dB) relative to the square wave's amplitude.  This means that one needs less than 1 dB of headroom to completely eliminate this issue.

Additionally, any clipping will manifest in the frequency domain as harmonic distortion.  This harmonic distortion is going to be miniscule.  I created a sine wave in Audacity and amplified it to the aforementioned +0.38 dBFS.  This is equivalent to taking only the first harmonic of a square wave at full scale.  The loudest harmonic was at -39 dBFS.  Chances are slim that this would be audible when masked by any other audio, and this is a worst-case scenario.

Re: How Much Do High Frequencies Matter In Practice?

Reply #10
I'll start with the point that I fully comprehend the Gibbs effect.  I understand the math.  I've created it; I've seen it .

What happens is that Gibbs 'ringing' can cause problems in subsequent equipment (mostly if analog) that CAN be audible.
And here is where the rubber meets the road.  An example of audible problems will speak volumes.

Re: How Much Do High Frequencies Matter In Practice?

Reply #11
The problem I see here is that audiophools are going to use comments in this discussion to justify the need for hi-res as a delivery format.

We've seen tinnito(?)phobia used a lot as justification  Sometimes 10+ pages when participants are afflicted with Dunning-Kruger.

Re: How Much Do High Frequencies Matter In Practice?

Reply #12
The problem I see here is that audiophools are going to use comments in this discussion to justify the need for hi-res as a delivery format.

We've seen tinnito(?)phobia used a lot as justification  Sometimes 10+ pages when participants are afflicted with Dunning-Kruger.

Archimago's Musings has shown that actual ringing in CD's is rare on modern equipment.  If it's likely to happen, it will happen in highly compressed, loud, poorly mastered music, which is not aiming for high fidelity anyways.

http://archimago.blogspot.com/2018/01/audiophile-myth-260-detestable-digital.html


Re: How Much Do High Frequencies Matter In Practice?

Reply #13
And the frequency of this ringing?

EDIT:
I didn’t think about whether clipping was created in the digital domain. This would result in imaging. This is different with an analog signal that was driven into clipping prior to digitization since it will be bandwidth limited at the converter in order to prevent aliasing.

Regardless, this doesn’t justify a need for “hi-res” media.

Re: How Much Do High Frequencies Matter In Practice?

Reply #14
The problem I see here is that audiophools are going to use comments in this discussion to justify the need for hi-res as a delivery format.

We've seen tinnito(?)phobia used a lot as justification  Sometimes 10+ pages when participants are afflicted with Dunning-Kruger.

Bingo (I agree.)  There ARE cases where higher resolution and/or sample rate are needed, but delivery isn't one of those cases.

For example, in a linear system -- if you start with 10kHz -- all you ever get is 10kHz -- not the harmonics unless you consider distortion (nonlinear.)  Likewise, if you start with 22kHz, that is all one gets -- 22kHz. If you start with 44.1kHz, the widest audio bandwidth will be 22.050kHz -- period.  So, if you start with a 44.1kHz sample rate, the issues of Gibbs effect are null unless there is a consumer sample rate conversion downward...

On the other hand, I am doing processing where I am multiplying a signal by a gain which varies sometimes fairly quickly.  So, if you start with 15kHz and have a gain control signal with a spectrum that can include 8000Hz, then there ARE 23kHz modulation products produced -- also 7kHz -- PERIOD.  That means that if I am working in the 44.1kHz sample rate world, then the result of the gain control goes above 22.050kHz, and therefore, I'll end up with a very audible beat frequency because of aliasing.  Such 'beat' frequencies or aliasing sound UGLY.
So, in my case -- for that specific issue, at least 48kHz is needed (24kHz signal handling capability) to avoid aliasing.  The aliasing at 44.1kHz WILL happen when the multiply (the gain change) occurs, and will NOT be fixed by any post-filtering (I do have some tricks that will minimize the problem, but it MUST still happen.)  My scheme suppresses undesired modulation products -- but must keep those necessary to effect the gain control -- very tricky, but works wonders.  Even in my scheme, aliasing must still be considered (and hopefully avoided.)

When we will suggest the sample rates for the DHNRDS decoder, it will be for 64kHz on up to 192kHz.  64kHz is a good minimum and the range between 64kHz through 96kHz are the optimum for various reasons.

BTW -- when I mentioned 8kHz as the maximum spectrum for the gain control -- it actually goes higher, but is very low level, and such aliasing because of 8kHz range isn't really all that great -- but still exists and is still ugly sounding.

Such dynamic gain control and other processing DOES occur in production, and there is BIG advantage to use 96kHz sample rate or above.  Keeping the signal at 96k and resolution 24bits or floating point avoids the (small) distortions from conversion and avoids small amounts of dithering noise from trickling in.   From my experience and the fairly aggressive processing in my project, there really is little reason (might be some, but uncommon) to need to go above 96k, but most of the time it wont hurt -- and MIGHT -- JUST MIGHT help.  With the new audio tape 'time base correction' scheme recently introduced, it might be useful to be able to pass higher frequencies, but hopefully the correction happens close to the tape machine -- so passing those very high frequncies shouldn't normally be needed (what is the highest typical bias frequency -- 600kHz? -- that would require a really high sample rate :-)).

Making the consumer deal with the higher frequencies just increases the burden.  Frankly, I don't like 44.1k, but 48k is certainly high enough for any consumer audio reason (unless the consumer is doing fancy dynamic gain control -- then sample rate conversion is a good thing.)

John

Re: How Much Do High Frequencies Matter In Practice?

Reply #15
Why don't you like 44.1? Is it the odd number that doesn't appeal to you or is it something else.

As the M&M study has demonstrated, they could find no evidence amongst using a wide range of listeners, from uni students, musicians, audiophiles and recording producers that could detect any difference between 44.1 and hi res under controlled conditions, using a wide variety of music, equipment, as much time as subjects needed and over hundred trials over a year.

Like they said, the burden of proof is on those that claim that 44.1 in any way degrades the sound.  So where is your proof?


Re: How Much Do High Frequencies Matter In Practice?

Reply #16
Why don't you like 44.1? Is it the odd number that doesn't appeal to you or is it something else.

As the M&M study has demonstrated, they could find no evidence amongst using a wide range of listeners, from uni students, musicians, audiophiles and recording producers that could detect any difference between 44.1 and hi res under controlled conditions, using a wide variety of music, equipment, as much time as subjects needed and over hundred trials over a year.

Like they said, the burden of proof is on those that claim that 44.1 in any way degrades the sound.  So where is your proof?

Note that I am WAY ahead of the matter of specious claims about 44.1k being bad in general -- I am NOT claiming that.

The reason why I don't like 44.1k isn't about distribution, it is about processing.  The high end of the audio is too close to the Nyquist frequency... I am not against 44.1k for distribution/delivery, it is that certain kinds of processing cauase distortion or gain control sidebands, whose spectrum can be tricky to avoid hitting the Nyquist frequency.   In pro applications, one can do lots of up and down conversion, but that degrades quality to a certain extent, so it is best to keep the signal at 48k or higher, where 96k or higher is best.  For my project, I have found about 64k is the sweet spot of easy processing and avoiding negative effects of aliasing in nonlinear situations.  96k just happens to be the next standard frequency (ingoring 88.2k, which also theoretically avoids the aliasing problems.)

For the consumer, my 'dislike' has little to do with the 'ringing' effects when chopping off the high frequencies -- because there is usually so litte music energy up there anyway.  The 'ringing' effects only occur when there is signal anyway.

If there is a matter of lots of energy in the 20kHz range, then it is better not to use 44.1k sample rate at all, unless one is willing to accept a smooth rolloff or some nonlinear processing to control the ringing effects.  Video is very sensitive to the ringing, and use of Thompson/Bessel filters to roll off before the brickwall is common.  Audio people don't like to do that kind of thing, but some special sauce is needed when starting with lots of energy in the 20k range.

The ringing effects at 20kHz are less of a problem for hearing, but rather can be troublesome for subsequent analog circuitry or even digital processing.

So, 44.1k is NOT a problem for consumer signal delivery, but for professional applications where there has been significant energy in the 20kHz range, the Gibbs effect CAN cause troulbe in subsequent equipment.  In fact, even some kinds of digital processing can be affected by it -- for example a compressor can be affected by the energy -- even though most peoples hearing is long gone by 20kHz.  The other thing is that analog electronics can have difficulty stomaching the energy in the higher frequencies, thereby causing in-band distortion effects (which cause the golden-ears who hear the difference to believe that having the wider response is better.  They are only hearing the difference in distortion.)

44.1k -- OKAY for consumer delivery, but not-so-good for professional production applications.  Nowadays, 48k is better because of fewer problems for any processing that might happen for the consumer, but 44.1k is certainly adequate for simple delivery.  Again, I am talking about avoiding up and down conversion for the processing.

John