Apodizing filter

[bigshot]

Ethan Winer didn't go beyond -40dB in his test, because he said it was pointless to worry about noise much beyond that. You can force yourself to hear further with gain riding on fade outs and using targeted test tones if you really want to, but I don't know what you're proving if you do that. And by the time you get down to -50dB in a commercial music recording you're starting to get into the sound of air conditioning in the recording studio and traffic driving by outside.

 

[Digital Enigma]

Ethan Winer didn't go beyond -40dB in his test, because he said it was pointless to worry about noise much beyond that. You can force yourself to hear further with gain riding on fade outs and using targeted test tones if you really want to, but I don't know what you're proving if you do that. And by the time you get down to -50dB in a commercial music recording you're starting to get into the sound of air conditioning in the recording studio and traffic driving by outside.

So what are you saying a DAC with a sinad of 50db is all you need or am I misinterpreting it?

 

[KinGensai]

So what are you saying a DAC with a sinad of 50db is all you need or am I misinterpreting it?

https://www.hypethesonics.com/dapti-database/

https://www.klippel.de/listeningtest/

As you can see in the database, even mediocre DAPs can produce sound with deviations from the source file below 50dB. Try that listening test @GoldenSound brought up, -50dB is a huge sound differential.

 

[bigshot]

If a DAC isn’t audibly transparent, something is set up wrong or the DAC is defective. Digital audio at 16/44.1 is audibly transparent. A DAC should be able to reproduce that. Most DACs are beyond transparent by a lot.

 

[GoldenSound]

https://www.hypethesonics.com/dapti-database/

https://www.klippel.de/listeningtest/

As you can see in the database, even mediocre DAPs can produce sound with deviations from the source file below 50dB. Try that listening test @GoldenSound brought up, -50dB is a huge sound differential.

I wouldn't use the hypethesonics site as an indicator in this regard. Doing null testing with multiple DACs without having the entire setup slaved to a single clock source is going to lead to differences that could look far larger in these tests than they are in reality. It's a methodology that is very rarely used for this and various other reasons.

 

[KinGensai]

I wouldn't use the hypethesonics site as an indicator in this regard. Doing null testing with multiple DACs without having the entire setup slaved to a single clock source is going to lead to differences that could look far larger in these tests than they are in reality. It's a methodology that is very rarely used for this and various other reasons.

From what I'm seeing here, digital headroom to prevent interpolation clipping errors seems like the most relevant design feature for practical use because the other metrics look to be typically 2x below a practical audible threshold of -50dB outside of occasional artifacts. Are those spikes typically what you are looking for when judging by listening?

 

[gregorio]

Let’s put this blind testing to the side for a minute because it’s straying off the path of my original question, what does an apodizing fast linear phase actually do.

There’s no answer to that question, or rather there’s no one answer. What a particular apodizing filter actually does is only known by those who programmed it, because technically “apodizing” just means modifying the shape of a mathematical function, so it could mean almost anything. In practice though it typically means the modification is designed to reduce ringing, although there’s no free lunch here, reducing the ringing necessitates degrading some other aspect of the filter’s performance, for example lengthening the transition band (the roll-off starting earlier in the freq spectrum) or by reducing the amount of attenuation above the Nyquist point. This isn’t always the case though. Usually we can deduce roughly what the filter is doing by pumping in some extreme test signals and looking at the filter’s response, say a Dirac Impulse, white noise or a sine sweep at full scale (or near full scale). However, this doesn’t characterise what you’ll actually get with music recordings because music recordings do not contain Dirac Impulses or full scale white noise or sine waves in the ultrasonic range.

So far I’ve learned that it reduces/removes problems caused upstream during the recording and mixing process by cutting off early before the nyquist frequency but the filter it’s self is not free of producing its own ripple/ringing in the process. Correct me if I’m wrong.

Where did you learn that? I’m not sure if you’re referring to the assertions made by GoldenSound about “proper” mastering/production but please note my responses. When I asked him for any examples of ADCs with such poor filters or the relevancy of ultrasonic content (when clipping during mixing/mastering), he didn’t/couldn’t answer and just changed the subject. And, even if there were “problems”, then an apodizing filter would not reduce/remove them. The only exception would be if we had some problem in the ultrasonic range, say some ringing at 22kHz or some aliasing at 20kHz and the apodizing filter had a stop band starting lower than that. However, you don’t need an apodizing filter to do that, any minimum or linear phase filter with a lower stop band would achieve that. Also, ringing is rare and very low magnitude, not to mention that it’s beyond the range of human hearing anyway, and we don’t get any significant aliasing in the audible range with modern ADCs, even modestly priced prosumer ones.

G

 

[KinGensai]

https://community.roonlabs.com/t/ap...mparator-results-no-audible-difference/257336

https://archimago.blogspot.com/2015/07/the-linear-vs-minimum-phase-upsampling.html?m=1

Some related reading I found.

My understanding as articulated previously was not complete. With regular linear phase filters, pre and post ringing are present from spectral leakage. When a filter is apodized, the practical effect is that the pre-ringing is eliminated, but the post-ringing is significantly lengthened.

I now realize why I found Meridian's filter settings worded strangely.

This is from my V60, but the V50 has the same filters and diagrams. The descriptions were about the filter itself, not the resulting impulse response.

So it seems that Meridian's "short" filter is an apodized filter. Like I said earlier, I preferred the "slow" filter because it sounded the clearest to me, while the apodized filter made everything sound dirtier for my library and transducers. Unlike the blind test archimago conducted, I listen to electronic music as well as acoustic music like jazz and metal, and the effect of these filters was more apparent with House and EDM because clean sine waves actually happen often in that kind of music compared to acoustic music.

 

[Digital Enigma]

https://community.roonlabs.com/t/ap...mparator-results-no-audible-difference/257336

https://archimago.blogspot.com/2015/07/the-linear-vs-minimum-phase-upsampling.html?m=1

Some related reading I found.

My understanding as articulated previously was not complete. With regular linear phase filters, pre and post ringing are present from spectral leakage. When a filter is apodized, the practical effect is that the pre-ringing is eliminated, but the post-ringing is significantly lengthened.

I now realize why I found Meridian's filter settings worded strangely.
This is from my V60, but the V50 has the same filters and diagrams. The descriptions were about the filter itself, not the resulting impulse response.

So it seems that Meridian's "short" filter is an apodized filter. Like I said earlier, I preferred the "slow" filter because it sounded the clearest to me, while the apodized filter made everything sound dirtier for my library and transducers. Unlike the blind test archimago conducted, I listen to electronic music as well as acoustic music like jazz and metal, and the effect of these filters was more apparent with House and EDM because clean sine waves actually happen often in that kind of music compared to acoustic music.

Wouldn’t the aliasing let in by a slow filter sound horrible though?

 

[gregorio]

Measurements available here: https://goldensound.audio/2022/05/26/chord-mojo-2-measurements/
Lower distortion/noise than apple dongle. Main difference would be the significantly better reconstruction filter (which is also linear phase) and significantly lower jitter

Right, so the “Main” differences at a reasonable listening level is 1. A significantly better reconstruction filter (actually anti-imaging filter), which is only different in the ultrasonic region and therefore audibly the same and 2. significantly lower jitter, which can’t even exist as sound, let alone be audible. So in terms of audibility your “main difference” is zero difference and “significantly better/lower” effectively means identical, so 1. No! and 2. Hell no! lol

The evidence in this phenomenon points overwhelmingly to people perceiving higher price with better quality

Sure but in terms of creating a bias when listening, price is only one determinant. For example, we could have unknown price or identical price, just a difference solely in appearance (EG. The identical DAC in two different cases) and yet still have a biased result. In addition, rather ironically higher price commonly actually gets you lower quality in the case of audiophile DACs and amps (if we define “quality” by fidelity).

It doesn't seem to me to matter upstream of the PCM to analog conversion if the track was mastered properly because the engineer uses higher sampling rates and a low pass filter of some kind to control aliasing.
IIRC of course, I'm not an audio engineer.

The engineer doesn’t necessarily use a higher sample rate. Generally we will only use a higher sample rate if delivery specifications require it (Eg. If the client wishes to market a higher sample rate version) so it’s sometimes the case with music but extremely rarely the case with sound for TV/Film (which is virtually always at 48kFs/S). However, there are some processors which could or need to use freqs higher than 22.05kHz, some modelled compressors and limiters for example, to emulate the resultant IMD. In such cases the processor (plugin) will upsample locally, typically outside the control of the engineer, although some provide the option to choose the oversample factor.

Answer seems clear to me: use the filter that sounds best to you. They all address aliasing, and IME the effect of changing filters is almost imperceptibly subtle if your transducers don't have an IR lower than 2ms.

None of them address aliasing. There is no aliasing in the DA conversion process, only in the AD conversion process. The filters actually address “images”, which is content above Nyquist, hence why they’re called anti-image filters rather than anti-alias filters, even though functionally anti-alias and anti-image filters might be identical. Just to be clear, aliases are content above the Nyquist Point that is folded-down (“mirrored”) into the spectrum below Nyquist. Aliases are therefore a much bigger theoretical problem than “images”, as they‘re very likely in the audible band rather than the ultrasonic band, although ultrasonic images are likely to cause IMD (in the amp or more likely, the transducers) in the audible band if not sufficiently attenuated.

I don't think there is sufficient study to conclude an audible transparency threshold with regards to reconstruction filters yet.

Of course science can’t prove a negative so there will never be absolute proof. Additionally, it is of course entirely possible to design an anti-imaging filter that’s effectively faulty (EG. A filter which is deliberately not audibly transparent) and some manufacturers actually do this, so there will be some exceptions. Assuming a typical, non-faulty filter though, there is plenty of reliable evidence to conclude audible transparency, especially if one uses inference from studies which indirectly address the issue. For example, if it’s not possible to tell the difference between 16/44.1 and say 24/192 or SACD, then by inference it’s not possible to tell the difference between the different filters being employed.

Looks like I get down to -39dB, don't know if android resampling has an effect or not.

I don’t know but bare in mind that a scientific/acceptable DBT or ABX will generally require a period of training to identify the specific artefact being tested. Equipment can certainly make a difference but training can make a very significant difference too, additionally, manipulating the test conditions can make a very big difference (for example, raising the level significantly). At one time I could hear certain distortion artefacts just below -70dB and I’ve seen it done to about -80dB but that was in a world class studio and at listening levels that were uncomfortably loud. I’d be surprised if you couldn’t do significantly better than -39dB with a bit of practice (unless there’s something preventing it with your playback setup).

So what are you saying a DAC with a sinad of 50db is all you need or am I misinterpreting it?

You’re misinterpreting it. The “-50dB” is a rather arbitrary figure. A difference of only -6dB can be completely inaudible in some cases, while in others, under rare conditions -70dB would be a more appropriate figure. It all depends on exactly what it is and how it relates to the content you’re testing, where in the spectrum it is, your listening conditions and also your listening skills. Unfortunately, SINAD does not tell us exactly what it is or where in the spectrum it is. So it’s possible, although unlikely, that a SINAD at -40dB would be inaudible while a SINAD of -50dB could be audible (for the same person in the same listening conditions). -50dB is a reasonable “rule of thumb” general threshold limit but may not always be the case.

Doing null testing with multiple DACs without having the entire setup slaved to a single clock source is going to lead to differences that could look far larger in these tests than they are in reality.

True to an extent, especially if using longer recording times but doing the test with the DACs slaved to an external clock will commonly degrade their performance and therefore lead to even larger differences. There are certainly issues with null testing DACs but it’s still a useful indicator.

When a filter is apodized, the practical effect is that the pre-ringing is eliminated, but the post-ringing is significantly lengthened.

That’s the effect of a minimum phase filter, it is often the effect of an apodizing filter too but not always/necessarily.

G

 

[GoldenSound]

significantly lower jitter, which can’t even exist as sound, let alone be audible

I'm not sure what you mean here. I'm measuring jitter from the analog output of the device. I'm not referring to input jitter and just assuming it's directly carrying across to the output.

Only time I measure jitter at the interface level is when testing DDCs

Of course science can’t prove a negative so there will never be absolute proof

I'm not asking for it to do so. Just simply saying that in the absence of any study concluding one way or the other, it wouldn't be scientific to make the claim that a reconstruction filter is definitely inaudible. It's an assumption.

Assuming a typical, non-faulty filter though

"non-faulty" is a grey area. If you mean "doesn't accurately conform to Nyquist theorem" then technically that applies to basically every filter. The vast majority of DAC filters either roll off early, attenuating treble under 20khz, or don't attenuate fully by Nyquist in the first place.

At what point a filter is 'sufficient' is entirely up for debate.

 

[KinGensai]

None of them address aliasing. There is no aliasing in the DA conversion process, only in the AD conversion process. The filters actually address “images”, which is content above Nyquist, hence why they’re called anti-image filters rather than anti-alias filters, even though functionally anti-alias and anti-image filters might be identical. Just to be clear, aliases are content above the Nyquist Point that is folded-down (“mirrored”) into the spectrum below Nyquist. Aliases are therefore a much bigger theoretical problem than “images”, as they‘re very likely in the audible band rather than the ultrasonic band, although ultrasonic images are likely to cause IMD (in the amp or more likely, the transducers) in the audible band if not sufficiently attenuated.

Ok, let me clear up something then. When I tried adding a high shelf from 15kHz up to pink noise with EQ, I noticed noise that was clearly below 15kHz being added to the signal. After listening to an example of aliasing happen in production and how engineers control aliasing by using higher sampling rates to push the nyquist frequency higher, I compared that to what I was getting on my DAP and think that's what's happening here.

Since I got aliasing from that high shelf, is the anti-image filter upstream of the software PEQ, thus unable to prevent that aliasing? The purpose then is to prevent images that are present in a source file from causing aliasing?

I don’t know but bare in mind that a scientific/acceptable DBT or ABX will generally require a period of training to identify the specific artefact being tested. Equipment can certainly make a difference but training can make a very significant difference too, additionally, manipulating the test conditions can make a very big difference (for example, raising the level significantly). At one time I could hear certain distortion artefacts just below -70dB and I’ve seen it done to about -80dB but that was in a world class studio and at listening levels that were uncomfortably loud. I’d be surprised if you couldn’t do significantly better than -39dB with a bit of practice (unless there’s something preventing it with your playback setup).

G

I did this test at my nominal listening level, so I could have pushed the gain up I suppose, but that would have been pretty uncomfortable.

This test was for identifying what sounded to me like artificially induced distortion in a music test track, so I listened for attack & decay of bass hits in particular because distortion in the bass harmonics are easier to identify consciously for me. After it got to around -33dB, discerning a difference became pretty fatiguing over iterations. Test tones probably would have been much easier because I used pure sine waves to learn what THD sounds like, but I wanted to know what my limits are in my everyday listening conditions.

 

[KinGensai]

Wouldn’t the aliasing let in by a slow filter sound horrible though?

I did not notice any aliasing occur during regular use. It did happen when I EQed treble up significantly as an experiment, but I don't know if any DAC filter could stop that.

 

[bigshot]

At what point a filter is 'sufficient' is entirely up for debate.

The filters used on most DACs are sufficient, because most DACs (excepting NOS) are audibly transparent. If you can’t hear it, it doesn’t matter.

 

[gregorio]

I'm not sure what you mean here. I'm measuring jitter from the analog output of the device.

Exactly, but of course that’s the analogue output of the device, not sound. If you have jitter artefacts at say -115dB (roughly that of the Apple Dongle) and your monitoring peak level is a reasonable, non-ear damaging 85dBSPL, what is the SPL level of the jitter artefacts? What about for the Mojo you measured?

Just simply saying that in the absence of any study concluding one way or the other, it wouldn't be scientific to make the claim that a reconstruction filter is definitely inaudible. It's an assumption.

It’s not an assumption, it’s an inference and how are inferences not scientific?

If you mean "doesn't accurately conform to Nyquist theorem" then technically that applies to basically every filter.

No it doesn’t, technically the vast majority of filters accurately conform to Nyquist/Shannon! They don’t conform “perfectly” to the Nyquist/Shannon Theorem but why do you think they need to? What do you think human hearing is capable of and what do you think the signals are that we’re recording and reconstructing?

The vast majority of DAC filters either roll off early, attenuating treble under 20khz, or don't attenuate fully by Nyquist in the first place.

Adult humans cannot hear a filter that starts to roll-off at say 19kHz or even 17kHz or 18kHz and is down by 2 or 3dB at 20kHz. In fact, after 6 years of testing about 1,800 students (mostly 18-21 years of age) we never found a single one who could identify a roughly 80dB difference at 20kHz, let alone just 2-3dB! Additionally, there is NO requirement to “attenuate” by Nyquist, there is only the requirement to band limit by Nyquist. So if the signal you’re recording has no content above 22kHz then absolutely no attenuation whatsoever will perfectly comply with the Nyquist/Shannon Theorem! Of course, you do need to attenuate somewhat in practice because there is almost always at least some content there but typically it’s at least two or three orders of magnitude below peak. You would still need an anti-imaging filter that attenuates in the DAC of course, due to the error signal but how much error signal do you think we have? In theory we don’t need an anti-image filter at all, because we can’t hear anything above the Nyquist point (22.05kHz) but in practice we’re likely to get IMD out of our speakers or amp. So if we had a full scale signal at 20,100Hz, our anti-alias filter attenuated by only 40dB by 24,100Hz (where our error signal “image” will be) and our speakers had 1% distortion at that freq, we’d get an IMD product at 4kHz, what would it’s level be? Without actually calculating it, I’m guessing somewhere around -80dB to -90dB. This example is for full scale though, how much content are we actually likely have at 20.1kHz to start with and what would be the distortion level then? I’ll give you a clue, with a violin at 1m distance, about 4% of it’s output is above 20kHz.

At what point a filter is 'sufficient' is entirely up for debate.

Only if one has a financial or other vested interest, for example, those selling, reviewing or justifying the purchase of products using anti-alias filter properties as a marketing point! For the rest of us it’s a debate that is either irrelevant or already done and dusted decades ago, with the only exception being the silly/faulty filter designs that have cropped up in the last few years for audiophile marketing purposes.

G