Apodizing filter
Feb 15, 2024 at 8:58 PM Thread Starter Post #1 of 221

Digital Enigma

100+ Head-Fier
Joined
Aug 23, 2023
Posts
223
Likes
409
Location
UK
My Arcam SA20 came with the apodizing linear phase filter as default, what is the difference between this and the linear fast and linear slow, somewhere in the middle?

Online it says “Its main advantage is that it removes most of the ringing that has been introduced upstream in the recording process when the original material was recorded and mastered.”

But if ringing is only in the ultrasonic range what’s the point of it?

Hopefully somone can help me understand, thank you in advance.
 
Feb 15, 2024 at 10:17 PM Post #2 of 221
From what I can see, it looks like an anti-imaging filter. If I'm right, it is supposed to prevent aliasing that happens when digital systems process signals approaching the nyquist frequency.

That's mostly taken care of by the mixing & mastering engineer in the source file, so that filter is preventing any complications introduced by your system, say something like a hardware graphic EQ.
 
Feb 15, 2024 at 11:01 PM Post #3 of 221
I’ve been swapping between the apodizing and linear fast filters specifically which are built in to the ESS9038 and the only audible changes I can detect so far is less sharpness in vocal sibilance with the apodizing filter.
 
Feb 16, 2024 at 4:32 AM Post #4 of 221
It's a deliberate choice to prioritize one aspect of a signal over another. Some try to get what the math tells them to get(very hard filter to attenuate the higher frequencies not recorded, as much and as fast as possible). One extreme of that would be Chord Electronics for example.
Some go for the opposite and would rather roll off the upper frequencies slowly(usually starting well within the audible range) and typically leave more aliasing in the signal, for the sake of an impulse response looking like it doesn't "ring"(my explanation:wink:, they would tell you it's to improve the time domain, which IMO is debatable). The so called apodizing filter belongs to that group. So depending on implementation and your hearing, it might be fairly to very noticeable in the treble to high frequency when listening to 44.1kHz files.

Everybody makes designs that are somewhere in between because of cost and physical constraints, and all is well despite the very many small to big changes in design for the reconstruction filter and oversampling choices of a DAC. IMO a big part of why we have such a situation is because it doesn't have that much of an impact. And the higher the sampling rate, the less it's going to matter.
If you find yourself enjoying one setting more for whatever reason(could be the nice sounding name of the filter for all I know), I suggest using that and be happy. We're at a point in DAC designs where even the wrong ways to do things lead to a pretty good sound.

Edit: I struck the part that was wrong. My lame excuse is that I started with 2 extremes, and then somehow found myself stuck between those 2 options, even though my initial idea was to discuss the in between stuff and apodizing. I'd like to say it's rare for me to have brain farts like those, but then again, just yesterday I took off my sneakers to put on working shoes, then put my sneakers back on and went to work with them...
 
Last edited:
Feb 16, 2024 at 7:28 AM Post #5 of 221
The Arcam SA20 is said to have an ESS Sabre 90382KM DAC chip. I haven't found this one on ESS's site, the closest match was the 9038Q2M.
https://www.esstech.com/wp-content/uploads/2022/09/ES9038Q2M-Datasheet-v1.4.pdf

If you check the filters sections (from page 55 to 58), you can see that the apodizing filter is strikingly similar to the fast roll-off linear phase filter, both in the time and frequency domain. Linear phase filters will add some pre and post ringing to the signal if the signal's frequency reaches the transition band. As you can see, the apodizing filter is a linear phase filter.

The apodizing filter starts to roll off ever so slightly earlier than the fast linear phase filter but both start to roll of above 20kHz. The apodizing filter also reaches the -40dB (the output amplitude is 1% of the input amplitude) attenuation point comparatively faster than the fast linear phase filter so it is also a bit steeper. However, the stopband attenuation is a bit worse than the fast linear phase filter. I lined up their graphs and marked the -3dB and -40dB points.
filters.PNG

To me, both of these filters look like they are adequate for filtering the output of the DAC. They don't significantly change the signal in the pass band so there's no ringing there and they attenuate the signal in the stop band well enough.
 
Last edited:
Feb 16, 2024 at 10:32 AM Post #6 of 221
Very detailed replies thank you. I’m happy to leave it on the default setting (apodizing) because of the less sharp vocal sibilance alone. I can’t get on with chord dacs for too long whether it’s the super long pre ring being audible I don’t know.
 
Feb 16, 2024 at 11:39 AM Post #7 of 221
From a purely theoretical standpoint, if you wanted to accurately reconstruct exactly what was in the original signal, you'd need an infinite length (or maximum number of taps possible with the number of samples in the track anyway) sinc filter, of a 'halfband' design.
This essentially means it perfectly adheres to Nyquist theorem, fully attenuating everything above the Nyquist frequency, and not attenuating anything below that.
As @castleofargh mentioned, Chord is a good example of this, they have the steepest halfband sinc filter of any DAC manufacturer currently with the DAVE+MScaler:

1708101128424.png


The problem in practice is that we can't actually implement a 'perfect' or infinite length filter, given as an infinite length filter would require infinite computing power (closest thing we have is PGGB which does not work in real-time and isn't cheap). But other options like MScaler or HQPlayer get pretty darn close practically speaking, with MScaler being 1 million taps, and HQPlayer offering 2 Million taps with the Sinc-L filter.

But also, this approach assumes that the signal itself was 'perfect', and since a lot of aspects of production won't be perfect, such as the filters in the ADC itself used for recording often being relatively basic, the signal will often contain some unwanted issues such as ringing that can be removed by having a filter that doesn't reconstruct all the way up to the Nyquist Frequency. (Since ringing itself occurs at the nyquist freq, so attenuating before that means any ringing in the recording originally is removed)

If you had a 'perfect' signal, a non-apodizing filter would be best.
If you had a 'problematic' signal, with ringing caused by the ADCs or DSP used in production/recording, then an apodizing filter is arguably preferable.

Something key to mention though: Ringing is not 'always there'. A lot of people think ringing is a really bad thing, when it's not. Ringing is the result of the filter WORKING. It's an effect of the filter removing out-of-band content, and it only occurs in the presence of an illegal signal.
If you have a 'properly' recorded and produced track, with no clipping or other illegal content, then there won't actually be any ringing anyway.

A lot of manufacturers and consumers will point to impulse responses of 'short' filters and say "look no ringing that's great!", but the reason there's no ringing is because that filter is shorter and isn't actually filtering effectively. If you wanted zero ringing you could just use a NOS DAC. But in terms of what is actually objectively most accurate, a longer sinc filter is the answer generally speaking.
 
Last edited:
Feb 16, 2024 at 12:15 PM Post #8 of 221
A non-apodizing filter makes most sense when it is sinc based, following Whittaker-Shannon interpolation, which would necessitate that it is non-apodizing. Any deviation from this will no longer guarantee perfect reconstruction as per Whittaker-Shannon. The idea is quite simple, you need to use all the information that is present in the band limited input signal if you have to reconstruct the signal completely at a higher rate, but within the same band. I consider this as a special case.

Outside of that context, it may even be better to use an apodizing filter to remove some of the aliasing introduced by ADC as there is no new advantage to be gained. Apodizing filter does not automatically mean, short with slow roll-off. Apodizing filters (in the sense of removing aliasing artifacts right at the edge of audible spectra) can be linear, or long and sharp too. So I would go with whatever sound best on the DAC for the type of music I listen to.
 
Feb 16, 2024 at 12:24 PM Post #9 of 221
Yep!

For example HQPlayer's 'Sinc M' filter is 1 million taps (upto 2 million if you use Sinc-Mx and have a PCM1.536Mhz capable DAC) but attenuates at 21khz. So you still get very good bandwidth and no early rolloff or remaining out of band content, plus hundreds of dB of stopband attenuation, but it is also apodizing.

1708104287231.png

1708104291864.png
 
Feb 16, 2024 at 1:53 PM Post #10 of 221
From a purely theoretical standpoint, if you wanted to accurately reconstruct exactly what was in the original signal, you'd need an infinite length (or maximum number of taps possible with the number of samples in the track anyway) sinc filter, of a 'halfband' design.
This essentially means it perfectly adheres to Nyquist theorem, fully attenuating everything above the Nyquist frequency, and not attenuating anything below that.
As @castleofargh mentioned, Chord is a good example of this, they have the steepest halfband sinc filter of any DAC manufacturer currently with the DAVE+MScaler:



The problem in practice is that we can't actually implement a 'perfect' or infinite length filter, given as an infinite length filter would require infinite computing power (closest thing we have is PGGB which does not work in real-time and isn't cheap). But other options like MScaler or HQPlayer get pretty darn close practically speaking, with MScaler being 1 million taps, and HQPlayer offering 2 Million taps with the Sinc-L filter.

But also, this approach assumes that the signal itself was 'perfect', and since a lot of aspects of production won't be perfect, such as the filters in the ADC itself used for recording often being relatively basic, the signal will often contain some unwanted issues such as ringing that can be removed by having a filter that doesn't reconstruct all the way up to the Nyquist Frequency.

If you had a 'perfect' signal, a non-apodizing filter would be best.
If you had a 'problematic' signal, with ringing caused by the ADCs or DSP used in production/recording, then an apodizing filter is arguably preferable.

Something key to mention though: Ringing is not 'always there'. A lot of people think ringing is a really bad thing, when it's not. Ringing is the result of the filter WORKING. It's an effect of the filter removing out-of-band content, and it only occurs in the presence of an illegal signal.
If you have a 'properly' recorded and produced track, with no clipping or other illegal content, then there won't actually be any ringing anyway.

A lot of manufacturers and consumers will point to impulse responses of 'short' filters and say "look no ringing that's great!", but the reason there's no ringing is because that filter is shorter and isn't actually filtering effectively. If you wanted zero ringing you could just use a NOS DAC. But in terms of what is actually objectively most accurate, a longer sinc filter is the answer generally speaking.

I agree that the super sharp linear phase is the correct filter and the most accurate but 90% of my CD collection is not well recorded it’s compressed rock and metal, in that case I think apodizing or even minimum phase is a better choice for me. I don’t have super high res audiophile grade recordings and masterings, that style of music isn’t my cup of tea anyway. Maybe that explains why I prefer the apple dongle over the mojo 1 for example.
 
Feb 16, 2024 at 2:05 PM Post #11 of 221
I agree that the super sharp linear phase is the correct filter and the most accurate but 90% of my CD collection is not well recorded it’s compressed rock and metal, in that case I think apodizing or even minimum phase is a better choice for me. I don’t have super high res audiophile grade recordings and masterings, that style of music isn’t my cup of tea anyway. Maybe that explains why I prefer the apple dongle over the mojo 1 for example.
Yeah in many cases apodizing could be a better option, especially with music that wasn't recorded/produced so well.

Minimum phase is a bit of a different thing though. That's more along the lines of "You should always use linear phase unless you need lower latency for production" imo.
 
Feb 16, 2024 at 5:31 PM Post #12 of 221
Something key to mention though: Ringing is not 'always there'. A lot of people think ringing is a really bad thing, when it's not. Ringing is the result of the filter WORKING. It's an effect of the filter removing out-of-band content, and it only occurs in the presence of an illegal signal.
If you have a 'properly' recorded and produced track, with no clipping or other illegal content, then there won't actually be any ringing anyway.

Could you please go into a bit more detail on what you mean by "an illegal signal" and "other illegal content". Thanks.
 
Last edited:
Feb 16, 2024 at 6:59 PM Post #13 of 221
Could you please go into a bit more detail on what you mean by "an illegal signal" and "other illegal content". Thanks.

Illegal signal is impulse response that doesn't conform to Nyquist bandwidth limits (>22 KHz). This exposes the filter used by DACs (based on the ringing and phase in the impulse response illegal signal)
 
Feb 16, 2024 at 7:37 PM Post #14 of 221
Illegal signal is impulse response that doesn't conform to Nyquist bandwidth limits (>22 KHz). This exposes the filter used by DACs (based on the ringing and phase in the impulse response illegal signal)

From what I understand, the Nyquist-Shannon sampling theorem simply states that the sample rate must be at least twice the bandwidth of the desired signal to avoid aliasing. So the 22.05kHz bandwidth limit is merely a function of having the sample rate used for 'CD quality' digital audio set at 44.1kHz. Now if we were talking about 'Hi-Res' digital audio with a 192kHz sample rate, then its Nyquist-defined bandwidth would be from 0Hz-96kHz.
 
Last edited:

Users who are viewing this thread

Back
Top