Bernhard said:
Bernard,
Simple...really...well, - let's assume that you might be using a ladder type D/A like the Philips used to make or older, but still great Analog Devices AD1862 etc.
With 44.1 kHz sampling rate the image frequency will centre at 44.1 kHz. Considering that the usable data extends to about 18 kHz you will have approximately 8 kHz of 'window' where your anti-aliasing filter will have to work to suppress the image frequency.
Even accounting for the ‘zero-order-hold’ of the D/A, which is a low-pass filter in the shape of a sin(x)/x or sinc function, contributing a few dB of attenuation, still the unti-aliasing low-pass filter will have to be of a huge order, completely impractical, in fact not doable.
You see, 16-bits of signal resolution translate into about 96 dB of S/N. This is what you need in order to recover the 16-bits. Assuming that the sin(x)/x gives you about 10 dB at 26 kHz (44-18) you still need to generate another shall we say 80 dB or so in order to hit the target S/N.
It is not possible to do an 80 dB low-pass filter in a space of 8-10 kHz. So, the over sampling is a must here. It will shift the image frequencies to a higher point. Alternatively you will have you live with a reality of much lower signal resolution. My guess is that even 15 bits is not achievable without an over sampling, - more like 12-13 bits is doable. The rest will be covered by noise.
Regards,
Vadim
Which algorithm should you use when upsampling and should you use an anti-alias filter when upsampling? I told someone on here yesterday not to use one when upsampling now after reading about NOS DACs and upsampling DACs my head freakin hurts.
So anyway lets say for this example I am using soundforge's batch converter. It gives you 3 algorithms - linear, lagrange, sin(x)/x. Which would be the best - if there is one - for upsampling? And what would the advantage be of using an antialias on an upsample?
So anyway lets say for this example I am using soundforge's batch converter. It gives you 3 algorithms - linear, lagrange, sin(x)/x. Which would be the best - if there is one - for upsampling? And what would the advantage be of using an antialias on an upsample?
Actually, any frequency component above Fs/2 is indistinguishable from a lower-frequency component, called an alias, associated with one of the copies. So anything over 22.05 kHz...
That's why a NOS DAC cannot function proper with most of the DIY filters that are shown on this site (more like no filters at all). Distortion level is extremely high with no abrupt filters.
Why those aren't so good either? Abrupt filters have the habit to make weird phase shifts in the signal. Not good for quality music reproduction.
Going with a simpler filter that has a decent attenuation of image and slow roll-off (low distortion) will be cutting the high-end of the audio band.
So the only way that you can have the cake (maximum bandwidth) and eat it (low distortion) is to use an over-sampling digital signal to relax the requirements for the analog low-pass filter.
LE: Even a filter like the one above that looks good (and hard to make), has only a -30dB attenuation at 30kHz - well into the image field. Thats' not even close of the -96dB that are easily obtainable with a 8x OS and simpler filters.
That's why a NOS DAC cannot function proper with most of the DIY filters that are shown on this site (more like no filters at all). Distortion level is extremely high with no abrupt filters.
Why those aren't so good either? Abrupt filters have the habit to make weird phase shifts in the signal. Not good for quality music reproduction.
Going with a simpler filter that has a decent attenuation of image and slow roll-off (low distortion) will be cutting the high-end of the audio band.
So the only way that you can have the cake (maximum bandwidth) and eat it (low distortion) is to use an over-sampling digital signal to relax the requirements for the analog low-pass filter.
LE: Even a filter like the one above that looks good (and hard to make), has only a -30dB attenuation at 30kHz - well into the image field. Thats' not even close of the -96dB that are easily obtainable with a 8x OS and simpler filters.
It's not about the quality of the components, it's the laws of physics that dictates wild phase shifts in the case of abrupt filters. At the best, soundstage gets completely ruined.
I don't deny that those aliasing distortions some might find enjoyable and give fancy names and qualifications for them. I did listen to a "well made" NOS and it sounded plain bad to me. Am matter of gustinbus non est disputandum, I guess
I don't deny that those aliasing distortions some might find enjoyable and give fancy names and qualifications for them. I did listen to a "well made" NOS and it sounded plain bad to me. Am matter of gustinbus non est disputandum, I guess
What good does removing the FS only? The images will be there and will be "mirrored" onto the real signal (high end will be translated into the low end and low into the high).
Cost for OS? Well... IMHO I think it is minimal compared with what happens with the NOS case. Actually that's the opinion of most of the audio industry.
Cost for OS? Well... IMHO I think it is minimal compared with what happens with the NOS case. Actually that's the opinion of most of the audio industry.
SoNic_real_one said:
Yes, it is, a signal that runs through such a filter, is totally lost.
SoNic_real_one said:
Did you hear that with your own ears or is it a worry ?
To my ears, the brickwall filter is no problem.
I have listened to everything, 1bit & multibit with all kinds of os rates & filters.
This is far the best IMHO, but sound is subjective.
Well... maybe you are right here. I didn't like the sound of the "brickwall" - maybe is not the phase shift that bothers me most but the incomplete rejection of the image. The best analog filters can do little to nothing for that. I don't know exactly, but also the numbers are not in favor of NOS. And it's not only me, all the audio industry people think the same since they make only OS DAC's now. They must be hearing something too.
Anyway, I know it is a "hot issue" I will leave it as being a matter of audio taste. Actually I choose to buy only SACD or DVD-A now to avoid completely those issues of CD given by the low Fs. SACD and DVD-A sound terrific compared with the CD versions IMHO.
Appreciate the good chat with you thou
PS: 4x OS means 2 extra bits in interpolation. 8x means 3 bits. A 24 bit DAC can loose those bits safely since only 18-20 will be translated in usable signal because of the analog limitations (2V output is one of them).
Anyway, I know it is a "hot issue" I will leave it as being a matter of audio taste. Actually I choose to buy only SACD or DVD-A now to avoid completely those issues of CD given by the low Fs. SACD and DVD-A sound terrific compared with the CD versions IMHO.
Appreciate the good chat with you thou
PS: 4x OS means 2 extra bits in interpolation. 8x means 3 bits. A 24 bit DAC can loose those bits safely since only 18-20 will be translated in usable signal because of the analog limitations (2V output is one of them).
Bernhard,
Judging from the overshoot of the filter response you posted it is an Elliptic or possibly Inverse Chebushev, - well, whatever the case may be, - the phase non-linearity makes this filter completely inappropriate for the application. But then again, - this is what you have to do in order to pay for the absents of over sampling. There is no free lunch here. As it was noted in earlier posts, that even a 7th order filter is not sufficient here.
Another thing is that you are under impression that the sampling frequency can be somehow removed.
This is not possible. What is possible is to diminish the effects of the sampling to a degree that is commeasurable with the precision of the data, - 16 bits. In our case it is a 16-bit S/N that serves as our target for the noise suppression. Doing better then that while admirable does nothing to your signal integrity.
In any event, there is only one solution possible here and that is setting an over sampling requirements to at least 8x or better 16x. With 8x over sampling and Bessel or Thomson type low-pass filter you can get away with approximately 6 order. No phase nonlinearities here and you can use the best op-amps available, like perhaps AD797, which is the only one I know that does sport 16-bit capability.
If you do the power supply regulation right, you may just sneak under measurable 16-bit S/N. I know it is possible as I have done it in the past. Shooting for better then 16-bits is extremely difficult and even expensive to measure properly. Better yet you can do a passive filter with shielded inductor and buffer it with a discrete extremely low-noise buffer. Not a simple task, but definitely doable.
I have a CD Player here dating back to 1990, - the Denon DCD-2650. It features 16x over sampling. I heavily modified it over the yeas and it does measure to about 100 dB S/N. It uses 4 AD1862 D/A in a balanced configuration. Great player, in fact a state of the art by modern standards, once you fix the power supply and it’s I/V converter along with analog filtering. I used 4th order Bessel to achieve the S/N target.
Vadim
Judging from the overshoot of the filter response you posted it is an Elliptic or possibly Inverse Chebushev, - well, whatever the case may be, - the phase non-linearity makes this filter completely inappropriate for the application. But then again, - this is what you have to do in order to pay for the absents of over sampling. There is no free lunch here. As it was noted in earlier posts, that even a 7th order filter is not sufficient here.
Another thing is that you are under impression that the sampling frequency can be somehow removed.
Bernhard said:
This is not possible. What is possible is to diminish the effects of the sampling to a degree that is commeasurable with the precision of the data, - 16 bits. In our case it is a 16-bit S/N that serves as our target for the noise suppression. Doing better then that while admirable does nothing to your signal integrity.
In any event, there is only one solution possible here and that is setting an over sampling requirements to at least 8x or better 16x. With 8x over sampling and Bessel or Thomson type low-pass filter you can get away with approximately 6 order. No phase nonlinearities here and you can use the best op-amps available, like perhaps AD797, which is the only one I know that does sport 16-bit capability.
If you do the power supply regulation right, you may just sneak under measurable 16-bit S/N. I know it is possible as I have done it in the past. Shooting for better then 16-bits is extremely difficult and even expensive to measure properly. Better yet you can do a passive filter with shielded inductor and buffer it with a discrete extremely low-noise buffer. Not a simple task, but definitely doable.
I have a CD Player here dating back to 1990, - the Denon DCD-2650. It features 16x over sampling. I heavily modified it over the yeas and it does measure to about 100 dB S/N. It uses 4 AD1862 D/A in a balanced configuration. Great player, in fact a state of the art by modern standards, once you fix the power supply and it’s I/V converter along with analog filtering. I used 4th order Bessel to achieve the S/N target.
Vadim
Key said:
I know how you feel. My heads hurts too when I think about these issues. It seems that there is a bit of a confusion in your post.
The algorithm you refer to is actually a reference to an interpolation, - you can’t NOT use it. It is a part of the over sampling process.
Key said:
It really does not matter which ‘algorithm’ you use, as I bet you can never tell them apart. The differences are more of a theoretical nature here.
The advantage of using an anti-aliasing filter is a noise suppression. If you do not do that then I would question the entire D/A process, - may as well forget about the digital audio and go back to records.
I think there is no confusion here. Maybe some "urban legends" raised at the level of religion 
NOS DAC's at 44.1 kHz Fs are just unmanageable on analog output.
16 bit OS DAC's originating from 44.1 KHz are "loosing" some of that resolution to the interpolation. As I said above, it means 2 bit for 4x, 3 bit for 8x, 4 bit for 16x.
24 bit OS DAC are "loosing" the same 3-4 bits but nobody will notice that since looks like the 20 bit dynamic range it is the analog limit those days.
NOS DAC's at 44.1 kHz Fs are just unmanageable on analog output.
16 bit OS DAC's originating from 44.1 KHz are "loosing" some of that resolution to the interpolation. As I said above, it means 2 bit for 4x, 3 bit for 8x, 4 bit for 16x.
24 bit OS DAC are "loosing" the same 3-4 bits but nobody will notice that since looks like the 20 bit dynamic range it is the analog limit those days.
Well I understand the idea when downsampling a signal because the side effects become glaringly obvious.
I think I will take you on and give it a try, upsampling to 192k 24-bit with all three and some material and see if I can hear a difference. Yeah yeah I know my gear probably sucks and is bottlenecked or whatever but I am trying it haha
er I think there is a bit of confusion in your post. I was asking if or if not I should use an anti alias filter, not an algorithm. I was also asking which algorithm was generally accepted to be the "best" if any.
I think I will take you on and give it a try, upsampling to 192k 24-bit with all three and some material and see if I can hear a difference. Yeah yeah I know my gear probably sucks and is bottlenecked or whatever but I am trying it haha
er I think there is a bit of confusion in your post. I was asking if or if not I should use an anti alias filter, not an algorithm. I was also asking which algorithm was generally accepted to be the "best" if any.
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.