It seems ygg-it wants to know if the oversampling and reconstruction filtering can somehow be separate processes even if the industry has chosen (?) to always combine them?
Oversampling is math - one number in, several out. Reconstruction is producing a voltage level corresponding to a number.
Is it not practical possible to split them?
//
Oversampling is math - one number in, several out. Reconstruction is producing a voltage level corresponding to a number.
Is it not practical possible to split them?
//
Not quite. First, it's the D/A unit which produces voltage or current levels which correspond to various numbers. Signal reconstruction, however, is better thought of as an frequency-domain low-pass filtering process where once all of the ultrasonic image bands are optimally removed the original signal automatically results. It's the filtering away of the image bands which reveals the original analog signal. The original signal only seems to need reconstruction because it is disguised by the presence of the image bands.
The idea of splitting playback up/oversampling from utilizing a digital filter doesn't make sense once you understand that interpolative oversampling is inherently an digital low-filter based process.
Upsampling can be done for a file of digital PCM words and stored in a new file, for later processing, so perfectly separable for reconstruction.
Reconstruction is a sinx/x filter doing it's integration thing to recreate the signal described by PCM samples. Sinc function - Wikipedia
"In a mixed-signal system (analog and digital), a reconstruction filter is used to construct a smooth analog signal from a digital input, as in the case of a digital to analog converter (DAC) or other sampled data output device." Reconstruction filter - Wikipedia
//
Reconstruction is a sinx/x filter doing it's integration thing to recreate the signal described by PCM samples. Sinc function - Wikipedia
"In a mixed-signal system (analog and digital), a reconstruction filter is used to construct a smooth analog signal from a digital input, as in the case of a digital to analog converter (DAC) or other sampled data output device." Reconstruction filter - Wikipedia
//
Last edited:
Upsampling is a form of reconstruction, so the two cannot be completely separated. An properly upsampled signal is necessarily, at least partly, reconstructed. Reconstruction is the low-pass filtering of the signal’s image bands. Upsampling also requires the low-pass filtering of the image bands.
No worries.
Part of the confusion, I think, is that we tend to conceptualize the reconstruction process only in the time-domain. As the signal would appear on an oscilloscope where it has a discrete (usually, the familiar stair-stepped looking) appearance to the eye. Which then tends to make us think in terms of adding missing samples to an discrete appearing signal in order to obtain an smoothly continuous looking signal. The key, I think, is in realizing that the expected smoothly continuous signal is already present BEFORE reconstruction. It doesn't require any creating, as such. It's just being disguised to the naked eye by the image bands. It's the presence of the image bands which makes the signal appear discrete, rather than smoothly continuous when viewed in the time-domain. This leads to mistaken notions that interpolation filters can only guess at or estimate the missing analog signal parts implied by the available sample values.
I find it more intuitive to understand how reconstruction works by viewing the signal in the frequency-domain. Then, it becomes easy to see that, even prior to reconstruction, the original analog signal band is already fully present. It is just accompanied by repeating copies of itself, which are termed image bands. So, if we remove all of these image bands, we are left only with the desired and already present smoothly continuous signal band. It's values didn't have to be guessed or estimated. It didn't have to be reconstructed. It was present all along, just disguised by the now filtered away image bands.
Last edited:
@TNT,
Ken is correct. Upsampling requires LP filtering at some point, and usually most of the LP filtering is done digitally. You may have noticed that some dacs, such as Sabre, have more than one filter you can choose from. All those filters are digital LP filters that define much of the sound of the dac, and the reason they affect the sound as much as they do is because they are doing a major part of reconstruction. If so much of it wasn't done digitally then we would need much more elaborate analog filters to do it in the output stage. However, if you look at a Saber output stage, for example, you can take a balanced output right after the I/V converter stage. Only one very gradual filter pole there, not close to enough for full reconstruction without any help from digital filtering.
Ken is correct. Upsampling requires LP filtering at some point, and usually most of the LP filtering is done digitally. You may have noticed that some dacs, such as Sabre, have more than one filter you can choose from. All those filters are digital LP filters that define much of the sound of the dac, and the reason they affect the sound as much as they do is because they are doing a major part of reconstruction. If so much of it wasn't done digitally then we would need much more elaborate analog filters to do it in the output stage. However, if you look at a Saber output stage, for example, you can take a balanced output right after the I/V converter stage. Only one very gradual filter pole there, not close to enough for full reconstruction without any help from digital filtering.
Last edited:
Trying to understand oversampling - Page 4 - KVR Audio
Since I'm too lazy to tool up matlab (or ____ software) to demonstrate this, I think the visuals in the post (hopefully it goes to the right one, #46), make it so much clearer. The aliases are there, regardless of what we do, but we can much more readily make "good" filters digitally that allow us to relax the analog antialiasing filter after upsampling/reconstruction filtration.
Since I'm too lazy to tool up matlab (or ____ software) to demonstrate this, I think the visuals in the post (hopefully it goes to the right one, #46), make it so much clearer. The aliases are there, regardless of what we do, but we can much more readily make "good" filters digitally that allow us to relax the analog antialiasing filter after upsampling/reconstruction filtration.
What produces the extra numbers? Think about it. You could have zero for the extra samples or repetition for the extra samples; in both cases it could be argued that no digital filter is involved.TNT said:
Then decide you want something else for the extra samples, presumably something related to the input samples. Anything you do to the input numbers in order to obtain the output numbers is a digital filter. Now you could make this digital filter something different from a reconstruction filter, but why? What do you gain by omitting a proper digital reconstruction filter when this is the very reason why oversampling was invented? What you are asking is 'can I do oversampling while avoiding the only reason for doing it?'. It would be like saying 'can I cross the road without actually arriving at the other side?'
Whether you do digital filtering online (in a chip) or offline (in a computer) makes no difference. It is still digital filtering.
Last edited:
In my post that you quoted I did not mention or discuss *filtering* at all - I just want to be clear about that.
I'm just trying to jive with thread starter. I think I understand that in virtually all existing implementations, OS incorporates filtering.
I just wanted to establish whether the 2 process are separable or not - and I think you stated that they are. If that is valuable or not is questionable - yes.
//
I'm just trying to jive with thread starter. I think I understand that in virtually all existing implementations, OS incorporates filtering.
I just wanted to establish whether the 2 process are separable or not - and I think you stated that they are. If that is valuable or not is questionable - yes.
//
Most likely DF96 meant it in a conceptual way; that different implementations can differ in quality is (should be) a given.
As i´ve written in post #12 there is a inconsistency in the nomencalture wrt the terms.
If the terms "oversampling" or "upsampling" are used to describe the process overall, it covers rasing of the sampling frequency (to a new one) _and_ digital low pass filtering to provide the interpolated (new calculated) samples between the original ones.
Unfortunately if the concept is illustrated, in the literature about system theory or digital signal analysis, by building blocks there often occurs a block called "upsampler" or "oversampler" followed (instantly or later down the chain) by another block called "interpolator" .
If the terms "oversampling" or "upsampling" are used to describe the process overall, it covers rasing of the sampling frequency (to a new one) _and_ digital low pass filtering to provide the interpolated (new calculated) samples between the original ones.
Unfortunately if the concept is illustrated, in the literature about system theory or digital signal analysis, by building blocks there often occurs a block called "upsampler" or "oversampler" followed (instantly or later down the chain) by another block called "interpolator" .
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.