It isn't clear to me if you are referring to a signal that has been digitally filtered or not, but I don't see how this could be very effective.Steve Eddy said:
Worst case image
Unfiltered: -4dB at 24kHz (and -3dB at 20kHz!!)
8x oversampled, no filter: -48dB at 336kHz
8x over, good digital LP filter: -125dB at 336kHz
Since digital images are inharmonically related, they look more like IMD than harmonic distortion/ Furthermore, even if your preamp/amp has some hundreds of kHz of bandwidth it still has some measurable IMD itself, and all of these images will produce IMD products throughout the spectrum. I would think the largest noticeable effect would be "blurring" or increased noise floor.
In fact, I don't think anyone has mentioned the effect of oversampling on noise:
tiroth said:
I was referring to no other filtering or processing (i.e. oversampling) having been applied. Just the straight 16 bits off the CD into the DA converter. It would only serve to help reduce the magnitude of the images above the audio band.
Sure. And a simple Bessel filter would help reduce that if one were to opt for a non-oversampling solution.
Dunno. Suffice to say that some seem quite satisfied with their non-oversampling, non-filtered systems.
That doesn't make a whole lot of sense. They seem to be implying that oversampling will reduce existing in-band noise. But since in-band noise is indistinguishable from in-band signal, I fail to see how oversampling that which has already been sampled equates to reduced noise.
se
annex666 said:
It beggars belief that you can write this given that the two are so fundamentally different. A square wave or any other wave exists in the analogue domain with continuous time and amplitude. A sample is a discrete time signal with an amplitude at one point in time. It exists in the digital domain. It cannot be any simpler. How on earth do you get from a binary value that represents a single instant to a square wave?
Impulses represent amplitude at a single point. That is all they do.
The bandwidth with information remains the same after oversampling and the images are moved further up the spectrum. But as the spectrum in the digital domain extends to +/- infinity one can hardly call that an increase. Prior to oversampling the images exist at multiples of Fs. After oversampling they exist at multiples of Fs_new. Simple moving something further up a road that already exists does not make the road any longer.
ray.
annex666 said:
Actually, most DSP's for audio use are still fixed point, and all of the integrated oversampling filters that you find inside DAC's are fixed point. This is because fixed point is cheaper to make and beceause a 32-bit fixed point number has higher resolution then a 32-bit floating point number. (A 32-bit floating point DSP has resolution equivalent to a 24-bit fixed point DSP)
tiroth said:
That is a good point you make about IMD distortion, much worse and much more audible than harmonic distortion. So the Bessel filter seems like a minimum, and could work well since most amplifiers have a 1-st order filter at their input as well.
As for lowering noise by oversampling: I believe this only applies to oversampling in the ADC, once the signal has been captured then oversampling again will not lower the noise floor further.
What upsampling *really* is?
Could somebody post an authoritative source that explains the difference between oversampling and upsampling?
I have read most Hifi magazine articles on this ("UP and over", Stereophile articles, letter section rants, etc). Some of these define differences as non-existant while some say there are "differences", without explaining them.
However, in none of my books (Principles of Digital Audio, DSP Book Introduction to Digital Audio, Digital Audio technology) is upsampling defined. Only sample rate conversion (various types), oversampling and re-quantisation.
To me, with my current knowledge, I have defined upsampling as follows:
Upsampling is oversampling (usually with a ratio that is a rational rather than integer number) and/or resampling combined with bandwidth transforms (needed re-quantisation) and dither/noiseshaping techniques on top of those.
Anybody else agree/disagree and could somebody PLEASE provide an authoritative source
Further, I can understand many of the values of oversampling in terms of filter application (without re-quantisation), but I cannot understand the value of resampling + re-quantisation + dither in terms of signal theory.
To me, upsampling (if defined as above) is often done ín practise as resampling, necessitating re-quantisation and in practise making jitter part of the noise floor while at the same time introducing further quantisation artifacts (which can be masked to a degree in audible terms with noise shaping).
However, in signal theory terms, I fail to see the benefits.
What does it gain that basic oversampling does not gain? In terms of signal theory or digital application (now leaving objective sound quality assessment aside for a while)?
regards,
Halcyon
PS I can surely be confused about the terminology, please feel free to set me right, but if you do, please provide that authoritative source on which the 'correct' terminology is based
Could somebody post an authoritative source that explains the difference between oversampling and upsampling?
I have read most Hifi magazine articles on this ("UP and over", Stereophile articles, letter section rants, etc). Some of these define differences as non-existant while some say there are "differences", without explaining them.
However, in none of my books (Principles of Digital Audio, DSP Book Introduction to Digital Audio, Digital Audio technology) is upsampling defined. Only sample rate conversion (various types), oversampling and re-quantisation.
To me, with my current knowledge, I have defined upsampling as follows:
Upsampling is oversampling (usually with a ratio that is a rational rather than integer number) and/or resampling combined with bandwidth transforms (needed re-quantisation) and dither/noiseshaping techniques on top of those.
Anybody else agree/disagree and could somebody PLEASE provide an authoritative source
Further, I can understand many of the values of oversampling in terms of filter application (without re-quantisation), but I cannot understand the value of resampling + re-quantisation + dither in terms of signal theory.
To me, upsampling (if defined as above) is often done ín practise as resampling, necessitating re-quantisation and in practise making jitter part of the noise floor while at the same time introducing further quantisation artifacts (which can be masked to a degree in audible terms with noise shaping).
However, in signal theory terms, I fail to see the benefits.
What does it gain that basic oversampling does not gain? In terms of signal theory or digital application (now leaving objective sound quality assessment aside for a while)?
regards,
Halcyon
PS I can surely be confused about the terminology, please feel free to set me right, but if you do, please provide that authoritative source on which the 'correct' terminology is based
rfbrw said:
Ok I'll try and explain myself simpler - there is a binary number that represents a single sample, this single sample is an approximation of the signal amplitude (approximation due to quantisation steps not being infinitesimal) at the start of the sample. This is sent to the DAC. The DAC outputs a voltage (or current depending on the DAC design) of that approximated level FOR THE DURATION OF THAT SAMPLE!!
If (as I stated in the original post) we were to look at a single sample it is the same as a primary harmonic and many other harmonics that follow those for a square wave.
Samples do not represent a single point in time - they would if the sampling frequency was infinite, but I'm afraid even modern electronics equiptment cannot defy basic physics.
annex666 said:
That's not the point! Analog signals have to be lowpass filtered before they are sampled by an AD-Converter. Meaning ther is an 24kHz lowpassfilter eleminating the signal beond 24kHz. Than one Sample exactly represents the signal for the holdtime. There is no loss of Signal! Only the Quantisation of the Amplitude, lets say to 24 bists, produced an Error.
If you do not apply a lowpassfilter, you will get aliasing effects distorting your signal in the digital domain.
achim
halcyon said:
1.
do whatever is needed to obtain a sample train at the desired sample frequency.
2.
low-pass filter this sample train at half of the original sample frequency.
All else is in the implementation and your insights, tastes, prejudices, ... of whatever constitutes proper low-pass filtering.
That's all there is to it, and yes, the same definition applies to oversampling just as well.
rfbrw,
had you read my earlier posts you would have understood that I have already read four books on the subject.
I think that constitutes as "researching the general topic of sample rate increase".
However, as long as we are discussing upsampling and I can clearly see that different people define the word's meaning differently, we are stuck on semantics.
Had we at least the very basic agreement that upsampling equals oversampling (not a subset, not a particular case, but equal), then we could perhaps go on with more meaningful discussion on the subject.
But, I'm not think we are there yet
That's why I'd really like to see if there is any real definition to upsampling in signal processing theory (or engineerin practise) that is well agreed upon.
Currently the word appears to me as a marketing term describing a certain type of oversampling combined with re-quantisation (most of the upsampling cd players at least work like this, according to their literature and/or chips involved). In signal theory oversampling does NOT necessitate or automatically include re-quantisation or bandwidth transform nor does it include dither and/or noise shaping automatically.
Therefor to me, based on how the word is used in general audio cd player market, upsampling (the marketing term) <> oversampling (the signal theory term).
However, that definition is far from conclusive or authoritative and will not stop our semantic quabble about what the word means and what are it's benefits and so forth.
So you see, my request for a good authoritative term springs up from my lack of understanding of sample rate conversion or resampling theory, but from a genuine effort to try and bring sense to this mess by having some common terms upon which we can build our knowledge collaboratively.
friendly regards,
Halcyon
had you read my earlier posts you would have understood that I have already read four books on the subject.
I think that constitutes as "researching the general topic of sample rate increase".
However, as long as we are discussing upsampling and I can clearly see that different people define the word's meaning differently, we are stuck on semantics.
Had we at least the very basic agreement that upsampling equals oversampling (not a subset, not a particular case, but equal), then we could perhaps go on with more meaningful discussion on the subject.
But, I'm not think we are there yet
That's why I'd really like to see if there is any real definition to upsampling in signal processing theory (or engineerin practise) that is well agreed upon.
Currently the word appears to me as a marketing term describing a certain type of oversampling combined with re-quantisation (most of the upsampling cd players at least work like this, according to their literature and/or chips involved). In signal theory oversampling does NOT necessitate or automatically include re-quantisation or bandwidth transform nor does it include dither and/or noise shaping automatically.
Therefor to me, based on how the word is used in general audio cd player market, upsampling (the marketing term) <> oversampling (the signal theory term).
However, that definition is far from conclusive or authoritative and will not stop our semantic quabble about what the word means and what are it's benefits and so forth.
So you see, my request for a good authoritative term springs up from my lack of understanding of sample rate conversion or resampling theory, but from a genuine effort to try and bring sense to this mess by having some common terms upon which we can build our knowledge collaboratively.
friendly regards,
Halcyon
halcyon said:
Had we at least the very basic agreement that upsampling equals oversampling (not a subset, not a particular case, but equal), then we could perhaps go on with more meaningful discussion on the subject.
But, I'm not think we are there yet
That's why I'd really like to see if there is any real definition to upsampling in signal processing theory (or engineerin practise) that is well agreed upon.
Currently the word appears to me as a marketing term describing a certain type of oversampling combined with re-quantisation (most of the upsampling cd players at least work like this, according to their literature and/or chips involved). In signal theory oversampling does NOT necessitate or automatically include re-quantisation or bandwidth transform nor does it include dither and/or noise shaping automatically.
Well said, I agree 100%
halcyon said:
I doubt that such a definition exists. Or that there is a real need for it in the field of signal processing theory, where marketing does not reign
My two-step definition above is simply based on the two mathematical operations that are strictly required for over/upsampling with the added constrained that the resultant signal spectrum should equal the original spectrum, i.e. no images allowed in the new passband.
All else, i.e. requantization, noise-shaping, fancy model-based interpolation etc. are add-ons that are not related to the *fundamental* act of oversampling. Hence my lumping this in the "insights, taste, prejudices" part.
Not quite. Research (much) older oversampling filters for audio. They included requantisation for the very simple reason that convolving a 16-bit sample stream with any filter yields a sample stream with wordlength in excess of 16 bits.
And the very first one for the 14 bit Philips players (obviously) included noise shaping.
From a mathematical point of view the distinction made in the marketing world is not valid. A sad state, but very very understandable as marketing needs to create 'new' things.
Werner,
good points. If we could agree that 'upsampling' (the marketing) term refers to a bunch of techniques that include oversampling, but do not end there.
Upsampling in signal theory could be agreed upon to be the same process as oversampling.
No qualms about that.
Now, if we want to talk about differences in modern (not old) cd players that use various methods of oversampling, re-quantisation, bandwidth transforms, dither and noise shaping (to mention some of the most important techniques), then I think we should use these terms and just drop the silly 'upsampling' (the marketing term) completely as it just confuses people. It's a loaded term. It can mean so many things to so many people.
BTW, thanks for the info on the older oversampling players! I didn't know that, but it makes sense. I wasn't trying to say that oversampling (old or new) cd players do not necessarily use other techniques, but that in signal theory terms oversampling does not necessitate re-quantisation, dither, noise shaping or lossy (in practise) bandwidth transforms. It's just that in practise these techniques are often combined with oversampling.
So oversampling (technique in signal theory) <> techniques used in an modern (or even old) oversampling cd players (a combination of techniques that are used in conjunction in practise).
1st term is a subset of the second.
I hope we are on firmer ground now. My intention is not to try and define the ultimate truth for words, but to gain common ground for the sake of this discussion (digital audio, particularly cd players).
best regards,
Halcyon
good points. If we could agree that 'upsampling' (the marketing) term refers to a bunch of techniques that include oversampling, but do not end there.
Upsampling in signal theory could be agreed upon to be the same process as oversampling.
No qualms about that.
Now, if we want to talk about differences in modern (not old) cd players that use various methods of oversampling, re-quantisation, bandwidth transforms, dither and noise shaping (to mention some of the most important techniques), then I think we should use these terms and just drop the silly 'upsampling' (the marketing term) completely as it just confuses people. It's a loaded term. It can mean so many things to so many people.
BTW, thanks for the info on the older oversampling players! I didn't know that, but it makes sense. I wasn't trying to say that oversampling (old or new) cd players do not necessarily use other techniques, but that in signal theory terms oversampling does not necessitate re-quantisation, dither, noise shaping or lossy (in practise) bandwidth transforms. It's just that in practise these techniques are often combined with oversampling.
So oversampling (technique in signal theory) <> techniques used in an modern (or even old) oversampling cd players (a combination of techniques that are used in conjunction in practise).
1st term is a subset of the second.
I hope we are on firmer ground now. My intention is not to try and define the ultimate truth for words, but to gain common ground for the sake of this discussion (digital audio, particularly cd players).
best regards,
Halcyon
halcyon said:
No need to limit yourself to new players: the variety in implementation and tricks has always been there:
Earliest Philips players had noiseshaping to allow the use of a 14bit DAC.
Some oversampling filters included phase-prewarp for specific analogue post-DAC filters.
Wadia Spline-based interpolation instead of Sinc(x)-approximation (note: Sinc is the prescribed and mandatory low-pass for a correct reconstruction, see Shannon).
Pioneer Legato: see Wadia.
Denon Alpha: 'creative' generation of new LSBs
HK had a scheme with two multibit DACs, the second delayed somewhat, and then a ramp generator doing linear interpolation in the analogue domain.
Luxman had a machine that emitted a bell-curve mind of shape for each incoming sample.
...
If only modern marketing was applied then
BTW, for those into such trivia ... by 1985 or so everyone had abandoned non-oversampling systems in favour of 2x and 4x oversampling. The non-oversampling retro thing was not something Japanese, not something PQ dreamed up, but was started in about 1988 by Burmester: he had a DAC that offered 16x or none at all, if I remember correctly.
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.