It must be remembered that there is a type of 'audiophile' who simply prefers (for reasons which are unclear) any system architecture which departs from whatever he considers to be 'mainstream' and therefore 'boring'. In many cases this forces him to return to architectures (e.g. NOS, SET, no global feedback, full range speakers) which were used in the beginning but were replaced by something better when that became possible. If NOS was mainstream (as it was in the early days of CD) he would prefer oversampling; if SET was mainstream he would prefer PP pentodes with feedback.
When you look at the leftmost oscilloscope picture of post #1, you see something that is much closer to a staircase than to a series of narrow impulses. It is not a perfect staircase, there is something that looks like exponential settling after each step. Presumably that comes from a first-order RC filter.
Anyway, to me it looks like zeroth-order hold and first-order RC filter, rather than like a bunch of Dirac impulses.
I've heard a different version of the story behind phase-linear CD players. It's all hearsay, but I did hear it from a former Philips employee (who had heard it from another Philips employee).
Originally Philips wanted CD to be a 14-bit system, but Sony insisted on 16 bit. Philips then looked for a trick so they could still use the 14-bit DAC they had developed, the TDA1540. They found that with four times oversampling and first-order noise shaping they could get 16-bit signal-to-noise ratio values out of their 14-bit DAC. On top of that, they could then get away with a much simpler analogue reconstruction filter.
As they happened to have a linear-phase FIR optimization program available, they decided to make a linear-phase oversampling filter. By using a Bessel analogue filter, the whole CD player became nearly phase linear. The marketers successfully used that as a selling point, and later other manufacturers followed suit.
Whatever the true story may be, it is indeed weird that having a minimum-phase response rather than a phase-linear response is nowadays sometimes used as a selling point. The steep analogue filters that were so disliked in the 1980's also had minimum-phase responses.
I'm sure the author has the misunderstanding of the theory of Nyquist. If you compare the top left(impulse response of OS) and the top right(impulse response of NOS), the top right is preferable because the waveform is as close as the digital input data. But the comparison has no meaning since it can't exist under the restriction of Nyquist theory.
CD system must not have more than 22.05kHz, though a little bit aliasing is allowed. But "true impulse" like the top right is not allowed(It has the infinite bandwidth). In other words, prefilter of ADC will eliminate most of out of band frequency when recording. What you can actually get from "true impulse " is exactly same as the top left because of the Nyquist theory. The top right can exist virtually but can't do in real music.
The waveform like "true impulse" can be digitally recorded in CD format, but doesn't exist in real world. You don't need to care about imaginary sound. If you have infinite bandwidth, you can have "true impulse." As long as you are under 22.05kHz, your impulse has the ringing like top left. So, the top left is exactly as same as the digital input. There is no trade-off. The comparison of an imaginary waveform is something like division by zero error.
Yes, looking at waveforms of impossible signals is a good way to confuse people. If you are sampling then you must have an anti-aliasing filter. Given that, pure impulses simply do not happen (and they never occur in music anyway, as air and microphones do not have infinite bandwidth). Sadly, confused people seem to often write about digital audio online and in magazines, thus propagating their confusion to others.
Last edited:
It seems that only an ideal low-pass filter with no errors in the time domain, can remove aliases and reveal the original signal before it hits the DAC. However, there are NOS filterless DAC's that do not have any filters. Yet, on subjective listening, these DAC's sound pretty good, and to many, better (subjectively) than what you state must be included in all DAC's. Care to elaborate your statement bit further?
I don't think an ADC without filter would sound acceptable, though, because ultrasonic sounds would alias into the 20 Hz to 20 kHz band. A DAC without filters only produces images at frequencies that are traditionally considered inaudible.
Anyway, like I wrote before, the slight treble roll-off caused by the zeroth order hold can very well explain the thread starter's subjective preference.
Anyway, like I wrote before, the slight treble roll-off caused by the zeroth order hold can very well explain the thread starter's subjective preference.
You are confusing the anti-aliasing filter (removes HF signals which would otherwise cause aliases) before the ADC with the reconstruction filter (removes images) after the DAC. The anti-alias filter does not need perfect time domain performance in order to remove aliases; it just needs perfect frequency domain performance - full attenuation above the Nyquist frequency. It needs good time domain performance to avoid messing with the audio signal.Extreme_Boky said:
I said that sampling needs an anti-alias filter. I did not say that reconstruction needs an anti-alias filter. Faithful reconstruction needs a reconstruction filter (which may be remarkably similar to the antialias filter).
Filterless NOS users have given up on the idea of faithfully reproducing the signal which went into the DAC at the studio. Instead, they prefer a sound which pleases their ears. This has two types of error:
1. HF images (i.e. sounds not present in the original music) - however, fortunately the strongest of these will be precisely time-aligned with HF sounds which were in the original music and as we have no sense of pitch at really high frequencies the combination may sound a bit like HF which was originally there but was removed by the anti-alias filter.
2. HF droop, due to ZOH. This may make the sound smoother.
Just to expand on the HF stuff: say the original music has a percussive sound with major components in the region 15-30kHz. Now most people cannot consciously hear much above 20kHz, but there is some weak evidence that sounds up here can somehow affect perception of sounds lower down in frequency. The anti-alias filter at the studio will remove stuff above around 20kHz, so our original 15-30kHz signal gets sampled and recorded as just 15-20kHz. A proper filtered DAC will then reproduce the 15-20kHz in your home.
A filterless NOS DAC will give you the 15-20kHz, but this will be accompanied by an image at 24.1-29.1kHz (the 15-20kHz reflected from 22.05kHz). This is not the original signal but it occurs at exactly the same time and so the brain could be fooled into thinking that it is. Given that few speakers and ears operate up here I suspect that most of filterless NOS popularity comes from the HF rolloff i.e. it is preferred because it does not sound like the original!
I was thinking of a digital (pre-equalisation) filter before DAC chip (it’s also called FIR filter I think…), but was indeed confused with the terminology of aliasing, anti-aliasing and various names for digital filters, FIR’s used before DAC…. which now I am beginning to classify in my head… slowly. Thank you for the explanation.
I think I got it now.
Anti-aliasing filter removes high-frequency content from an analogue signal before it hits the ADC, the band-pass/cut-off high frequency will depend on a sample rate we want to record at. Later on, we can use a gentle analogue filter after a DAC, to remove what’s left of aliases as the result of sampling of an analogue signal at the recording studio. Please correct this if I’m wrong.
From what I measured, it seems the drop is around -3dB at 20kHz. The input signal was 20kHz wav file ripped off from a Burson CD (sampled at 44.1kHz in studio)
I did an extensive comparison/listening tests with NOS DAC (no digital pre-filtering and no analogue filtering after DAC of any kind) with a DAC that can process PCM (up to 358kHz) and DSD natively at 2.8Mhz (but it will accept double that as well).
My finding is that if I send the 8X oversampled signal to NOS DAC, the frequency response gets ruler-flat far beyond 20kHz. This mitigates the -3dB drop at 20kHz side-effect of ZOH in NOS DAC. Subjectively, the sound had the same high-frequency extension as the other DAC that was processing PCM signal at the same frequency. NOS DAC sounded very nice and extremely spacious – closest to an analogue recording (played on a record player).
The other DAC that was processing PCM signal had well-defined high frequencies, which sounded nice… but after an hour of listening to it, it was easy to realise that its sound was sterile and over-processed. The bass, in particular, was unnatural – it had an initial attack, but it was disappearing too quickly. The sound, as a whole, did not appear to be complete.
However, once I went back to the original sample rate of 44.1kHz, the definition of instruments and voice returned to what my brain tells me is correct, with a NOS DAC. Bass also returned to correct attack, decay and definition. I was able to enjoy the sound again – but the high frequencies were lacking – no doubt about it.
The 8X oversampled signal fed to NOS DAC sounded expansive, but unnatural.
Compared to DSD-capable DAC processing (44.1kHz original signal converted to 2.8kHz DSD in JRiver), the sound was nice, somewhat wide, but it lacked definition. It was easy on the ear, but unnatural. My brain did not know what to think about this sound…. I did not dislike it though. Maybe my brain is wired to a PCM digital sound.
I am not sure which DAC sounds an “original” any more…. None, if you ask me. But, if I must listen to DAC, then I enjoy the NOS DAC with no filters of any kind. I also prefer no oversampling (JRiver selected to “None” oversampling, i.e. original). I also have (very few) “original” 24/192 songs, which sound good through NOS DAC. But, this 24/192 material was produced in a studio where they used reel-to-reel analogue tape as a source…
The DAC's AIF (anti-imaging filter) can do nothing at all about aliases introduced at the recording stage. That's because they're in the audio band. The AIF is entirely concerned with removing images introduced by the DAC because those are all above the audio band.
Yes the 'NOS droop' is about -3.2dB @ 20kHz.
With 8X OS you'll get negligible 'NOS droop'. I have also found increased spaciousness moving from NOS to 2X OS.
Try a passive filter straight after the DAC chip itself. It won't need to be particularly complex given that the frequency domain space available for the transition band is huge, compared to NOS. A two inductor, three capacitor filter probably will do an excellent job.
FIR (Finite Impulse Response) is a type of digital filter. Equalisation is a filter application.Extreme_Boky said:
First part right; second part wrong. Anti-alias filter before ADC to prevent aliasing. Reconstruction filter after DAC to remove images. Images are not aliases or the leftovers of aliases. The reconstruction filter cannot be too gentle.
That would be much worse as a source of hi-fi than almost any competent digital audio chain.
Yes, a reasonably sharp analog filter could be some sort of good compromise between NOS and NONOS.
Look at my attached images. The Butterworth LP filter cuts at around 30khz.
At 2khz square wave output you can notice that there is no pre ringing as in a typical OS DAC.
At 20khz output, the signal resembles a sine wave pretty well.
Now to a second interesting thing, related to this:
If one implements the reconstruction filter as IIR, would that be less aggressive when it comes to "smearing" and pre ringing etc...
The filter below is very easy to implement as IIR. Could it be so that IIR filters are more natural to our ears than FIR filters? The transfer function is similar to a typical IIR. Actually, I think I must experiment a bit with this.
Back to Matlab...
Look at my attached images. The Butterworth LP filter cuts at around 30khz.
At 2khz square wave output you can notice that there is no pre ringing as in a typical OS DAC.
At 20khz output, the signal resembles a sine wave pretty well.
Now to a second interesting thing, related to this:
If one implements the reconstruction filter as IIR, would that be less aggressive when it comes to "smearing" and pre ringing etc...
The filter below is very easy to implement as IIR. Could it be so that IIR filters are more natural to our ears than FIR filters? The transfer function is similar to a typical IIR. Actually, I think I must experiment a bit with this.
Back to Matlab...
Attachments
One more interesting issue.
We all know the dreaded preringing of a pulse in a OS DAC. But the thing is that a single pulse that goes from zero to one and back to zero ( for one sample duration) is impossible to record using normal aliasing filters.
If you feed an ADC with a sharp spike with the duration of around one sample, it will be "smeared out" by the anti aliasing filter quite a bit and there will be no ( or very small)pre ringing in the following DAC.
But I don't really get how such a spike would look like ( in the digital domain) after it has been aliased. Would it exhibit some ringing similar to what the reconstruction filter of the DAC does?
Furthermore, if so are these ringings adding up or are they cancelling out each other. If the latter is true, then the whole NOS-thing will fall into the placebo bin.
But it's more likely that they adds up?
This should be fairly easy to experiment with.
We all know the dreaded preringing of a pulse in a OS DAC. But the thing is that a single pulse that goes from zero to one and back to zero ( for one sample duration) is impossible to record using normal aliasing filters.
If you feed an ADC with a sharp spike with the duration of around one sample, it will be "smeared out" by the anti aliasing filter quite a bit and there will be no ( or very small)pre ringing in the following DAC.
But I don't really get how such a spike would look like ( in the digital domain) after it has been aliased. Would it exhibit some ringing similar to what the reconstruction filter of the DAC does?
Furthermore, if so are these ringings adding up or are they cancelling out each other. If the latter is true, then the whole NOS-thing will fall into the placebo bin.
But it's more likely that they adds up?
This should be fairly easy to experiment with.
I think you should read Peter Craven's article on apodizing filters:
Peter G. Craven, "Antialias filters and system transient response at high sample rates", Journal of the Audio Engineering Society, vol. 52, no. 3, March 2004, pages 216...242
In a nutshell, when you put a filter with controlled pre-ringing at one and only one place in the signal chain, ensure that its stopband starts before the transition bands of all other filters in the chain, and use phase-linear filters with very little passband ripple everywhere else, then the pre-ringing of the entire chain is controlled by that single filter with controlled pre-ringing.
Peter G. Craven, "Antialias filters and system transient response at high sample rates", Journal of the Audio Engineering Society, vol. 52, no. 3, March 2004, pages 216...242
In a nutshell, when you put a filter with controlled pre-ringing at one and only one place in the signal chain, ensure that its stopband starts before the transition bands of all other filters in the chain, and use phase-linear filters with very little passband ripple everywhere else, then the pre-ringing of the entire chain is controlled by that single filter with controlled pre-ringing.
Last edited:
When you make a minimum-phase filter, you get the smallest amount of pre-ringing for a given magnitude response, but especially for filters with a steep roll-off, you also get a far worse phase response than with a linear-phase FIR filter.
IIR filters and lumped continuous-time filters are minimum-phase filters as long as you don't include any all-pass sections. FIR filters can be minimum phase, linear phase, some compromise between the two (such as asymmetrical Wilkinson filters), whatever you want.
SAMPLED IMPULSE:
Theoretical:
When you have a Dirac impulse exactly at a sampling instant and a perfect anti-aliasing filter with a cut-off frequency of fs/2, then all other samples are taken at moments when the filter's impulse response passes through zero, so you get precisely one pulse on the digital side.
Practical:
When the impulse does not exactly occur at a sampling instant and/or the filter is not perfect and/or the filter's cut-off frequency is not exactly fs/2, then you indeed get sampled pre- and post-ringing on the digital side.
CASCADE OF LINEAR-PHASE FILTERS:
Theoretical:
A cascade of two ideal low-pass filters with equal cut-off frequencies is also an ideal filter, and therefore has exactly the same impulse response as a single filter.
Practical:
A cascade of two linear-phase FIR filters with similar cut-off frequencies has longer pre- and post-ringing than a single filter, but the part with the strongest ringing is very similar to what you get from a single filter.
Marcel, it's clear that you know a lot about the math behind digital processing.
And it's also clear that I didn't understand you. What I'm trying to find out is how that "smearing" effect relates to human hearing and what we perhaps can do to improve the subjective impressions.
If a traditional brick wall FIR filter smears things, what can we do to make some kind of compromise between OS and NOS?
So what I want is an open mind.
And it's also clear that I didn't understand you. What I'm trying to find out is how that "smearing" effect relates to human hearing and what we perhaps can do to improve the subjective impressions.
If a traditional brick wall FIR filter smears things, what can we do to make some kind of compromise between OS and NOS?
So what I want is an open mind.
Abraxalito, that PDF you mentioned:
http://www.diyaudio.com/forums/attachments/digital-line-level/643828d1509791550-nos-actually-sense-hierarchical-approach-archiving-distribution-pdf
That was the most interesting document about psycho acoustics I have ever witnessed!
Take this quote that relates to the NOS issue:
"The impulse response of an ‘ideal’ Shannon-sampled
system is a ‘sinc’ function which has a fairly sharp central
pulse but also a pre-ring and a post-ring, which build up
and die away slowly.
Some may wonder how a time-domain analysis can tell
us anything different from a more conventional
frequency-domain analysis, since it is known that the
frequency-domain and time-domain descriptions of a
linear system are completely equivalent. If a human
cannot hear above say 18 kHz, how can a pre-ring at a
frequency of 20 kHz or 22 kHz be of any consequence?
One answer is to consider that a Fourier analyzer uses a
window that extends both forwards and backwards in
time. Thus although the two descriptions are equivalent
if one considers the global signal, the frequency-domain
description is very unhelpful in thinking about the
situation at a particular point in time when the future of
the signal is not known. A neuron has to make a decision
on whether or not to fire on the basis of what it sees now."
http://www.diyaudio.com/forums/attachments/digital-line-level/643828d1509791550-nos-actually-sense-hierarchical-approach-archiving-distribution-pdf
That was the most interesting document about psycho acoustics I have ever witnessed!
Take this quote that relates to the NOS issue:
"The impulse response of an ‘ideal’ Shannon-sampled
system is a ‘sinc’ function which has a fairly sharp central
pulse but also a pre-ring and a post-ring, which build up
and die away slowly.
Some may wonder how a time-domain analysis can tell
us anything different from a more conventional
frequency-domain analysis, since it is known that the
frequency-domain and time-domain descriptions of a
linear system are completely equivalent. If a human
cannot hear above say 18 kHz, how can a pre-ring at a
frequency of 20 kHz or 22 kHz be of any consequence?
One answer is to consider that a Fourier analyzer uses a
window that extends both forwards and backwards in
time. Thus although the two descriptions are equivalent
if one considers the global signal, the frequency-domain
description is very unhelpful in thinking about the
situation at a particular point in time when the future of
the signal is not known. A neuron has to make a decision
on whether or not to fire on the basis of what it sees now."
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.