TJF said:
For me it is distortion. Or may be the jitter *spectrum*. Whatever it is, I cannot stand how sigma-delta DAC sound. Somebody more tech savvy than me can find scientific reasons. I trust my ears.
OS/NOS for me is not really an issue, it is actually an overrated debate. Take a state of the art multibit DAC, put a 8x OS filter, remove it and the difference is SMALL, really small.
I have had the pleasure to listen to Peter's system. It sounds better without any filter. Maybe another system will sound better with OS.
My take on CICS:
1) Closed platform. I dont want to use that hardware and I believe that most people want to use whichever computer they have or they can assemble. And still make it sound great, with just a few optimizations.
The more you optimize, the better.
2) I dont believe some of those optimizations correspond to a SQ change. One for all: memory. I want all the possible ram, in excess to 4GB to preload high-res files. CICS cannot. It is not future-proof. It's good to play only redbook material.
What are these win7 latencies?
IIRC they have improved the audio stack. I have W7 build 7022 x64 and for what I can tell, xxhe sounds the same as in Vista (sp1), but i couldnt ABX, since it is a dual boot machine.
PS: I never install beta service packs.
I don't understand your point. Changing between negative full-scale and positive full-scale in a single sample is absolutely a valid piece of data (how else would you represent a full-scale sine wave up near the nyquist frequency?). Oversampling is irrelevant to any of this discussion - it's not needed to resolve this signal. It's all perfectly* there already.
(* or at least, perfect to within the tolerance of the sample depth you're using. But that's not relevant either.)
Of course it is. I was just explaining that this happens, that those transients just are there, and that they should be respected (since we can't know better anyway). However, there is my point about the data being molested ...
If we
- Accept such data as being there frequently (we must, because it just is);
- Turn this into a test signal, which is a square wave (the difference would be that in music data the amplitude doesn't stay high/low as long and even as a square wave);
- Run this through a fs64 DAC;
Then we see no transient left. A sine is the result, if only the frequency is high enough. Of course, the sine still has a transient, but from a dead bird. Exactly how OS sounds btw (ok, let the latter be subjective
)So, besides one may not like that for SQ (harmonic response) result, it is the worst for measurements. There would be no way I can measure back the transient fed, when no transients come out, right ?
Also note that while music data will not contain the squares with lasting amplitude as from test data, it does contain the steep transients, and just because (!) the amplitude doesn't stay long, the effect will be that the amplitude gets lower. Another error from OS. So, draw a triangle with one of the corners at the top, and start rounding (oversampling). The top will get lower and lower and lower.
Btw, I am explaining what I perceive from sound, not the other way around. And it all fits, so far.
But I am open to it.
Peter
Source is sampled at 44100 at all times.
Oversampling 3000Hz square (ESS Sabre).
Filtered of course, but unrelated.
Oversamping 10000Hz square (ESS Sabre).
Ditto.
NON-Oversampling 22050Hz square (just to overdo it). Filterless. As steep as electronics permit. Post-ringing (PCM1704).
NON-Oversampling 22050Hz square. Filtered. Transient is less, ringing has gone (PCM1704).
A long time ago, when NOS was rather "new" there was a long thread at Hydrogen about squares could not be heard through speakers anyway. Hmm. Look above and you know why.
This, btw, was the same thread which was about the famous UDial test tone (an around 19.x KHz varying tone, inaudible to most, but impeeding for phone-dialing tones which really weren't in the data -> aliasing ... and which could burn tweeters).
Anyway, this is what I'm talking about, and by itself it's nothing new I guess. The only thing I say is :
When everything is deformed like you see above, and knowing that harmonics spring from squariness hence transients, the harmonics are lost by heavy oversampling.
This is very much audible, and hardly subjective. One just needs to know what to listen for.
As said, the downsides of it are in quite another league, and to my firm opinion we can't - or should not - say that those downsides are to be preferred over loosing the harmonics.
We better try to keep them right, instead of throwing them away in advance and all is lost.
Please correct me wherever I'm wrong.
Peter
PS: Square waves like you see above spring from speakers, just travel through air as well, can be captured by microphone just the same, and form their own harmonics in that mid air. It really does. You won't do that by PC speakers though.
Oversampling 3000Hz square (ESS Sabre).
Filtered of course, but unrelated.
Oversamping 10000Hz square (ESS Sabre).
Ditto.
NON-Oversampling 22050Hz square (just to overdo it). Filterless. As steep as electronics permit. Post-ringing (PCM1704).
NON-Oversampling 22050Hz square. Filtered. Transient is less, ringing has gone (PCM1704).
A long time ago, when NOS was rather "new" there was a long thread at Hydrogen about squares could not be heard through speakers anyway. Hmm. Look above and you know why.
This, btw, was the same thread which was about the famous UDial test tone (an around 19.x KHz varying tone, inaudible to most, but impeeding for phone-dialing tones which really weren't in the data -> aliasing ... and which could burn tweeters).
Anyway, this is what I'm talking about, and by itself it's nothing new I guess. The only thing I say is :
When everything is deformed like you see above, and knowing that harmonics spring from squariness hence transients, the harmonics are lost by heavy oversampling.
This is very much audible, and hardly subjective. One just needs to know what to listen for.
As said, the downsides of it are in quite another league, and to my firm opinion we can't - or should not - say that those downsides are to be preferred over loosing the harmonics.
We better try to keep them right, instead of throwing them away in advance and all is lost.
Please correct me wherever I'm wrong.
Peter
PS: Square waves like you see above spring from speakers, just travel through air as well, can be captured by microphone just the same, and form their own harmonics in that mid air. It really does. You won't do that by PC speakers though.
Pictures nb. 1 and 2 are correct - that is the continuous-system waveform of the samples your discrete-system generator with fs=44.1kHz has produced. Do not forget your discrete system is limited to 22kHz and cannot generate squares. It is only the deficiency of your graphical software that it displays straight lines/squares instead of the correct waveforms as on the first two pictures.
Pictures nb. 3 and 4 are obviously wrong. System with frequency-limited transfer function cannot produce straight lines. If it does, the waveform is "straightened" and "sharp-edged" by aliases causing serious distortion of the original signal (which could not be squares in the first place).
You chose 22050Hz for obvious reasons - the squares look just nice (exactly 2 points per period). Why don't you take pictures e.g. for 20kHz "squares"? For NOS you will hardly recognize the shape (if you make your scope synchronize at all), while OS will produce decent curly squares - just as it should be. Maybe you will want to post the resultant pictures for others to see.
Pictures nb. 3 and 4 are obviously wrong. System with frequency-limited transfer function cannot produce straight lines. If it does, the waveform is "straightened" and "sharp-edged" by aliases causing serious distortion of the original signal (which could not be squares in the first place).
You chose 22050Hz for obvious reasons - the squares look just nice (exactly 2 points per period). Why don't you take pictures e.g. for 20kHz "squares"? For NOS you will hardly recognize the shape (if you make your scope synchronize at all), while OS will produce decent curly squares - just as it should be. Maybe you will want to post the resultant pictures for others to see.
Hi Peter
That's a nice demonstration of your point, so I see what you mean now.
However, you've missed a fundamental property of sampling theory: the square waves you show aren't really meant to be square waves. A group of positive-full-scale samples placed directly after a group of negative-full-scale samples (and repeating) is not a square wave. It looks a lot like one if you join them up with a smooth line, but that's misleading, and isn't what the samples mean. They've come out as a square wave in the unfiltered case because those are actually distorted by the lack of filtering. There are harmonic frequencies there which the digital system is not capable of controlling, and what you're seeing is distortion.
For example, try generating a full-scale sine wave just near the Nyquist frequency - it'll come out as a square wave. Clearly not what you want.
That's a nice demonstration of your point, so I see what you mean now.
However, you've missed a fundamental property of sampling theory: the square waves you show aren't really meant to be square waves. A group of positive-full-scale samples placed directly after a group of negative-full-scale samples (and repeating) is not a square wave. It looks a lot like one if you join them up with a smooth line, but that's misleading, and isn't what the samples mean. They've come out as a square wave in the unfiltered case because those are actually distorted by the lack of filtering. There are harmonic frequencies there which the digital system is not capable of controlling, and what you're seeing is distortion.
For example, try generating a full-scale sine wave just near the Nyquist frequency - it'll come out as a square wave. Clearly not what you want.
Ok ...
1. This thread indeed is not about that. On that matter all last posts are not. So, I'll stick with some misconceptions for now.
2. phofman, you are right. But only because I indeed was unlucky to use a frequency I should not have. I did not think about that, and the 22050 was just because it is the max to be used.
3. Wingfeather, of course I agree, but I look from the other standpoint. So, besides, again, I used pictures I shouldn't have (and which were taken for a different purpose really), I just *want* the 22050 as a sine to look as it does on those last pictures. That is how the data is, and I don't want it smoothened.
And I don't want that because everything will be smoothened because of it otherwise. That this leaves me with a pile of harmonic distortion ... well, I've said it earlier.
All 'n all when the data shows transients, those transients should come out. If we only can agree on that (not difficult), we can debate on how to do it (ok, not here), what will destroy it in the first place (I say heavy OS) and whether it will ever be possible without a pile of distortion when you start off right. And I know, shifting Nyquist to infinity will, but this is infinit OS. The wrong start.
IMO.
1. This thread indeed is not about that. On that matter all last posts are not. So, I'll stick with some misconceptions for now.
2. phofman, you are right. But only because I indeed was unlucky to use a frequency I should not have. I did not think about that, and the 22050 was just because it is the max to be used.
3. Wingfeather, of course I agree, but I look from the other standpoint. So, besides, again, I used pictures I shouldn't have (and which were taken for a different purpose really), I just *want* the 22050 as a sine to look as it does on those last pictures. That is how the data is, and I don't want it smoothened.
And I don't want that because everything will be smoothened because of it otherwise. That this leaves me with a pile of harmonic distortion ... well, I've said it earlier.
All 'n all when the data shows transients, those transients should come out. If we only can agree on that (not difficult), we can debate on how to do it (ok, not here), what will destroy it in the first place (I say heavy OS) and whether it will ever be possible without a pile of distortion when you start off right. And I know, shifting Nyquist to infinity will, but this is infinit OS. The wrong start.
IMO.
Peter
I sort've see what you mean. But I don't see why - why would you want what is meant to be a sine wave to come out as a square wave? That's just a very heavily-distorted sine anyway.
It's fundamental to the operation of the digital system that you have the filtering in there, or else you get distortion. It's totally integral. And so, anything that "looks a bit like" a fast transient in the digital data (i.e. one that contains any components over Fs/2) isn't actually that sort of transient at all. Drawing a smooth line through the points is not the correct way to evaluate what the data means. Those samples don't represent a square wave (because they can't ever mean that) and aren't supposed to be interpreted that way.
I can see why, from a purist standpoint, one might not want what looks like "smoothing" of their data, but for DSP this filter is really not optional. Taking it off simply breaks everything and makes your samples look like a square wave that they never were. The most important point is that nothing is gained from doing this (in terms of transients or anything else), you just make your output plain wrong.
I sort've see what you mean. But I don't see why - why would you want what is meant to be a sine wave to come out as a square wave? That's just a very heavily-distorted sine anyway.
It's fundamental to the operation of the digital system that you have the filtering in there, or else you get distortion. It's totally integral. And so, anything that "looks a bit like" a fast transient in the digital data (i.e. one that contains any components over Fs/2) isn't actually that sort of transient at all. Drawing a smooth line through the points is not the correct way to evaluate what the data means. Those samples don't represent a square wave (because they can't ever mean that) and aren't supposed to be interpreted that way.
I can see why, from a purist standpoint, one might not want what looks like "smoothing" of their data, but for DSP this filter is really not optional. Taking it off simply breaks everything and makes your samples look like a square wave that they never were. The most important point is that nothing is gained from doing this (in terms of transients or anything else), you just make your output plain wrong.
If I may second the comprehensive explanation of Wingfeather - the conversion between discrete (digital) and continuous (analog) systems is NOT done by joining the samples with straight lines. I think this is the fundamental principle which causes a lot of confusion on this website. For performance reasons, graphical programs chart the waveforms this way and people wrongly expect to see the same shapes coming from analog outputs. I used to make the same mistake, eventhough I have masters from the theory of systems - bad me http://www.diyaudio.com/forums/showthread.php?postid=918499#post918499
http://www.diyaudio.com/forums/showthread.php?postid=918499#post918499
Well, I stopped reading that thread here : http://www.diyaudio.com/forums/showthread.php?postid=920377#post920377 which just says it all, when it comes down to my opinion.
In the end, phofman, we're all on the same boat, come from the same origine, and head towards the same objectives.
I, for one, clearly don't like upsampling just because of the sound coming from it. So, consensus on the "straight lines", assuming that I understand what you mean by it. Do note though, that I always look at the data itself (the numbers) which may not be all that obvious or common.
Wingfeather, I know and understand what you mean, and you are right were it for the theories. This, however, does not hold back that there is a differences between those theories disobeying the net result for sound quality, and the not-theories disobeying the theories but sound better.
It is difficult. Like the knowledgeable in that thread say : distortion/error might be of a flavor one likes. Of course this is the way out for those who disobey the theories and want to be right on the "it sounds better". But I am to deep in this all to believe that.
It is difficult because it is worse than apples and oranges, and only the net result of all brings us "sound". For (stupid) example, do we exactly know that the harmonic distortions are indeed audible ? They are lower than the original signal ...
Taken into account that I am into 100% representation of instruments only (I even buy them to compare !) I cannot recognize that the distortions (which are way bad indeed !!) make sound worse or less natural for that matter. No such thing. But when I even slightly oversample (my own, rabbit, SRC in the DAC, whatever) the sound is killed. It just is.
The last picture of those I showed earlier is from a nice 2nd order Bessel, and it doesn't show the anomalies we in the end are talking about. An FFT everybody would like. But it does not sound better, and I mean : instruments get LESS real from it.
I must add to it, and I think this is important, that my amps cope to 200KHz easily, IOW they at least won't get stressed by hf (unreal !) harmonics. Then I will also add that I don't even use a preamp to filter out (flatten) undesired (again UNREAL) alisasing coming from it, so indeed I am listening to the mess the FFT clearly shows.
Lastly, because of the ongoing work on XXHighEnd I just learned from experience that the better those stupid digital waves are passed through to the analogue domain, the better it gets. So yes, I kind of like to present a 2-step 22050 Hz sine frequency to the analogue chain as a pure square wave. Stupid theory : no matter the digital steps, pass them through 1:1 and net sound gets better. And no, I don't say that I'd rather listen to a 22050 square instead of the original sine (might I be able to hear it), but this is the difference with "net".
The opposite, the ESS Sabre (in Buffalo implementation) does nothing nothing nothing. I cannot help that. It shows the exact opposite of natural instruments. It sounds very "easy" though. Much more easy than NOS. Easy for matters of not much requiering from the rest of the chain.
There is more to this than we currently know, and I just want to find out what it is.
As I told in between the lines earlier, I am very confident that the harmonics coming to my ears for a large part emerge in mid air. I mean : have base tones (but including the "square" data) in the data, throw them out of the speakers, and the harmonics will form in air anyway. Think about this please.
It might even go as far that what happens in air cancelles out the wrongness of it. Oh yes, highly pretentious that is.
Guys, please trust me, this all goes way way further than just comparing theories, most probably because we don't know the proper theories yet. It also goes as far as using the PROPER digital volume in order to leave out the preamp. Avoid that decent volume, and you have nothing to start with. Who knows that ? who actually *uses* that ? I do ... and it is a base for leaving out the preamp which is the very base to ever hear what's in there. Again, please trust me. I'm working on this too long now for not to know ...
Wasn't this thread about bit perfect ? Well, if this would come down to the one stupic LSB being dithered (as how XP does it), no problem. I think I can hear that, but I am far from sure.
This is very different from Vista, where resampling takes place (if you're careless that is) and a 44.1 is resampled to 96 and back. Do I hear that ? you bet I do.
Earlier I think TJF said : XXHighEnd = NOS. He got that right.
It is my only discrete (hence possible) means to improve on sound. If that weren't so a. XXHE would sound lousy and b. I would be working on DSPs instead.
Now let's all move on. As said, we're really on the same boat.
Wingfeather, thank you. I am sure you are more educated on these matters than I am, but at least now you know why I think like this.
Thanks,
Peter
In the end, phofman, we're all on the same boat, come from the same origine, and head towards the same objectives.
I, for one, clearly don't like upsampling just because of the sound coming from it. So, consensus on the "straight lines", assuming that I understand what you mean by it. Do note though, that I always look at the data itself (the numbers) which may not be all that obvious or common.
Wingfeather, I know and understand what you mean, and you are right were it for the theories. This, however, does not hold back that there is a differences between those theories disobeying the net result for sound quality, and the not-theories disobeying the theories but sound better.
It is difficult. Like the knowledgeable in that thread say : distortion/error might be of a flavor one likes. Of course this is the way out for those who disobey the theories and want to be right on the "it sounds better". But I am to deep in this all to believe that.
It is difficult because it is worse than apples and oranges, and only the net result of all brings us "sound". For (stupid) example, do we exactly know that the harmonic distortions are indeed audible ? They are lower than the original signal ...
Taken into account that I am into 100% representation of instruments only (I even buy them to compare !) I cannot recognize that the distortions (which are way bad indeed !!) make sound worse or less natural for that matter. No such thing. But when I even slightly oversample (my own, rabbit, SRC in the DAC, whatever) the sound is killed. It just is.
The last picture of those I showed earlier is from a nice 2nd order Bessel, and it doesn't show the anomalies we in the end are talking about. An FFT everybody would like. But it does not sound better, and I mean : instruments get LESS real from it.
I must add to it, and I think this is important, that my amps cope to 200KHz easily, IOW they at least won't get stressed by hf (unreal !) harmonics. Then I will also add that I don't even use a preamp to filter out (flatten) undesired (again UNREAL) alisasing coming from it, so indeed I am listening to the mess the FFT clearly shows.
Lastly, because of the ongoing work on XXHighEnd I just learned from experience that the better those stupid digital waves are passed through to the analogue domain, the better it gets. So yes, I kind of like to present a 2-step 22050 Hz sine frequency to the analogue chain as a pure square wave. Stupid theory : no matter the digital steps, pass them through 1:1 and net sound gets better. And no, I don't say that I'd rather listen to a 22050 square instead of the original sine (might I be able to hear it), but this is the difference with "net".
The opposite, the ESS Sabre (in Buffalo implementation) does nothing nothing nothing. I cannot help that. It shows the exact opposite of natural instruments. It sounds very "easy" though. Much more easy than NOS. Easy for matters of not much requiering from the rest of the chain.
There is more to this than we currently know, and I just want to find out what it is.
As I told in between the lines earlier, I am very confident that the harmonics coming to my ears for a large part emerge in mid air. I mean : have base tones (but including the "square" data) in the data, throw them out of the speakers, and the harmonics will form in air anyway. Think about this please.
It might even go as far that what happens in air cancelles out the wrongness of it. Oh yes, highly pretentious that is.
Guys, please trust me, this all goes way way further than just comparing theories, most probably because we don't know the proper theories yet. It also goes as far as using the PROPER digital volume in order to leave out the preamp. Avoid that decent volume, and you have nothing to start with. Who knows that ? who actually *uses* that ? I do ... and it is a base for leaving out the preamp which is the very base to ever hear what's in there. Again, please trust me. I'm working on this too long now for not to know ...
Wasn't this thread about bit perfect ? Well, if this would come down to the one stupic LSB being dithered (as how XP does it), no problem. I think I can hear that, but I am far from sure.
This is very different from Vista, where resampling takes place (if you're careless that is) and a 44.1 is resampled to 96 and back. Do I hear that ? you bet I do.
Earlier I think TJF said : XXHighEnd = NOS. He got that right.
It is my only discrete (hence possible) means to improve on sound. If that weren't so a. XXHE would sound lousy and b. I would be working on DSPs instead.
Now let's all move on. As said, we're really on the same boat.
Wingfeather, thank you. I am sure you are more educated on these matters than I am, but at least now you know why I think like this.
Thanks,
Peter
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.