Does audio upsampling really work or is it more of a marketing term?

I am looking into a new CD player, and will ultimately listen myself, but I was curious to know what the general consensous is. The marketing says it can take a regular CD and make is sound close to a SACD or DVD A. Well I am one of the guys who notices a HUGE difference in CD vs DVD A/ SACD so if this is the case i would see huge value in this feature.

On a side not IF it does work and makes a difference why has there been no audio upsampling available in home theater processors for 5.1 like there has been for CD players?

Does audio upsampling really work or is it more of a marketing term?

I am looking into a new CD player, and will ultimately listen myself, but I was curious to know what the general consensous is. The marketing says it can take a regular CD and make is sound close to a SACD or DVD A. Well I am one of the guys who notices a HUGE difference in CD vs DVD A/ SACD so if this is the case i would see huge value in this feature.

On a side not IF it does work and makes a difference why has there been no audio upsampling available in home theater processors for 5.1 like there has been for CD players?

You would notice a difference between CD and the hi-res if the CD and the hi res are differently mastered or EQed. Otherwise, hardly and a recent JAES paper demonstrated that.


http://www.mlssa.com/pdf/Upsampling-theory-rev-2.pdf

http://www.thetadigital.com/upsampling.htm

Upsampling/resampling is not marketing if the player has the hardware and is actually changing the original signal from 44.1 redbook to the new target rate. But whether or not upsampled digital audio sounds better than 44.1 redbook is pretty subjective, especially if you've heard the album as vinyl, or redbook cd originally.

I've heard Anagram licensed asynchronous sample rate conversion implemented in a friend's Audio Aero player. That definitely changed the sound of an ordinary redbook CD. I thought it sounded great. It was also really expensive...

You might not like the effect is has on your favorite albums, however. I was really leaning towards getting an AA Capitole mKII for a while, but kept away due to the high cost of the unit.

I think most multichannel receivers do not feature audio upsampling due to cost (licensing+extra system parts) and additional software complexity. Most folks, given the choice of video upscaling from 480 to 1080P (vs audio resampling to DSD) would opt for video.

- Steve O.

Cambridge Audio's CD player 840C and 740C is a great entry point (well undr $2k) for someone looking for Anagram 16-bit/44.1kHz CD data to 24-bit/384kHz upsampling.

http://www.audioholics.com/education/audio-formats-technology/upsampling-vs-oversampling-for-digital-audio

Upsampling and oversampling are effectively synonymous, although oversampling is usually 1) limited to integer multiples and 2) refers to what is done in many DAC chips themselves.

I for one am of the opinion that non-integer upsampling should not be done. The implementation of a non-integer upsampler is more difficult, increasing the likelihood of error and/or degradation, and there is no benefit to a non-integer multiple. So I would not trust anyone that claims it is better to upsample RedBook to 96/192/384kHz instead of to the nearest integer counterparts 88.2/176.4/352.8kHz. Note that DSD's sample rate is 64*44.1kHz, so upsampling to DSD is fine, in this regard at least.

As to whether it is worth doing at all, I think the debate is still out. I tend to favor it because I think it makes the analog design easier. There is really nothing magic about it, though, and I do think that standalone upsamplers are kind of a waste. But the proof is in the (blind) listening :)

EDIT: An interesting tidbit in the article is that some people advocating doing no analog filtering at all. This is interesting because from a signal processing standpoint that leaves a lot of artifacts in the signal. But those artifacts are all in the >20kHz range and should be inaudible---as long as it doesn't affect our perception in other ways (or our tweeters). I'd love to see some studies done on that question.

http://www.audioholics.com/education/audio-formats-technology/upsampling-vs-oversampling-for-digital-audio

Upsampling and oversampling are effectively synonymous, although oversampling is usually 1) limited to integer multiples and 2) refers to what is done in many DAC chips themselves.

I for one am of the opinion that non-integer upsampling should not be done. The implementation of a non-integer upsampler is more difficult, increasing the likelihood of error and/or degradation, and there is no benefit to a non-integer multiple. So I would not trust anyone that claims it is better to upsample RedBook to 96/192/384kHz instead of to the nearest integer counterparts 88.2/176.4/352.8kHz. Note that DSD's sample rate is 64*44.1kHz, so upsampling to DSD is fine, in this regard at least.

As to whether it is worth doing at all, I think the debate is still out. I tend to favor it because I think it makes the analog design easier. There is really nothing magic about it, though, and I do think that standalone upsamplers are kind of a waste. But the proof is in the (blind) listening :)

EDIT: An interesting tidbit in the article is that some people advocating doing no analog filtering at all. This is interesting because from a signal processing standpoint that leaves a lot of artifacts in the signal. But those artifacts are all in the >20kHz range and should be inaudible---as long as it doesn't affect our perception in other ways (or our tweeters). I'd love to see some studies done on that question.

In the early days, or soon after the first CD players came on the market, the makers started to advertise that their players oversampled by 2x on up to 8x.
I don't think this practice went away, just the advertising did.

Agreed. Nowadays it seems like oversampling is just one of several "tools" in the digital to analog conversion process, along with noise shaping, single-bit/multi-bit/full-bit, etc.

IIRC many users of the DCS Purcell prefer even multiple upsampling, that is 24/176.4 over 24/192.
( When I had a Purcell mated to a Wadia 27ix, I prefered 24/88.2 over 24/96 )

The Purcell also offers several options for noise-shaping and other digital parameters, that can affect the sonic outcome.

- Andy

EDIT: An interesting tidbit in the article is that some people advocating doing no analog filtering at all. This is interesting because from a signal processing standpoint that leaves a lot of artifacts in the signal. But those artifacts are all in the >20kHz range and should be inaudible---as long as it doesn't affect our perception in other ways (or our tweeters). I'd love to see some studies done on that question.

Hi Michael,

If you search out the work David Greisinger did on some of the ultrasonic "higher resolution" formats and intermodulation distoriton, you will find some interesting things that relate to the matter you bring up. My recollection of his testing was that some electronics would have enough non-linearity to create sub-harmonics that actually get into the realm of audibility, although not very easily audible unless the circuit is clipped.

Here is David Geisinger's personal website (http://world.std.com/~griesngr/) where he had a PPT presentation on Intermodulation Audibility (http://world.std.com/~griesngr/intermod.ppt#1). I seem to recall another paper/presentation by David/Lexicon that I read a few years ago, but didn't have time to search if the above is the same. Even D.G. doesn't assert this research to be all encompassing, but it does correct some major assumptions or throw plenty of "gotchas" into casual investigations and observations made by listeners. :rolleyes:

The implementation of a non-integer upsampler is more difficult, increasing the likelihood of error and/or degradation, and there is no benefit to a non-integer multiple.

I don't believe Benchmark would agree with this. They resample everything in their DAC1 to 110kHz. This A(asynchronous)SRC is claimed to have benefits but I don't recall what they are (sorry). It might be on their website which has quite a bit of information. I believe the Lavry Black, another pro model DAC, also does ASRC.

You know, that's really fascinating---110kHz? What an odd number to choose! I have great respect for those guys so I have to guess they had a very specific reason why they chose it---maybe having something to do with their oscillator/clock design goals. If anyone could make sure to do ASRC right it's them.

When Head-Fi.org is up again, I can point you to the specific post where Benchmark explains why they picked 110 kHz, but apparently through their measurements, that's where the DACS they used perform best. The ASRC also helps in their jitter rejection system. The DAC-1 is one of the most amazing pieces of audio engineering I've seen at any price.

--Andre

As promised, here's a link to Head-Fi where a Benchmark engineer explains why they use 110 kHz:

http://www.head-fi.org/forums/f7/benchmark-dac1-now-available-usb-223006/index68.html#post3039703

Basically, currently available DAC chips don't perform well at 192 kHz.

--Andre

it works. for little $$$ you can add the behringer ultramatch src 2496 to your system, and upsample 16/44.1 to 24/96 (asynchronous). async upsampling helps reduce jitter (which tends to harden the high frequencies). the difference is definitely audible as a smoother, more natural, less harsh sound.

theories abound as to why it works. i don't know for sure why. my latest theory goes like this: i think it has to due partly with the quantization resolution at high frequencies. it's hard to explain, but if you draw a sinewave on a piece of paper and imagine it to represent a wave at 1/2 the sampling frequency (22.05KHz) and at full-amplitude, you only get to record 2 amplitude measurements per one full wave cycle. depending on when you start to sample, you may sample all the zero-crossings. or you may sample the absolute max/min of the peaks and troughs. or any +/- amplitude pair in between. in other words, you have lost all amplitude resolution accuracy. ok, that's at 1/2 the sample rate -- where we know we are beyond the theoretical limits of sampling. what about at 1/4 the sample rate (11.025KHz). now you get 4 measurements per wave cycle. but again, depending on where in the wave cycle to start, you will get quite a bit of variation on values, which could hit all the zeros and max/min points. or may hit the points where the +/- values are of equal distance above and below zero (e.g., +16000, -16000, +16000, -16000 -- i made those values up). so i'd say you have compromised amplitude accuracy. mathematically upping the amplitude fineness (more bits) while upping the rate (higher frequency before amplitude resolution is materially compromised) should yield a waveform closer to the original analog waveform.

i know this is a bit of a hand-waving argument -- and i invite anyone to tell me why it's full of water -- but i began to arrive at this conclusion after working on a program to perform my own jitter tests at home which mimick the ones performed in stereophile magazines test measurements. it was in discussing the difference between the maximum signal level and the word value for an 11.025KHz test jitter signal with the esteemed editor of that magazine that got this thought going.

food for thought...

my latest theory goes like this: i think it has to due partly with the quantization resolution at high frequencies. it's hard to explain, but if you draw a sinewave on a piece of paper and imagine it to represent a wave at 1/2 the sampling frequency (22.05KHz) and at full-amplitude, you only get to record 2 amplitude measurements per one full wave cycle.


I don't think this is the reason, because you've just described a broken sampler. The highest frequency must strictly be less than half the sampling rate, so a 44.1 kHz system cannot sample a 22.05 kHz signal correctly. It can correctly sample lower frequencies.

--Andre

I don't think this is the reason, because you've just described a broken sampler. The highest frequency must strictly be less than half the sampling rate, so a 44.1 kHz system cannot sample a 22.05 kHz signal correctly. It can correctly sample lower frequencies.

--Andre

Damn that Nyquist guy always butting in to conversations. ;)

And in practice there's a steep input filter to prevent aliasing of the signal with inputs 1/2fs and higher.

Up sampling... It all might be a mute point. HD-AAC was just announced as a viable option to replace CDs. The new media discs would be digital lossless songs at 24-bit/96KHz which is superior to CD.

Plus who says double dipping is only exclusive to the movie studios?

It may allow one format so all media players can play them. No more AAC vs. MP3 issues.

Watch them mess this one up with DRM too.

Oh forget it... DVD-AUDIO and SACD were superior too... look what happened to them. And don't forget they'll charge double the money to prevent people from really adopting the new format.

I would like to see a higher resolution format adopted by the public though.

yes, i'm well aware of the nyquist limit, and anti-aliasing pre-sampling filtering -- but i was attempting to explore why that limit is there. but please ignore that, and consider the attached image. imagine it represents the sampled 1/4fs sinewave at maximum level (which [I]is[I] within the range of what the system should be able to sample correctly). note that the sample values (yellow dots) are not occurring at the maximum word values (+/- 32767), but are instead less than that. (let's assume they're around +/- 16000).

now imagine that you dial down the amplitude of the sinewave in 32767 equal steps. to capture all of those different amplitude levels would require 32767 integers between the maximum level signal values an zero. but because our maximum signal level is already giving sample values less than the maximum word value, there are not enough values between it and zero. i.e., the quantization accuracy is less than is implied by a 16 bit sample word.

oddly, if the exact moment of sampling were delayed by 1/4 the sinewave cycle (phase shifted by 90 degrees), then the sample values would be at 0, +32767, 0, -32767 -- which would retain full 16 bits of resolution if the above example were repeated (32767 steps in amplitude down to zero).

Provided you're within the Nyquist criteria, which implies band limiting, with those two points, there can exist one and only one function which passes through those points.

As far as HD-ACC, just what the consumer needs. To act as a cash cow for the companies that will be doing lousy mastering while looking to funnel money into the politicos on both sides of the specturm to limit your rights by creating software and hardware that does so.

Agim, the problem you describe is a quantization issue, not a sampling rate issue. It's actually a pretty good description of one of the idiosyncrasies of quantization, and one of the reasons why extra bits and dithering are so important when doing any sort of mixing or word size adjustment.

As promised, here's a link to Head-Fi where a Benchmark engineer explains why they use 110 kHz:

http://www.head-fi.org/forums/f7/benchmark-dac1-now-available-usb-223006/index68.html#post3039703

Basically, currently available DAC chips don't perform well at 192 kHz.

--Andre

Alas, no ABX tests to show whether it makes any audible difference or not.

Dan Lavry's own white papers have outlined the reasons why 192 kHz sampling is pointless and perhaps even detrimental. He suggests that practical and theoretical problems with Redbook sampling rate are solved when the rate is moved up to ~ 60 Hz. If asynchronicity is considered an issue then the logical rate would be 88.2 kHz.

Agim, the problem you describe is a quantization issue, not a sampling rate issue. It's actually a pretty good description of one of the idiosyncrasies of quantization, and one of the reasons why extra bits and dithering are so important when doing any sort of mixing or word size adjustment.
i suppose that's fair enough -- maybe i am just realizing for the first time that the quantization error is realted to the sampling rate. DSD encoding takes it to the limit, where a 1 bit quantization value paired to a very high sampling rate provides the equivalent information of a multi-bit word at a lower sampling rate.

If you search out the work David Greisinger did on some of the ultrasonic "higher resolution" formats and intermodulation distoriton, you will find some interesting things that relate to the matter you bring up. My recollection of his testing was that some electronics would have enough non-linearity to create sub-harmonics that actually get into the realm of audibility, although not very easily audible unless the circuit is clipped.


You are totally correct. Anyone who has studied electronics will attest the effect of beat frequencies and inadvertent ring modulation and their ability to introduce or reintroduce AF information from non AF components. Even when not audible as AF it can manifest as phase cancellation and add to standing wave issues.

could i ask, how many on this thread have actually inserted an upsampler into their home system? i use one, and can reliably pick it out as sounding better in a blind test (though i can't think how to arrange a double-blind test). and for $130, anyone else can try it too -- using the behringer ultramatch src2496 for 24bit/96KHz (or 88.2KHz) upsampling. set the unit for the internal clock (not "dig in"), 24 bits, dither, pre-emphasis off -- then flip back and forth between it and the straight source (if you use the clock button to toggle between "digital in" and "internal", the unit will revert to the sample rate of the source when set to "digital in" and then back to 96KHz or 88.2KHz when you toggle back to "internal").

i think that you'll easily hear the improvement. what i don't know, however, is whether it is truly due to the upsampling, or rather due to a reduction of jitter in the source signal. i worked extensively to minimize the jitter in my system, performing the same tests that are routinely performed in stereophile reviews (the 11.025KHz special test signal and FFT spectral analysis of the output), and can see that the 96KHz upsampled signal has demonstrably less jitter than the source signal. as an aside, this testing also prompted me to remove a jitter-reducing device which proved superfluous or even detrimental, and helped to choose the digital cabling scheme that performed best from a jitter perspective.

Whether it is truly due to the upsampling, or rather due to a reduction of jitter in the source signal.Or due to RMS level differences due to subtle scaling performed during the upsampling process, or due to differences in the DAC's analog performance at different sample rates. I'm not sure that test is really a single-variable A/B test as it may appear.

Quite so Michael. There are more things of a crude nature to measure (RMS level differences) that can provide additional insight. Unless your jitter rises to pathological levels, reducing it may be more an excercise in technical accomplishment that doesn't rise to the level of audibility.

It's also a Behringer, which is never a good indicator for good implementation or transparency.

--Andre

Or due to RMS level differences due to subtle scaling performed during the upsampling process, or due to differences in the DAC's analog performance at different sample rates. I'm not sure that test is really a single-variable A/B test as it may appear.
true - these are potential variables. but i would submit that it is impossible to eliminate the factor of the DAC performing differently at different sampling rates, as every conceivable test is subject to at least that variable (or the greater variable of using a different DAC entirely).

btw, the RMS levels measure identically.

but i'll ask again -- who here has actually listened to an upsampler and concluded that it was not efficacious? and please state which brand/model was auditioned.

but i would submit that it is impossible to eliminate the factor of the DAC performing differently at different sampling rates, as every conceivable test is subject to at least that variable (or the greater variable of using a different DAC entirely).I'm happy to concede that for the sake of argument. But that necessarily means that you can't conclude that the digital process of upsampling is responsible for the differences you heard.

But I still think it is possible to better narrow down the causes. If the differences were RMS level differences, then I think it's reasonable to say the test was flawed. But maybe it wasn't: maybe it was jitter reduction, for instance. But in that case, there are other ways to reduce jitter that don't require upsampling.

Or your particular DAC might very well perform better at a higher sampling rate; but others might do just the opposite. After all, designing low-noise DACs is more difficult at higher sampling rates. As AndreYew's head-fi link points out, the Benchmark folks use this argument to justify their choice of a 110kHz target rate.

I do seem to remember that AVS Forum member tzucc concluded that on his dCS stack, upsampling by an integer factor (44.1->176.4) outperformed noninteger upsampling (44.1->192).

I'm happy to concede that for the sake of argument. But that necessarily means that you can't conclude that the digital process of upsampling is responsible for the differences you heard.

But I still think it is possible to better narrow down the causes. If the differences were RMS level differences, then I think it's reasonable to say the test was flawed. But maybe it wasn't: maybe it was jitter reduction, for instance. But in that case, there are other ways to reduce jitter that don't require upsampling.

Or your particular DAC might very well perform better at a higher sampling rate; but others might do just the opposite. After all, designing low-noise DACs is more difficult at higher sampling rates. As AndreYew's head-fi link points out, the Benchmark folks use this argument to justify their choice of a 110kHz target rate.

I do seem to remember that AVS Forum member tzucc concluded that on his dCS stack, upsampling by an integer factor (44.1->176.4) outperformed noninteger upsampling (44.1->192).
nor would it be possible to conclude that it isn't.

so where are we then? the variables possibly in play are:
1) upsampling itself
2) jitter attenuation of asynchronous conversion
3) the DAC performing less optimally at 44K than at 96
4) scaling to a different RMS level (below measurable difference in my experiment, but generally could be in play)
5) placebo effect

that article about the benchmark DAC1 stated that the DAC performance at 110KHz was experimentally found to be optimal -- and, in that person's opinion, the asynchronous upsampling was beneficial both as a means of jitter attentuation, and because it provides a better match to the rate at which the DAC's analog performance was optimized.

it's hard to imagine a way to separately test for these variables, since the playback chain inextricably binds them together. any ideas?

I take it that you've recorded some of your physical measurements as well as a record of the blind tests that were performed, agim. Can you share those?

Quite so Michael. There are more things of a crude nature to measure (RMS level differences) that can provide additional insight. Unless your jitter rises to pathological levels, reducing it may be more an excercise in technical accomplishment that doesn't rise to the level of audibility.
btw, what level of jitter would you consider "pathological"?

We had a good discussion on this over in this thread here:

http://www.avsforum.com/avs-vb/showthread.php?t=908665&highlight=jitter

It's not a long thread and there are a number of useful links, too.

true - these are potential variables. but i would submit that it is impossible to eliminate the factor of the DAC performing differently at different sampling rates, as every conceivable test is subject to at least that variable (or the greater variable of using a different DAC entirely).

btw, the RMS levels measure identically.

but i'll ask again -- who here has actually listened to an upsampler and concluded that it was not efficacious? and please state which brand/model was auditioned.



And I'll ask, what were your blind test scores? And how did you implement blinding and level-matching?

While marketing, here is pretty good write up on D/A upsampling.
Personally, I agree with the premise that the only way to be truly sucessful with upsampling, is to make sure you remove the most amount of jitter from the signal, before its upsampled. The only way to do that consistently, would appear to be with a Word Clock, clock link, master clock, etc...

http://en.wikipedia.org/wiki/Word_clock

The very best upsampling CD players I have ever heard, have all had this function. I will say the Esoteric units, but there is also Wadia and DSC, among others, I am sure.
So, I fully disagree that you just tacking on an upsampling device will improve the sound of the music, as it will essentially magnify, the good and the bad.

Here is a pro model.

http://www.apogeedigital.com/products/bigben.php

Personally, I agree with the premise that the only way to be truly sucessful with upsampling, is to make sure you remove the most amount of jitter from the signal, before its upsampled.I'm not sure what you mean here. Upsampling is an entirely digital process, and as such jitter is completely irrelevant to the quality of the process. Jitter intrudes only at the point of digital to analog conversion.

btw, what level of jitter would you consider "pathological"?

That would depend on its distribution, no? But it seems to me that most players out there are having no problem hitting a few hundred ps. If a player were doing a couple of orders of magnitude more, then I'd worry about the vendor. Michael has provided a link for a general discussion.

And I'll ask, what were your blind test scores? And how did you implement blinding and level-matching?

I'm hesitant to do so, b/c I'm certain it will be picked apart for its imperfection -- but bear in mind, this is not my day job, so I don't have a whitepaper on the experiment, just scribbled notes on a pad.

Level matching by running the output into an edirol USB 24/96 sound card, feeding into freeware named Visual Analyzer 8, and using the RMS checkbox on the voltmeter in that software to test that levels were matched while playing a pink-noise file generated in Cool Edit Pro.

Blinding was accomplished by closing my eyes. My assistant would mute the preamp, flip a coin, and for heads would set the unit to 96KHz internal clock, for tails set it to the "digital in" clock, meaning bypass upsampling. She would also toggle various buttons on an off to throw off any unconcsious attempts to guess at the current settings by listening for button presses. Each trial consisted of playing back both settings -- the coin flip randomly determined which came first.

BTW, the DAC in the preamp was used, not the DAC in the upsampler. RCA S/PDIF cables were used from a Squeezebox to the upsampler, then to the preamp.

The source track was a rip of Nickel Creeks "A Lighthouse's Tale" performed with Exact Audio Copy software.

The results: in 9 of 10 trials, the setting I identified as sounding better was the 96KHz setting.

I'm sure this won't convince anyone -- but here's what I think is missing from this conversation: actual experience using upsamplers. It is an interesting exercise to theorize about whether upsampling could do anything positive, and I fully appreciate healthy skepticism on a scientific basis. However, I think that the scientific approach normally starts with an observed phenomenon, then attempts to explain it with theories that can be refuted. Here, I see no evidence of anyone observing the phenomenon in the first place.

It's good to doubt any claims that defy obvious logic -- but when people who care about sound quality (e.g., the writers for stereophile magazine) claim to hear improvements, I for one trust that it may be possible that they are hearing a real effect. So I test to see whether I can hear it first, then later consider why I can. Of course there's no magic invovled -- but that doesn't mean that all of the relevant factors have been measured or adequately understood either. The original post asked whether upsampling was anything more than marketing hype -- I guess the answer is that we don't know. It might be only hype, or it might be that different implementations are more or less effective based on the algorithm used, or it may be that the upsampling itself has little to do with any observed improvement, and rather changes to jitter or better/worse matching to the DAC is responsible for the change in sound.

And while my $130 upsampler improves the sound in my system, I can also report that my experience with the upsampler built into the Outlaw 990 pre/pro (a $1000 unit) did nothing to improve the sound thru that unit, which to my ears completely mangles the source signal, performing obvious scaling prior to the DAC.

agim, we've had discussions about upsampling before---for example, I mentioned some previous discussions about tzucc's dCS upsampler; we've also had discussions about DSD conversion with live experience with dCS and Meitner stacks. That this thread doesn't bring them all back out is simply more an indication that it's not an active topic now. So yes, the phenomenon has been observed that upsamplers can make a difference, though opinions about what the best approach is are varied.

As for me I've had the privilege of listening to both systems but haven't done the kind of testing required to offer an experienced view. I'm really not questioning that you heard differences more than I am offering skepticism about the reasons why. Seems to me that you're offering a reasonable amount of skepticism yourself which means there's not all that much disagreement here.

Seems to me that you're offering a reasonable amount of skepticism yourself which means there's not all that much disagreement here.

agreed. i just wanted to make sure the starting point is observing a difference, without which it's merely an academic discussion. indeed i am skeptical as to the reason for the difference. you have to be, given that various factors which may influence the result cannot really be tested independently.

there is one more thing that is interesting. i mentioned i had performed jitter spectrum analysis akin to those performed by JA in stereophile. while i don't know the algorithm for determining the peak-to-peak mesaurement in picoseconds, i was able to graphically compare the various settings. i found that there were many more sidebands around the 1/4fs signal at either 44.1 or 88.2 than at 48 or 96, and that 48 and 96 appeared identical from a jitter perspective, as did 44.1 and 88.2. SUBJECTIVE STATEMENT TO FOLLOW: if this is a good indicator that the level of jitter is, in fact, identical for 48 and 96, then from listening i would have to conclude that about 75% of the difference is down to jitter, with the remaining 25% due to 96 vs 48. if you grant that the 25% difference is real (an unproven assertion, i know), and were to conclude that it is merely due to the DAC performing better at 96 than at 48, would it be fair to say that upsampling is not merely hype, but can lead to improved performance apart from jitter considerations and amplitude scaling?

agim,

IMO, while your thirst for knowledge on this topic and willingness to experiment are admirable, you likely won't find many other people around here as excited about the topic as you are. Many people, as you have likely noticed already (I say likely, because I have a lot of those people on ignore, so I'm not following their parts of the argument except where your quotes reveal parts of their posts), are going to be more interested in discrediting any conclusions of audible difference on your part. Even if you prove there is an audible difference, they will attempt to discredit any significance to the difference... Not due to any overwhelming interest in the truth on their parts sadly enough. :(

I'm not referring to Grant of course, he tends to be one of the few "true objectivists" around on these fora. I am enjoying the discussion you both are having.

Wow, they changed the "edit" timer. I guess I don't have 4 to 5 minutes to edit my posts anymore... No more Mr. Lazy.

I don't plan on staying subscribed to this thread, but I'll enjoy reading it every once in a while as it gets larger.

And while my $130 upsampler improves the sound in my system, I can also report that my experience with the upsampler built into the Outlaw 990 pre/pro (a $1000 unit) did nothing to improve the sound thru that unit, which to my ears completely mangles the source signal, performing obvious scaling prior to the DAC.

I've got to give you mad props for at least trying to cover the bases, far beyond what most such claimants do. I'd rather see something like 20 trials, and double- rather than single- blinding (did your 'proctor' leave the room after finally setting the source?). And I'm afraid it's still not clear from your tests whether upsampling itself is the 'culprit' in the difference you heard.

Is oversampling as well as upsampling taking place in your signal chain?

agim, can you outline how you did your hookups to the Behringer unit? It's unclear to me from your posts. Thanks.

I've got to give you mad props for at least trying to cover the bases, far beyond what most such claimants do. I'd rather see something like 20 trials, and double- rather than single- blinding (did your 'proctor' leave the room after finally setting the source?). And I'm afraid it's still not clear from your tests whether upsampling itself is the 'culprit' in the difference you heard.

Is oversampling as well as upsampling taking place in your signal chain?

when i have more time, i make take up a more rigorous test. but for now i've satisfied my original goal, which was increased listening enjoyment -- not necessarily to prove upsampling one way or the other. i agree it's not absolutely clear whether upsampling itself is the culprit -- only that the upsampler device has had a positive effect.

the DAC is 96/24, so i do not think oversampling is taking place (which i understand to be possible only at integer multiples of the base sampling rate). i find no mention of oversampling in the specs or user manual.

agim, can you outline how you did your hookups to the Behringer unit? It's unclear to me from your posts. Thanks.

sure.

[squeezebox]---coax s/pdif---[behringer]---coax s/pdif---[preamp]

for the record, i tried various hookups, including coax/toslink, toslink/coax, toslink/toslink -- and at least a couple of different makes of each. i was actually surprised that the purpose-built s/pdif coax cable sounded and measured no better than a plain markertek 75-ohm coax cable (as in, for cable TV) that had F-to-RCA connectors tightly screwed onto either end (with pliers) -- in fact, it measured ever-so-slightly worse.

i have a question -- after thinking about that benchmark DAC and its 110KHz async sample rate conversion. does anyone have a theory as to why there would be an optimal match at such a rate? is it like over-clocking a PC -- in that you can push it only so far before the electrical circuit's settle-time exceeds the clock-tick interval, at which point it is no longer stable? if so, that would explain why you wouldn't want to mae it run too fast -- but doesn't explain why slower wouldn't be just as good.

what other variables may be involved? i'm thinking about the DAC's rise-time and its ability to hold the output steady at a given current or voltage level without ringing or sagging as potential factors -- but maybe someone with more EE insight could comment.

agim, there is a clear reason to run the DAC as fast as possible: the higher the DAC's sample rate, the gentler the analog reconstruction filter can be. So I'm assuming that Benchmark wanted to run their DAC as fast as they could without encountering other issues with its performance. Your theories about what those issues might be sound good.

agim, there is a clear reason to run the DAC as fast as possible: the higher the DAC's sample rate, the gentler the analog reconstruction filter can be. So I'm assuming that Benchmark wanted to run their DAC as fast as they could without encountering other issues with its performance. Your theories about what those issues might be sound good.

somehow, my post from yesterday disapeared. in a nutshell, we could theorize:

1) we want to run the DAC at a high sampling rate, to push the reconstruction filter's efffects farther out of the audio band.
2) we must not run it so high that the electrical properties of the components are "overclocked" to the point that the do not have enough settle-time, rise-time, too much ripple when holding between samples, or something in that realm.

now, finding the optimal balance point would be tricky -- but assuming that it doesn't happen to coincide with a integer multiple of the 44.1K original rate -- the implication is that asynchronous upsampling is necessary to achieve the optimal performance from the DAC & reconstruction filter.

All our posts from 5AM on disappeared due to a server failure...

agim, there is a clear reason to run the DAC as fast as possible: the higher the DAC's sample rate, the gentler the analog reconstruction filter can be.

That makes sense, but someone should tell that to the chip companies, who all seem to build brickwall filters only. Even Wolfson who go on about how much street cred their chips have don't use a shallower filter. I believe the Benchmark has a brickwall, too.

--Andre

Andre, I think you're confusing analog and digital filters here, because I was just looking at Wolfson's data sheet for their highest-end stereo DAC and I saw nothing about analog filters at all---only digital filters.

A proper reconstruction filter pretty much must be a brick wall, at least for 44.1kHz and 48kHz. (*) The problem is that it is very hard and/or very expensive to build an analog brickwall filter, let alone one that avoids significant phase issues. Upsampling allows the bulk of the filtering to be done in the digital domain where the phase issues can be avoided and achieving a steep brick wall is not nearly as difficult.

(*) (I know there are some upstarts who think that ripping out the reconstruction filter altogether has its advantages, but I think they are a tiny minority, even among high-end designers.)

In fact, before the AVS server crash, I had posted a question here that went something like this: does anyone even make all-analog reconstruction filters for audio anymore? Many of the higher-end DACs like the Wolfson use a sigma-delta architecture, which are upsampling by design, and necessitate digital filtering.

Another less-frequently mentioned advantage of upsampling and digital filtering is that it provides an opportunity to correct for sin(x)/x rolloff in the DAC. A standard (i.e., not sigma-delta) native-rate 44.1kHz DAC is going to be down about 3.2dB at 20kHz. So assuming you care to fix that, the subsequent analog filter must not only provide for a brick wall between 20-22kHz but provide an inverse sin(x)/x response between 0-20kHz. If you're doing upsampling, that sin(x)/x rolloff reduced---to 0.75dB with just 2x upsampling, for instance. And you can fix that by rolling the correction into the same phase-friendly digital filter you're using to interpolate.

Andre, I think you're confusing analog and digital filters here, because I was just looking at Wolfson's data sheet for their highest-end stereo DAC and I saw nothing about analog filters at all---only digital filters.


i think there is generally a lot of confusion surrounding analog and digital filters / anti-aliasing and reconstruction filters. anti-aliasing has to do with the A/D (recording) part of the process, reconstruction with the D/A (playback) part. there was a thread lost in the crash where someone posted about anti-aliasing filters, and how the professional market "knew" that a properly implemented one obviated the need for above-44.1K sampling -- but that is irrelevant to the playback side of the chain, where we're already stuck with 44.1K and whatever errors were made in the recording chain, but are wondering how to optimize the D/A conversion for playback.

note: in my setup, the behringer device is used as digital-in and digital-out -- i.e., neither the ADC nor the DAC is in the playback chain (i.e., neither its anti-aliasing filter nor reconstruction filter are in the loop).

Agim, you fail to remember the lost post where Chu answered your objection; he was making a point about filtering per se, not anti-aliasing filtering particularly. The 'optimization' at the D/A end for redbook, involves oversampling (resampling) and anti-imaging filters. No benefit from upsampling -- which is equivalent of leaving the signal in its 'oversampled' state at output, rather than downsampling back to 44.1 -- has ever been demonstrated.

I also made a point that Dick Pierce's observations re: the importance of well-designed filtering stages over higher sample rates, are fully in line with those of Dan Lavry, who designs high-end ADCs and DACs for a living, and the results of tests by Bob Katz, as documented in his book 'Mastering Audio', where the 'benefits' of 96 kHz sampling were found to reside not in the higher sample rate per se, but in the filtering involved.

When you feed the Behringer a 44.1 signal, does it do a straight bypass to the preamp or does it 'mess' with it somehow? I take it your assessment that the 96 was better than 44.1 was predicated on passing the same signal through the Behringer and then doing a subjective comparison. Yes?

Andre, I think you're confusing analog and digital filters here, because I was just looking at Wolfson's data sheet for their highest-end stereo DAC and I saw nothing about analog filters at all---only digital filters.

Maybe we're talking about different things. The analog filters after oversampling won't be brickwalls, because that's the point of oversampling DACs. The digital filters used for 44.1 kHz oversampling DACs have to be brickwalls because the transition band is so tiny. But what about 96 kHz DACs whose digital filters could start attenuating right above 20 kHz gently down to 48 kHz?

--Andre

When you feed the Behringer a 44.1 signal, does it do a straight bypass to the preamp or does it 'mess' with it somehow? I take it your assessment that the 96 was better than 44.1 was predicated on passing the same signal through the Behringer and then doing a subjective comparison. Yes?

there are two modes that can be used -- there is the one where it resamples to 44.1 using its internal clock, and another where it simply passes the clock on the digital input. i use the latter for comparison -- in that case, it should not be messing with the signal, though one could imagine it may be altering the jitter performance. i need to look back at my jitter test to see if i tested that path, and if so, how it compared to physically bypassing the unit.

Agim, you fail to remember the lost post where Chu answered your objection; he was making a point about filtering per se, not anti-aliasing filtering particularly. The 'optimization' at the D/A end for redbook, involves oversampling (resampling) and anti-imaging filters. No benefit from upsampling -- which is equivalent of leaving the signal in its 'oversampled' state at output, rather than downsampling back to 44.1 -- has ever been demonstrated.

I also made a point that Dick Pierce's observations re: the importance of well-designed filtering stages over higher sample rates, are fully in line with those of Dan Lavry, who designs high-end ADCs and DACs for a living, and the results of tests by Bob Katz, as documented in his book 'Mastering Audio', where the 'benefits' of 96 kHz sampling were found to reside not in the higher sample rate per se, but in the filtering involved.

as i recall, the dr. pierce reference was specifically discussing anti-alias filters. regarding oversampling/upsampling -- filtering needs are what push for higher rates (over- or up-sampling should either fit the bill), but DAC electrical properties may set an upper limit for optimal performance (rise,settle,hold,etc). so unless that sweet spot just happens to be at an integer multiple, upsampling would be needed to reach it.

fyi - i'm leaving this thread. while i am impressed by the amount of knowledge and research and thought going into many of these posts, i am rather turned off by the agressive and sometimes pissy attitude of some posters.

my interest lies not in reading what so-called experts said, or what some company says, or what anyone says, per se; nor in theorizing in a vacuum. it lies in actually experimenting, with the goal of making changes to my system to increase my listening enjoyment. i would encourage anyone actually interested in musical satisfaction to experiment for themselves, and take the conversation out of the vacuum. that behringer i've used is dirt-cheap (the cost of a dinner out), and if you don't like what it does for the sound, i'd bet you can unload it on ebay and recoup at least half the cost. what's the downside?

anyway, thanks for the conversation... farewell.

there are two modes that can be used -- there is the one where it resamples to 44.1 using its internal clock, and another where it simply passes the clock on the digital input. i use the latter for comparison -- in that case, it should not be messing with the signal, though one could imagine it may be altering the jitter performance. i need to look back at my jitter test to see if i tested that path, and if so, how it compared to physically bypassing the unit.
I'd like to see the graphs if you've got them. What happens if you don't bypass but resample 44.1 to 44.1? Redundant and silly I know, but...

Also were you able to take measurements at the speaker terminals in both sampling cases to check the voltages when running a test tone? It would at least rule audible differences due to level out.

Oh, you're leaving? Bye.

But what about 96 kHz DACs whose digital filters could start attenuating right above 20 kHz gently down to 48 kHz?Andre, filters really can't just "start attenuating" in any particular place; their attenuation is a continuous, smooth function of frequency (with the exception of exact nulls). A filter can't have exactly 0dB of attenuation from 0-20kHz and then start dropping off from there; in order to "start" attenuating at 20kHz it must necessarily have been attenuating before that, too. This effect is more pronounced the "gentler" (lower-order) the filter is.

So to keep the attenuation negligible in the 0-20kHz range, you really do need to make the filter high order, and start the transition band as far away from audible as possible. I really see no downside to doing that. I would if the filters were analog, but not for digital filters whose phase response can be tightly controlled (or even made completely linear phase).

EDIT: Also, an octave isn't a whole lot of room for a gentle filter to work in anyway. I mean, a 20-22kHz transition band is supertight, don't get me wrong, but 20-48kHz isn't exactly a walk in the park. You still need a fairly high-order filter for that.

[Oh, nevermind... not worth it.]

Andre, filters really can't just "start attenuating" in any particular place; their attenuation is a continuous, smooth function of frequency (with the exception of exact nulls).

Yes, you're right --- sorry for the sloppy wording. How about "the place where the filter qualitatively changes behavior?" :)

An octave is still very tight, especially if you have to drop 120+ dB, but I wonder if there's value to relaxing the transition band as much as possible so that its time-domain behavior is as short as possible, and doesn't interact with the top critical band of the ear. JJ's conjectured about this before, and seems to believe a 5-kHz-ish transition would be sufficient (implying that a 50 kHz sampling rate should be enough):

http://www.pgm.com/pipermail/proaudio/2007-April/003290.html
http://www.pgm.com/pipermail/proaudio/2007-April/003302.html

Both explain the basic situation, but that whole thread contains more subtle things throughout.

--Andre

It is entirely possible (in fact, easy) to design a filter that minimizes impulse response width or pre-ringing energy subject to requirements on passband ripple and stopband attenuation. But I'm just not sure that's necessary. Time and frequency domain effects are intimately related, as you know; and pushing the brick wall as far as possible into the ultrasonic ought to in turn push the pre-ringing well into the inaudible range.

I haven't done the MATLAB analysis jj is suggesting, though. I don't have the signal processing toolbox he's proposing to use, though I could design the equivalent filters using my own optimization tools, which are far more flexible. Maybe I'll sit down and do that at some point.