Hi Ken,
I fully agree that listening should be an emotional and not at all an intellectual experience, being unique to each and every individual.
But coming back to the earlier questioned title of the thread, did you notice that Lampie even preferred the sound of his Nos dac after first having oversampled the content?
So I’m curious to hear what consequence this could have for your initial opinion regarding the sound differences between NOS and OS.
Hans
I fully agree that listening should be an emotional and not at all an intellectual experience, being unique to each and every individual.
But coming back to the earlier questioned title of the thread, did you notice that Lampie even preferred the sound of his Nos dac after first having oversampled the content?
So I’m curious to hear what consequence this could have for your initial opinion regarding the sound differences between NOS and OS.
Hans
Getting back to the question in the title of the thread: probably the slight treble roll-off due to the zeroth order hold, although it could also be related to intersample overshoots or pre-echoes related to passband ripples.
Experiments you could do to find out:
1. ABX test of a NOS DAC and an identical DAC running at four times the sample rate that gets the same sample four times in a row. Maybe you can use the same DAC for A and B and just switch the clocks.
2...n. If you don't notice any difference in 1, do ABX tests with an FPGA or DSP running various filter algorithms: repeating the sample four times, linear-phase filter with 0.1 dB passband ripple, filter with 0.00001 dB ripple, the 0.00001 dB filter convolved with uniform weighting over four samples and so on. Use attenuated digital signals when you want to check the sound without intersample overshoots and peak sample normalized signals when you do want to include intersample overshoots.
Checking the hypothesis of Hans would require an analogue filter after the DAC that you can switch into and out of the signal path in an ABX manner, while keeping the passband gain constant to within 0.1 dB.
Experiments you could do to find out:
1. ABX test of a NOS DAC and an identical DAC running at four times the sample rate that gets the same sample four times in a row. Maybe you can use the same DAC for A and B and just switch the clocks.
2...n. If you don't notice any difference in 1, do ABX tests with an FPGA or DSP running various filter algorithms: repeating the sample four times, linear-phase filter with 0.1 dB passband ripple, filter with 0.00001 dB ripple, the 0.00001 dB filter convolved with uniform weighting over four samples and so on. Use attenuated digital signals when you want to check the sound without intersample overshoots and peak sample normalized signals when you do want to include intersample overshoots.
Checking the hypothesis of Hans would require an analogue filter after the DAC that you can switch into and out of the signal path in an ABX manner, while keeping the passband gain constant to within 0.1 dB.
Last edited:
You get to have the last word on the above, as I don't wish us to further divert the thread.
Hi, Hans,
I haven't overlooked your suggestion. The truth is, I was not sure what response to give you because, I have a DIY AD1865 based DAC which I designed and built specifically for conducting various experiments in NOS. Well, I recently performed one experiment too many on the PCB of that DAC and trashed it.
So, I need to build a new NOS DAC.
I also have a DIY PCM1794A based DAC which I also use for experiments and which still functions - for now. One of the lesser reasons I launched this thread is because I want to design and build another NOS DAC incorporating the latest thinking on them from all of you first.
As far as your proposed experiment, perhap someone on this thread might perform it and report their results back to us?
Okay, gang,
I've collected a list of what we have suggested as suspected technical reasons for why NOS sounds different from OS. Please add to the list any suspected reason which I've overlooked. After the list is finished, I hope we can apply some deductive reasoning (Mr. Holmes, I presume?) to arrive at one or two prime suspect reasons. Then, devise experiments which prove or disprove their guilt.
Here is what I have so far. In no particular order:
1) SINC aperture based -3dB droop @ 20KHz.
2) Lack of an FIR interpolation-filter, freeing the DAC from certain processing 'artifacts' , such as:
a) time-domain signal echoes produced within Equiripple on-chip FIR filters.
b) impulse response ringing (pre or post)
c) half-band filters plainly violating Nyquist.
3) Phase-modulation of the baseband signal due to insufficiently suppressed image-bands. In other words, because the signal waveform is not fully reconstructed according to the sampling theorem requirements.
4) The unsuppressed image-bands are, somehow, producing audible IM products directly within the ear.
5) Different jitter impact due to fewer D/A conversion cycles per second.
6) Reduced supply and ground noise due to slower clock rates.
7) Converter settling-time becomes a smaller percentage of each conversion period as the conversion rate becomes slower.
I've collected a list of what we have suggested as suspected technical reasons for why NOS sounds different from OS. Please add to the list any suspected reason which I've overlooked. After the list is finished, I hope we can apply some deductive reasoning (Mr. Holmes, I presume?) to arrive at one or two prime suspect reasons. Then, devise experiments which prove or disprove their guilt.
Here is what I have so far. In no particular order:
1) SINC aperture based -3dB droop @ 20KHz.
2) Lack of an FIR interpolation-filter, freeing the DAC from certain processing 'artifacts' , such as:
a) time-domain signal echoes produced within Equiripple on-chip FIR filters.
b) impulse response ringing (pre or post)
c) half-band filters plainly violating Nyquist.
3) Phase-modulation of the baseband signal due to insufficiently suppressed image-bands. In other words, because the signal waveform is not fully reconstructed according to the sampling theorem requirements.
4) The unsuppressed image-bands are, somehow, producing audible IM products directly within the ear.
5) Different jitter impact due to fewer D/A conversion cycles per second.
6) Reduced supply and ground noise due to slower clock rates.
7) Converter settling-time becomes a smaller percentage of each conversion period as the conversion rate becomes slower.
2c and 3 actually don't make sense, and 2d is missing:
2c Much as I dislike halfband filters, they do suppress images better than a zeroth order hold and they have less roll-off below the Nyquist frequency. That is, they violate Nyquist less than a zeroth order hold.
2d NOS DACs are usually less prone to clip on peak sample normalized recordings, the intersample overshoot issue.
3 Inadequately suppressed images don't phase-modulate the baseband signal, they just add a lot of ultrasonic crap to it. A zeroth order hold is a phase-linear low-pass filter, albeit a very poor one.
2c Much as I dislike halfband filters, they do suppress images better than a zeroth order hold and they have less roll-off below the Nyquist frequency. That is, they violate Nyquist less than a zeroth order hold.
2d NOS DACs are usually less prone to clip on peak sample normalized recordings, the intersample overshoot issue.
3 Inadequately suppressed images don't phase-modulate the baseband signal, they just add a lot of ultrasonic crap to it. A zeroth order hold is a phase-linear low-pass filter, albeit a very poor one.
Last edited:
This is a very good idea for test and I think most will find oversampled one doesn't sound as good.
Foobar resampler has ZOH filter you can use for it
4 consecutive correct identifications on an ABX test is nothing like definitive evidence:
You'll see that I use an illustrative example there based on codecs.
I am a big fan of ABX tests as they are the closest to being scientific that we can be using our ears and perceptual machinery not measuring instruments. What is being measured is the ability to hear a difference which is the crucial first step before we get into the complicated world of preferences, ranking etc. They do take very careful setting up though, and a correct analysis which is not always obvious. I used to do this as part of my role for KEF in the early 90s when Laurie Fincham and Richard Small headed up the R&D dept, and then later as a consultant to a University tonmeister course for research.
The ability to set these up in the digital domain with levels set exactly equal and the double blind/choice capture completely automated is very appealing.
You'll see that I use an illustrative example there based on codecs.
I am a big fan of ABX tests as they are the closest to being scientific that we can be using our ears and perceptual machinery not measuring instruments. What is being measured is the ability to hear a difference which is the crucial first step before we get into the complicated world of preferences, ranking etc. They do take very careful setting up though, and a correct analysis which is not always obvious. I used to do this as part of my role for KEF in the early 90s when Laurie Fincham and Richard Small headed up the R&D dept, and then later as a consultant to a University tonmeister course for research.
The ability to set these up in the digital domain with levels set exactly equal and the double blind/choice capture completely automated is very appealing.
There a number of established sensory test protocls; a brief introduction: https://www.apps.fst.vt.edu/extensi...ory Analysis/Sensory Analysis - Section 4.pdf
A/B for sound identification to me should probably work as follows:
1) test subject may listen to A and or B as many times as desired.
2) when test subject feels ready, either A or B is randomly played.
3) test subject chooses that random sound is either A or B.
4) if no forced choice, then test subject can pass on answering: "I don't know."
Also, one of the most important things about perceptual testing is to train the test subject to reach a steady level of proficiency with the test apparatus using test sounds of graded difficulty. Once the test subject plateaus, then they are ready for a blind test.
In this manner, the experimenter can be sure what is being measured is not the test subject's comfort level with the apparatus and or the methodology.
A/B for sound identification to me should probably work as follows:
1) test subject may listen to A and or B as many times as desired.
2) when test subject feels ready, either A or B is randomly played.
3) test subject chooses that random sound is either A or B.
4) if no forced choice, then test subject can pass on answering: "I don't know."
Also, one of the most important things about perceptual testing is to train the test subject to reach a steady level of proficiency with the test apparatus using test sounds of graded difficulty. Once the test subject plateaus, then they are ready for a blind test.
In this manner, the experimenter can be sure what is being measured is not the test subject's comfort level with the apparatus and or the methodology.
Last edited:
On a first look the analysis of your A/B protocol would be identical to ABX except if (4) is allowed which makes it a censored sample that can be handled but is quite a bit more complex. For those interested, plot of the p(h) for 2, 3, 4 correct out of 4 tests is attached. Worth bearing in mind that the expected number of correct IDs from a monkey or random number generator without any ability to make considered choice would be 2, and this is reflected in the plot.
We did pre-checking and training of subjects and found that we had some listeners who were consistently more/less sensitive than others. In preference tests that we did with European partners they became part of the ANOVA analysis which removed their variation as a covariate. Preference tests are far more uncertain given realistic sample sizes - one reason that I preferred ABX-type tests which established if a difference could be heard.
We did pre-checking and training of subjects and found that we had some listeners who were consistently more/less sensitive than others. In preference tests that we did with European partners they became part of the ANOVA analysis which removed their variation as a covariate. Preference tests are far more uncertain given realistic sample sizes - one reason that I preferred ABX-type tests which established if a difference could be heard.
Attachments
Hi, Marcel,
Yes, some candidates on the suspect list certainly are more questionable than others, on their face. At this stage, however, I'm just hoping to compile the various suggested reasons, before beginning the process of elimination.
Half-band filters, for example, are certainly superior at image-suppression to having only the ZOH SINC envelope comb-filter effect. I included it because, at this point, I suggest that we first identify the technical differences between OS and NOS, not which is theoretically superior (OS, of course) nor which produces a better sound. Only, what might be producing the characteristically different NOS/OS sound which many of us feel we hear.
The possible-modulation of the baseband notion comes from a suggestion by Don Hill. I include it because I once looked presentation slides by Bruno Putzys, where he claims that NOS features a time-variant impulse-response. However, it was not clear to me exactly what he meant. My current understanding is that the baseband is intact before reconstruction, and that reconstruction only involves filtering the image-bands away.
@hop333,
Here is a problem I have with ABX: When I experimented with foobar ABX I found that when I was unsure, I tended to err by picking the wrong choice more often than not. In other words, my subconscious appeared to hear something but somehow the choice became inverted. The effect tended to reduce my score to less than what I could have done if not forced. Don't know if anyone else has noticed such a tendency.
Here is a problem I have with ABX: When I experimented with foobar ABX I found that when I was unsure, I tended to err by picking the wrong choice more often than not. In other words, my subconscious appeared to hear something but somehow the choice became inverted. The effect tended to reduce my score to less than what I could have done if not forced. Don't know if anyone else has noticed such a tendency.
@Lampie,
OS is theoretically superior to NOS simply because it can largely reconstruct the original signal. I say, largely, because it cannot completely reconstruct the signal. It mostly moves the image-bands up in frequency, but it cannot completely eliminate them. Being a discrete-time engine, an digital filter can only output discrete-time samples. To completely reconstruct the original signal into continuous-time requires a final analog filter. Even so, OS gets the signal closer to reconstruction than without one.
Of course, the discrete-time signal could, instead, be band-limited by an analog filter, however, the narrow 2KHz transition/guard band of the CD format makes that extremely difficult to do without resort to FIR filtering.
OS is theoretically superior to NOS simply because it can largely reconstruct the original signal. I say, largely, because it cannot completely reconstruct the signal. It mostly moves the image-bands up in frequency, but it cannot completely eliminate them. Being a discrete-time engine, an digital filter can only output discrete-time samples. To completely reconstruct the original signal into continuous-time requires a final analog filter. Even so, OS gets the signal closer to reconstruction than without one.
Of course, the discrete-time signal could, instead, be band-limited by an analog filter, however, the narrow 2KHz transition/guard band of the CD format makes that extremely difficult to do without resort to FIR filtering.