Yea, I am also confused by that..
I can at least tell you that I could not hear the difference any more, and that I had to guess...
It seems that should one of the files produce a subjectively more dynamic sound, it would lead you to choose that file in a loudness ABX test, even though the pair now measure the same. It would be interesting to submit to Hans for scoring, which 'God Give Me Strength' file you chose in your ABX test. This would determine whether the orignal, or the resampled version was the one subjectively louder to you. I'm not certain of the significance of whichever the result, but identifying a correlation must indicate something relevant.
By the way, is that 0.35dB loudness difference a peak measure, or an average?
Last edited:
Was that test for detection of RIAA variances? I remember that Lipshitz did extensive testing on the RIAA, hence the variances could be related to phase shifting at the corner frequencies.
In my first test I compared file 7 with 24 with conclusion that 7 is louder than 24
The second test I boosted file 24 by 0.35 dB and compared that with the earlier file 7. Conclusion was that I could not hear any difference between the files
The 0.35dB loudness difference is an everage over the whole file.
This value is also the optimum value for minimising the difference (file 7 minus file 24) between the files. So I am quite confident there is global scaling in one of the files by plus or minus 0.35dB depending if file 7 or 24 was the original. @Hans: what tools was used for the processing? Maybe we can find out what causes the level difference?
Ok results for all files:
7 God Give Me Strength.wav -16.45 dB
24 God Give Me Strength.wav -16.80 dB
Difference: -0.35 dB
8 Bach_ Pastorale In C Minor.wav -22.04 dB
13 Bach_ Pastorale In C Minor.wav -22.22 dB
Difference: -0.18 dB
10 Day 0.wav -23.60 dB
14 Day 0.wav -23.60 dB
Difference: 0 dB
16 Still Got The Blues.wav -17.64 dB
22 Still Got The Blues.wav -17.64 dB
Difference: 0 dB
So indeed, it looks like the other files at least are so close, the level wont have an influence on the tests.
Still would like to know why the level changes are there in the first place
Last edited:
In view of this I retested both of these files from an auditory perspective. After each test I turned the volume completely down, started up the files and raised the amplitude. The character differences were consistent over a large volume range, hence doesn't seem dependent of differences, particularly for such small differences. Interestingly the differences seemed more obvious with this file as compared to some others.
The PGGB resampler, indeed, has an inter-sample-over gain scaling correction feature, which is defeatable. I had forgotten about that feature, and it seems likely that this is where the small level difference is coming. I suspect that the 'God Give Me Strength' track is one of the two problem files which ZB mentions in the below excerpt from one of his emails about our resampled files.
________________________________________________________________
...During upsampling to 88.2kHz, you can have intersample overs, i.e., signals can exceed full scale (in fact two of your test tracks do this), so the signals need to be scaled to avoid clipping, this too will ensure the conversion will not result in the original 44.1kHz signal even if the filters used are perfect. This can become audible if the level difference is significant...
-ZB
Ok, did this test: The peaks:
(inter samples when upsampling to 8fs with audacity)
7 God Give Me Strength -0.35 dBFS
(edit: I now also checked the other 4... no other > 0.0 dBFS samples found)
8 Bach_ Pastorale In C Minor: +0.192 dBFs
10 Day 0 -0.635 dBFS
16 Still Got The Blues -0.29 dBFS
So only file 8 has an issue like this
Last edited:
Could be, I use the 'best quality' resampling of audacity.
44.1 to 352.8
Have no idea what the other one is...
While people are hopefully still listening to the 4+4 test, I have some info's on I-V conversion, after Ken appointed me to "chief I-V"
I-V conversion may be an important ingredient in producing a sonic profile after all.
My research concentrated around a PCM179X Dac, but several results may be valid also for other Dac's. This very Dac produces +/-4mA Fs output current on a 6mA bias.
I have investigated 4 versions:
1) A passive conversion with a 25R resistor, producing the most faithful I-V conversion, but giving a S/N after active amplifying that is quite a bit less than other active variants.
2) A fast voltage feedback amp, a LT1468 with a 1k//1.5nF first order LP filter, thereby presenting a virt. gnd to the Dac's current output.
3) A second order active MFB topology, presenting a 10nF cap to the Dac's output, meant to absorb eventual current transients and allowing for a lower specified amp hereafter.
4) A fast current conveyor that is shifting the current 1:1 to the conveyors output where a simple resistor can do the I-V conversion, eventually followed by a buffer.
Contrary to the other active solutions, this type of converter has no feedback.
A) First findings using the 25R resistor as I-V converter.
- The PCM1792 in my set, was installed to convert all input to 192/24.
However, image pairs do not appear at multiples of 192Khz, but at 8 times as high at multiples of 1.536Mhz, making the demands for an analogue reconstruction filter very relaxed.
- The Dac's analogue output update happens at multiples of 81.38nsec or 1/(64*192Khz).
-Starting each new 81.38nsec timeslot, independent whether current changed or not, a 1mA 2nsec transient is produced followed by a number of smaller ones.
It is obvious that this poses serious demands on an active I-V converter not to become overloaded resulting in unwanted distortion.
B) The active second order MFB topology
Loaded with the information under A), I started my search with the MFB topology, hoping that the 10nF cap at the Dac's output would attenuate the 2nsec transients significantly and that this version could become the first choice for I-V converters, although never before published to my knowledge, as we will see with a reason.
However, the Dac responded in an unexpected way.
HF noise increased considerably, against all expectations and sound was noticeably worse.
More bass, but mid end was a far cry from what it was before.
All recorded spectra showed significant differences, so this variant was rejected.
Long story short, the Dac obviously doesn't like getting loaded with a capacitor.
C) The current conveyor or CC.
This topology can have a very high BW with ordinary transistors, so input overload will never be a problem, and when choosing for the right circuit diagram, input impedance can be low over a large BW.
With 2nsec transients, input impedance rises to above 20R, but starts being constant over a large frequency band somewhere in the low Ohm region.
Simulations may show ultra-low distortion versions, but the limiting factor will be in most cases the Dac distortion itself.
So opting for a ppm distortion version with a more complex circuitry, makes hardly sense.
Depending on various choices to be made, this may be a good solution.
D) The first order voltage feedback amp.
Most important issue for this version was to find out whether a fast op-amp could handle said current transients of 1mA@2nsec.
It turned out that the voltage at the input did not exceed a measured ca. +/-30mV, a good deal below the +/-50mV where the op-amp could become oversteered.
Zin for these transients is thus around 30R when using the LT1468 op-amp.
At audio frequencies Zin is the lowest of all other variants in the very low mOhm range ramping up with 20dB/decade, to ca. 1R at 100K.
For those seeing the constant Zin up to 100Khz as a plus for the CC, a 0.5R resistor in series with the Dac output will cause the same constant Zin up to 100Khz.
So a fast op-amp, slew rate >=25v/usec and around 100Mhz BW, is just as well able to handle the Dac’s transients as a CC, with the advantage that they take much less space.
The only difference between the two is the 100% global feedback versus no global feedback.
E) CMRR and PSRR.
For the op-amp variant PSRR will be most likely good enough, but for the CC it largely depends on the used circuit diagram. A simple version may have a rather bad PSRR.
But CMRR, defined as 20*log(Adiff/|Acm|) is a reason for concern for both the op-amp and for the CC solution.
The Dac and the I-V converter have to do their jobs in a hostile environment with many digital signals around so CMRR is important.
A 1% mismatch between the two resistors converting the Dac’s Iout+ and Iout- , already results in -40dB CMRR, whereas 0.1% gives below -60dB.
The same is true for the amp behind the two I-V converters, responsible for turning the two opposite signals into one diff or SE signal.
Resistors around this amp should also be within 0.1%.
The 1.5nF capacitor // to the 1K resistor, giving the first order LP function is also critical and should match on both sides within 1%.
F) NOS Dac with the PCM179X.
Since this Dac is a very complicated machine, I still have to investigate what happens when switching off the FIR up-sample filter by connecting two pins.
Hans
I-V conversion may be an important ingredient in producing a sonic profile after all.
My research concentrated around a PCM179X Dac, but several results may be valid also for other Dac's. This very Dac produces +/-4mA Fs output current on a 6mA bias.
I have investigated 4 versions:
1) A passive conversion with a 25R resistor, producing the most faithful I-V conversion, but giving a S/N after active amplifying that is quite a bit less than other active variants.
2) A fast voltage feedback amp, a LT1468 with a 1k//1.5nF first order LP filter, thereby presenting a virt. gnd to the Dac's current output.
3) A second order active MFB topology, presenting a 10nF cap to the Dac's output, meant to absorb eventual current transients and allowing for a lower specified amp hereafter.
4) A fast current conveyor that is shifting the current 1:1 to the conveyors output where a simple resistor can do the I-V conversion, eventually followed by a buffer.
Contrary to the other active solutions, this type of converter has no feedback.
A) First findings using the 25R resistor as I-V converter.
- The PCM1792 in my set, was installed to convert all input to 192/24.
However, image pairs do not appear at multiples of 192Khz, but at 8 times as high at multiples of 1.536Mhz, making the demands for an analogue reconstruction filter very relaxed.
- The Dac's analogue output update happens at multiples of 81.38nsec or 1/(64*192Khz).
-Starting each new 81.38nsec timeslot, independent whether current changed or not, a 1mA 2nsec transient is produced followed by a number of smaller ones.
It is obvious that this poses serious demands on an active I-V converter not to become overloaded resulting in unwanted distortion.
B) The active second order MFB topology
Loaded with the information under A), I started my search with the MFB topology, hoping that the 10nF cap at the Dac's output would attenuate the 2nsec transients significantly and that this version could become the first choice for I-V converters, although never before published to my knowledge, as we will see with a reason.
However, the Dac responded in an unexpected way.
HF noise increased considerably, against all expectations and sound was noticeably worse.
More bass, but mid end was a far cry from what it was before.
All recorded spectra showed significant differences, so this variant was rejected.
Long story short, the Dac obviously doesn't like getting loaded with a capacitor.
C) The current conveyor or CC.
This topology can have a very high BW with ordinary transistors, so input overload will never be a problem, and when choosing for the right circuit diagram, input impedance can be low over a large BW.
With 2nsec transients, input impedance rises to above 20R, but starts being constant over a large frequency band somewhere in the low Ohm region.
Simulations may show ultra-low distortion versions, but the limiting factor will be in most cases the Dac distortion itself.
So opting for a ppm distortion version with a more complex circuitry, makes hardly sense.
Depending on various choices to be made, this may be a good solution.
D) The first order voltage feedback amp.
Most important issue for this version was to find out whether a fast op-amp could handle said current transients of 1mA@2nsec.
It turned out that the voltage at the input did not exceed a measured ca. +/-30mV, a good deal below the +/-50mV where the op-amp could become oversteered.
Zin for these transients is thus around 30R when using the LT1468 op-amp.
At audio frequencies Zin is the lowest of all other variants in the very low mOhm range ramping up with 20dB/decade, to ca. 1R at 100K.
For those seeing the constant Zin up to 100Khz as a plus for the CC, a 0.5R resistor in series with the Dac output will cause the same constant Zin up to 100Khz.
So a fast op-amp, slew rate >=25v/usec and around 100Mhz BW, is just as well able to handle the Dac’s transients as a CC, with the advantage that they take much less space.
The only difference between the two is the 100% global feedback versus no global feedback.
E) CMRR and PSRR.
For the op-amp variant PSRR will be most likely good enough, but for the CC it largely depends on the used circuit diagram. A simple version may have a rather bad PSRR.
But CMRR, defined as 20*log(Adiff/|Acm|) is a reason for concern for both the op-amp and for the CC solution.
The Dac and the I-V converter have to do their jobs in a hostile environment with many digital signals around so CMRR is important.
A 1% mismatch between the two resistors converting the Dac’s Iout+ and Iout- , already results in -40dB CMRR, whereas 0.1% gives below -60dB.
The same is true for the amp behind the two I-V converters, responsible for turning the two opposite signals into one diff or SE signal.
Resistors around this amp should also be within 0.1%.
The 1.5nF capacitor // to the 1K resistor, giving the first order LP function is also critical and should match on both sides within 1%.
F) NOS Dac with the PCM179X.
Since this Dac is a very complicated machine, I still have to investigate what happens when switching off the FIR up-sample filter by connecting two pins.
Hans