This was in response to my post #525. Sorry for the delayed response Mark, my internet has been down. Your response seems asserting that my arguments are unequivocally faulty, being based upon unequivocal faulty assumptions (by the use of the word “none”). Further that my type of unequivocal fallacy of argument has been historically dismissed of others having presenting it by authorities you appeal. Notwithstanding that your response is suggestive of fallacies of argument as well, it seems promoting that human hearing is impossible to correlate with visual spectrum data, thereupon equally asserting that visual spectrum data cannot be correlated of any scientific legitimacy to human hearing. This is considered true in absence of referencing spectrum data having arbitrary numbered scaling to physical SPL levels used in conducting listening tests.
A frequency spectrum is often displayed with an arbitrary scale beginning from 0dB down to some negative dB value. If this 0dB corresponds to the generation of a physical 160dB SPL, a subject would report auditory experiences differently if that same 0dB corresponded to -40 dB SPL. Given a variance of 200db SPL being imposed on these subjects it becomes difficult to correlate their auditory responses in comparison to just a visual spectrum displayed having an arbitrary numbered scale. Notwithstanding the difficulties in correlating human hearing to displayed spectrums as you suggest, referencing is essential in supporting conclusions based upon auditory responses to frequency spectrums, those usually presented implying auditory significance.
My intent was not to suggest referencing a physical SPL to the arbitrary 0dB point on a frequency spectrum, rather to reference 0dB SPL, defined at the limit of human hearing, to some point on the -dB scale of a visual spectrum presentation. For example if 0dB SPL corresponds to -80dB on a displayed spectrum, being of intensity corresponding to the sound of a mosquito in the room at 3 meters, this becomes the point below which spectrum data presented can be ignored as not possible to hear under those SPL levels of auditory testing.
Questions often arise as to where the boundary is of acceptable auditory noise, that for example in my case was recently done in testing an RIAA preamplifier having an input stage closely approx. the noise of an LT1028. This was fed from a turntable using a moving coil cartridge of 0.2mV. In conducting tests the listening level of variant records was turned up to the highest level of my personal expectation to generate. In lifting the stylus no noise was heard, representing a SNR value being acceptable regardless of value. This is to suggest that the magnitude of SNR has no necessary relevance in an auditory reality, though it would seem in my case the SNR would be around 75 to 80dB. As a beginning general rule the SNR does not need to exceed the level of SPL being produced, as 0dB is at the threshold of the human hearing of that noise.
Perhaps this (long!) interview with Grimm audio is interesting in the light of this discussion. They made a great streamer, with very long and precise filters and ultra low jitter. Basically they said that they wanted to achieve what is great about NOS, but with an improved oversampling solution...
Alpha Audio - Interview Grimm Audio | Facebook
Alpha Audio - Interview Grimm Audio | Facebook
A 0.2mV cart with a LT1028 can never exceed a S/N of:
20*log(0.2mV/(1nV/rtHz*sqrt(20Khz)) = 63dB.
This figure will be improved by the Riaa correction by 7dB max, giving an overall maximum achievable S/N of 70dB.
Hans
Attached, are the slides taken from a 2007 (wow, that long ago?) AES presentation by Bruno Putzeys, regarding digital filters and their issues in audio application. While the slides do not constitute a technical paper, and obviously are meant for Bruno to speak to in front of an audience, they touch on the DF issues which are in our investigation outline.
Hi Hierfi,
Sorry to hear your internet was down.
Maybe I can try to explain better what I was trying to get at: There is something called threshold of hearing. It is not a hard limit. It is an estimate of an average SPL for a particular frequency at which half of the population of people cannot hear the sound because it is too low SPL, but the other half of the population can still hear it.
Thresholds of hearing are measured with sine waves.
If human hearing were linear, time-invariant, and stationary, then Fourier would hold and we could think about the frequency domain and the time domain as two different ways of describing the exact same thing. Actually, we could use Fourier if hearing were at least time invariant and stationary. Unfortunately, hearing is not time invariant nor stationary.
It means we can't extrapolate from sine wave measurements how human hearing will work in all time domain cases. We have to measure hearing for each individual case.
In addition, it is already well known and uncontroversial that some people can hear if CDs are made without dither. 16-bit dither noise is down around -93dBFS, and truncation occurs at -96dB. Some people can hear that stuff at very modest playback SPL levels. They don't have to listen to CD playback at 100dB SPL to hear dither noise or truncation distortion. That's because Fourier doesn't hold, because of non-time invariance, and because of non-stationarity. We can't apply sine wave hearing thresholds as though humans were as time invariant and as stationary as an amplifier.
In addition, there is an AES paper (Frindle, 1997) that describes experiments and evidence showing that some people can hear truncation distortion even when it is hidden in dither noise. IIRC the dither noise in that case extended below -105dB.
IME it is also possible for some humans to hear harmonic distortion from a dac down around -120dBFS. Again, it is not necessary or helpful to play back music at 120dB SPL to hear it. In fact the opposite is true; too much SPL is deafening and interferes with hearing acuity.
How can that be true? IIUC it has never been studied to find out exactly why, but something very interesting has been studied and it may be suggestive of a type of mechanism to look for in human hearing: Stochastic resonance (sensory neurobiology - Wikipedia)
In the case of human perception of very low level harmonic distortion, I have said before that it could be that music signal raises the distortion SPL above some threshold SPL. Maybe it is somewhat like what happens in stochastic resonance, except for distortion perception of music it seems to me it might be more of a 'high-correlation resonance' type of effect, where the distortion is highly correlated with the music signal.
Last edited:
Noise floor of the room is usually much higher, but even so, people with good ears usually play music rather quietly in my experience. I'm pretty sure that human brain can separate meaningful information way below noise floor, which current measurement equipments can't. AI can? Maybe, if it can read CAPTCHA.
Last edited:
That’s not the case.
Just like an FFT, our ears/brains have a filter width or bin width.
When noise floor over 20Khz is A dB, and the bin width is B Hz,
Then we can look 10*log(20Khz/B) “ deeper into noise floor” A.
Hans
Digital Filter Experiments
Hans Polak has suggested an experiment to determine the subjective effect of a given OS digital FIR interpolation-filter. In addition, there is a second related experiment. These two experiments have the potential to be the most significant we perform.
You will need a computer to rip and playback a music track test file.
You will need a NOS capable DAC which can play music files from that computer.
You will need some sort of sampling rate conversion computer software, such as Audacity or Foobar, both of which utilize the SoX sample rate conversion utility. Another very intriguing sample conversion tool is the 'PGGB' offline resampling software. The PGGB has the potential to be the most accurate and transparent sample rate converter, which, as you'll recall, are essentially digital interpolation-filters. It features up to 8 billion taps, and looks to be the best example of a full-convolution Windowed-SINC interpolation-filter of which I'm aware. See my post #632 for more information on the PGGB. https://www.diyaudio.com/forums/digital-line-level/371931-makes-nos-sound-64.html#post6683260 SoX and PGGB both appear to utilize non-half-band Windowed-SINC interpolation-filters.
Hans' up/down resampling experiment should reveal whether the particular FIR interpolation-filter utilized in the rate conversion is audibly transparent. At least, that's our expectation. Here is the procedure, and possible subjective results.
Experiment 1:
1. Select a 44.1KHz music track for the experiment.
2. Rip an up-sampled copy of that track to your computer. 2FS (88.2KHz) is fine.
3. Then down-sample the copy back to 1FS (44.1KHz).
4. Compare the original 1FS track versus it's up/down sample copy, via your NOS DAC.
5. Listen only for whether they sound different, or the same.
Possible subjective results of experiment 1:
A) The sound of the up/down sampled file still sound like the original track when both are played back NOS. Meaning, they both sound like NOS characteristically does. This would indicate that the particular FIR filter utilized in the up/down sampling can be considered free of producing typical OS sound.
B) The sound of the up/down sampled file sound different than the original track when both are played back NOS. Meaning, the original track sounds like NOS characteristically does, however, up/down sampled file sounds like OS characteristically does. This would indicate that the particular FIR filter utilized in the up/down sampling is implicated as a subjective culprit.
==================================================
For the second experiment, you will listen for whether the up-sampled file sounds like typical OS CD-playback does to you, or whether it sounds like NOS playback, or whether it, perhaps, sounds better than either. In addition to the test set-up used for experiment 1, you will need an typical digital OS interpolation-filter based DAC or CD-player, which produces characteristic sounding OS playback to your ears.
Experiment 2:
1. Rip an up-sampled CD music track. The ratio is not critical, so long as your NOS DAC will accept the higher sample rate.
2. Listen to the sound of that up-sampled track played back via your NOS DAC.
3. Judge whether the up-sampled file played back via a NOS DAC sounds like, a) typical OS playback, or b) typical NOS playback, or c) something else. You only need identify the characteristic sound produced by OS playback and NOS playback. Your test result will either be: Sounds like OS, sounds like NOS, or sounds like something else.
Possible subjective results of experiment 2:
A) Up-sampled file playback via the NOS DAC sounds characteristically like typical CD playback via your OS DAC or CD-player. The reason may be that the up-sampling digital filter of the rate conversion software utilized may be essentially of the some type utilized in the OS digital interpolation-filter within your CD-player or DAC.
B) Up-sampled file playback via the NOS DAC still sounds characteristically like NOS does. This may indicate that you are getting the high-performance image-band suppression of a digital filter along with the subjective sound character of NOS.
C) Up-sampled file playback via the NOS DAC sounds different than OS or NOS typically do. Possibly, better than either. This result is what we hope to discover, how to obtain playback that is subjectively superior to either OS or NOS. Essentially, the best of the two combined.
If you try either experiment, please identify in your test report which sample rate conversion software you utilized. To determine that we hope to, we will need to identify the type of digital filter utilized (Meaning, Windowed-SINC, Equiripple, etc.) in your sample rate conversion software. In addition, different sample rate converter products feature interpolation-filters of different length. The PGGB software resampler, for example, features a user selectable filter length (taps). Making it possible to additionally experiment with the effect of interpolation-filter length. Longer filters produce a theoretically more accurate signal reconstruction.
Hans Polak has suggested an experiment to determine the subjective effect of a given OS digital FIR interpolation-filter. In addition, there is a second related experiment. These two experiments have the potential to be the most significant we perform.
You will need a computer to rip and playback a music track test file.
You will need a NOS capable DAC which can play music files from that computer.
You will need some sort of sampling rate conversion computer software, such as Audacity or Foobar, both of which utilize the SoX sample rate conversion utility. Another very intriguing sample conversion tool is the 'PGGB' offline resampling software. The PGGB has the potential to be the most accurate and transparent sample rate converter, which, as you'll recall, are essentially digital interpolation-filters. It features up to 8 billion taps, and looks to be the best example of a full-convolution Windowed-SINC interpolation-filter of which I'm aware. See my post #632 for more information on the PGGB. https://www.diyaudio.com/forums/digital-line-level/371931-makes-nos-sound-64.html#post6683260 SoX and PGGB both appear to utilize non-half-band Windowed-SINC interpolation-filters.
Hans' up/down resampling experiment should reveal whether the particular FIR interpolation-filter utilized in the rate conversion is audibly transparent. At least, that's our expectation. Here is the procedure, and possible subjective results.
Experiment 1:
1. Select a 44.1KHz music track for the experiment.
2. Rip an up-sampled copy of that track to your computer. 2FS (88.2KHz) is fine.
3. Then down-sample the copy back to 1FS (44.1KHz).
4. Compare the original 1FS track versus it's up/down sample copy, via your NOS DAC.
5. Listen only for whether they sound different, or the same.
Possible subjective results of experiment 1:
A) The sound of the up/down sampled file still sound like the original track when both are played back NOS. Meaning, they both sound like NOS characteristically does. This would indicate that the particular FIR filter utilized in the up/down sampling can be considered free of producing typical OS sound.
B) The sound of the up/down sampled file sound different than the original track when both are played back NOS. Meaning, the original track sounds like NOS characteristically does, however, up/down sampled file sounds like OS characteristically does. This would indicate that the particular FIR filter utilized in the up/down sampling is implicated as a subjective culprit.
==================================================
For the second experiment, you will listen for whether the up-sampled file sounds like typical OS CD-playback does to you, or whether it sounds like NOS playback, or whether it, perhaps, sounds better than either. In addition to the test set-up used for experiment 1, you will need an typical digital OS interpolation-filter based DAC or CD-player, which produces characteristic sounding OS playback to your ears.
Experiment 2:
1. Rip an up-sampled CD music track. The ratio is not critical, so long as your NOS DAC will accept the higher sample rate.
2. Listen to the sound of that up-sampled track played back via your NOS DAC.
3. Judge whether the up-sampled file played back via a NOS DAC sounds like, a) typical OS playback, or b) typical NOS playback, or c) something else. You only need identify the characteristic sound produced by OS playback and NOS playback. Your test result will either be: Sounds like OS, sounds like NOS, or sounds like something else.
Possible subjective results of experiment 2:
A) Up-sampled file playback via the NOS DAC sounds characteristically like typical CD playback via your OS DAC or CD-player. The reason may be that the up-sampling digital filter of the rate conversion software utilized may be essentially of the some type utilized in the OS digital interpolation-filter within your CD-player or DAC.
B) Up-sampled file playback via the NOS DAC still sounds characteristically like NOS does. This may indicate that you are getting the high-performance image-band suppression of a digital filter along with the subjective sound character of NOS.
C) Up-sampled file playback via the NOS DAC sounds different than OS or NOS typically do. Possibly, better than either. This result is what we hope to discover, how to obtain playback that is subjectively superior to either OS or NOS. Essentially, the best of the two combined.
If you try either experiment, please identify in your test report which sample rate conversion software you utilized. To determine that we hope to, we will need to identify the type of digital filter utilized (Meaning, Windowed-SINC, Equiripple, etc.) in your sample rate conversion software. In addition, different sample rate converter products feature interpolation-filters of different length. The PGGB software resampler, for example, features a user selectable filter length (taps). Making it possible to additionally experiment with the effect of interpolation-filter length. Longer filters produce a theoretically more accurate signal reconstruction.
Last edited:
Everyone, Marcel makes a valid point. It's technically possible that some resampling tools might not move the original sample values during up-sampling and down-sampling. Rendering your test result null. To absolutely avoid that possibility, up-sample to some rate which is not an integer multiple of 44.1KHz. So, any useable rate other than, 88.2KHz, 132.3KHz, 176.4KHz, etc.
Last edited:
At times it can be of questionable advantage to raise the SPL somewhat in order to diminish the residual background noise of a loved one... er...
It doesn't necessarily follow that hearing was caused by dithering as suggested, rather instead that the act of dithering a 0dBFS signal could have generated other artifacts being heard, being incorrectly attributed to dithering. To test this hypothesis is to perform alternative testing to reject this possibility.
Conventional testing seems was done by testing A vs. B, whereupon A was an un-dithered signal being sonically generated and compared to a sonically
generated B, the dithered version. An alternate test would be to isolate the dither itself by subtracted A from B as in (B-A), then maintain A as a separated source for both tests and sonically sum in the independently generated difference as (B-A). Hence B is generated as (A +(B-A)) that by logical reduction is still B.
This method rejects unknown artifacts stemming from changing a high level A to a high level B by using A in both tests, whereupon the difference signal (as the dither itself) is of low level magnitude suggested can be resolved.
High-correlation appears loosely along the lines suggested of temporal coherence found interesting at the following website.
Temporal coherence structure rapidly shapes neuronal interactions | Nature Communications
Even from the abstract it introduces some interesting thoughts. For example, in testing of digital source material that is undithered and thereupon more discontinuous, this could be limiting mechanisms of neuro alignment as affecting temporal coherency, as well as introducing neuro mechanisms in counter competition to that coherency.
In an analog world, if neuro alignment occurs on mass (in being focused on something) it can dig out sound below the noise floor, seemingly of depth reflected by the numbers of neutro aligned elements, much like using numbers of paralleled detectors.
In an analog world, if neuro alignment occurs on mass (in being focused on something) it can dig out sound below the noise floor, seemingly of depth reflected by the numbers of neutro aligned elements, much like using numbers of paralleled detectors.
With some proudness I can say that I have successfully modified one channel of my Dac from active into hybrid.
First impression is that it seems to be working properly.
After having performed some tests with positive outcome, I will also modify the Left channel.
Hans
.
First impression is that it seems to be working properly.
After having performed some tests with positive outcome, I will also modify the Left channel.
Hans
.
Attachments
Last edited: