HydrogenAudio

Hydrogenaudio Forum => Scientific Discussion => Topic started by: Arnold B. Krueger on 2015-04-20 13:10:09

Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-20 13:10:09
In the MQA hype I've found the following claim:

MQA Hype (https://mrapodizer.wordpress.com/2015/01/13/mqa-what-is-meridian-hiding/)

"However from more recent research it turns out that although humans can not hear frequencies above 20kHz, they are sensitive to timing of sounds to about 10 microseconds. So first you notice the arrival of a sound (quick change in air pressure, a very high frequency) and later on you actual hear what sound it is. To preserve this timing info in the audio signal 96kHz is therefor not enough, we actually need 192kHz."

Both the claim that "(humans) are sensitive to timing of sounds to about 10 microseconds" and

"To preserve this timing info in the audio signal 96kHz is therefor not enough, we actually need 192kHz"

seem to be completely wrong, based on what I know about human hearing and digital audio.

What are they talking about?
Title: Audibility of phase shifts and time delays
Post by: Wombat on 2015-04-20 14:36:41
What are they talking about?

"So first you notice the arrival of a sound (quick change in air pressure, a very high frequency) and later on you actual hear what sound it is."
This must be their way of describing the terrific PRE-ringing!
Title: Audibility of phase shifts and time delays
Post by: xnor on 2015-04-20 15:11:45
(http://xserv.compress.to/xnor/audio/images/cd10us.png)

x-axis is in us (microseconds). Notice how the dots are spaced at 44.1 kHz or roughly 23 us apart.
y-axis is linear in 0.1 divisions, with the max shown being ~0.6 or about -4.4 dBFS.

We take some impulsive signal, quantize it and visualize it as red dots. The corresponding blue line shows the upsampled version of the signal before quantization (the 'ideal' or target).
We then delay this signal by 10us, quantize it and visualize it as green dots. The blue line is again the upsampled, this time delayed signal before quantization.

The dots match our target... Even if we attenuate this signal by 60 dB first, we will see the effects of quantization and dither, yes, but the 10us delay does not disappear.
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-20 16:01:58
(http://xserv.compress.to/xnor/audio/images/cd10us.png)

x-axis is in us (microseconds). Notice how the dots are spaced at 44.1 kHz or roughly 23 us apart.
y-axis is linear in 0.1 divisions, with the max shown being ~0.6 or about -4.4 dBFS.

We take some impulsive signal, quantize it and visualize it as red dots. The corresponding blue line shows the upsampled version of the signal before quantization (the 'ideal' or target).
We then delay this signal by 10us, quantize it and visualize it as green dots. The blue line is again the upsampled, this time delayed signal before quantization.

The dots match our target... Even if we attenuate this signal by 60 dB first, we will see the effects of quantization and dither, yes, but the 10us delay does not disappear.



In short, 10 microseconds is nowhere near the actual abilities of 4416 to convey timing differences among
signals is:  1/(44100*65536) =

3.460042871315193e-10  seconds. Less than a nanosecond.
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-20 16:05:20
The information that I am aware of related to the audibility of time differences is summarized here:

David L. Clarks AES paper about the audibility of time (http://ethanwiner.com/phase.html)

"
Despite the possibility of "digital effects," the modest quality of the speakers, and the fact that the listeners were in effect being tested, every one of approximately 12 participants was able to match the time delays to within +/- 40 microseconds (about 1/2 inch). This finding confirms the clams of those who espouse the virtues of arrival time compensated loudspeakers.
"
Title: Audibility of phase shifts and time delays
Post by: saratoga on 2015-04-20 16:37:06
Cool, words about audio from someone who doesn't understand sampling.
Title: Audibility of phase shifts and time delays
Post by: pdq on 2015-04-20 16:46:55
In the MQA hype I've found the following claim:

MQA Hype (https://mrapodizer.wordpress.com/2015/01/13/mqa-what-is-meridian-hiding/)

"However from more recent research it turns out that although humans can not hear frequencies above 20kHz, they are sensitive to timing of sounds to about 10 microseconds. So first you notice the arrival of a sound (quick change in air pressure, a very high frequency) and later on you actual hear what sound it is. To preserve this timing info in the audio signal 96kHz is therefor not enough, we actually need 192kHz."

The link is now broken.
Title: Audibility of phase shifts and time delays
Post by: xnor on 2015-04-20 16:53:27
signals is:  1/(44100*65536) =

3.460042871315193e-10  seconds. Less than a nanosecond.


Well using that formula we'd tread into the >1us 'resolution' range 70 dB down*. 90 dB down would be 10us.


*) This does not translate directly into what you see in a spectrogram. A 1ms long tone burst that reaches full scale would show up below -50 to -60 dB with a 65k FFT size in Audition.
Title: Audibility of phase shifts and time delays
Post by: xnor on 2015-04-20 17:06:49
Arny, they are not talking about phase in an absolute sense, but the phase relationship of the content within a track.
Title: Audibility of phase shifts and time delays
Post by: pdq on 2015-04-20 17:23:36
In the MQA hype I've found the following claim:

MQA Hype (https://mrapodizer.wordpress.com/2015/01/13/mqa-what-is-meridian-hiding/)

"However from more recent research it turns out that although humans can not hear frequencies above 20kHz, they are sensitive to timing of sounds to about 10 microseconds. So first you notice the arrival of a sound (quick change in air pressure, a very high frequency) and later on you actual hear what sound it is. To preserve this timing info in the audio signal 96kHz is therefor not enough, we actually need 192kHz."

The link is now broken.

Working again...
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-20 21:17:45
Arny, they are not talking about phase in an absolute sense, but the phase relationship of the content within a track.


Exactly.  Re: Clark's article David L Clark's article about the audibility of timing and phase (http://ethanwiner.com/phase.html)
Title: Audibility of phase shifts and time delays
Post by: xnor on 2015-04-20 22:27:28
I didn't mean allpass filters either. That would be a constant phase shift at a given frequency.

I guess what they're talking about is some sort of quantization distortion along the x-axis. I doesn't make sense, but it doesn't need to for FUD.
Title: Audibility of phase shifts and time delays
Post by: Woodinville on 2015-04-21 03:52:07
(http://xserv.compress.to/xnor/audio/images/cd10us.png)

x-axis is in us (microseconds). Notice how the dots are spaced at 44.1 kHz or roughly 23 us apart.
y-axis is linear in 0.1 divisions, with the max shown being ~0.6 or about -4.4 dBFS.

We take some impulsive signal, quantize it and visualize it as red dots. The corresponding blue line shows the upsampled version of the signal before quantization (the 'ideal' or target).
We then delay this signal by 10us, quantize it and visualize it as green dots. The blue line is again the upsampled, this time delayed signal before quantization.

The dots match our target... Even if we attenuate this signal by 60 dB first, we will see the effects of quantization and dither, yes, but the 10us delay does not disappear.



In short, 10 microseconds is nowhere near the actual abilities of 4416 to convey timing differences among
signals is:  1/(44100*65536) =

3.460042871315193e-10  seconds. Less than a nanosecond.


Well, 1/ ( 2 pi bandwidth number_of_levels)   

This hardly disproves your point, naturally!
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-21 10:50:30
(http://xserv.compress.to/xnor/audio/images/cd10us.png)

x-axis is in us (microseconds). Notice how the dots are spaced at 44.1 kHz or roughly 23 us apart.
y-axis is linear in 0.1 divisions, with the max shown being ~0.6 or about -4.4 dBFS.

We take some impulsive signal, quantize it and visualize it as red dots. The corresponding blue line shows the upsampled version of the signal before quantization (the 'ideal' or target).
We then delay this signal by 10us, quantize it and visualize it as green dots. The blue line is again the upsampled, this time delayed signal before quantization.

The dots match our target... Even if we attenuate this signal by 60 dB first, we will see the effects of quantization and dither, yes, but the 10us delay does not disappear.



In short, 10 microseconds is nowhere near the actual abilities of 4416 to convey timing differences among
signals is:  1/(44100*65536) =

3.460042871315193e-10  seconds. Less than a nanosecond.


Well, 1/ ( 2 pi bandwidth number_of_levels)   

This hardly disproves your point, naturally!



Thanks for the correction, JJ. Your correction seems to make the resolution even  more than my calculation said or 1.0990333934666109990166543854704e-10  seconds - about 0.1  nanoeconds.  No perceiving that with the ears, or its effects on frequency response.
Title: Audibility of phase shifts and time delays
Post by: xnor on 2015-04-21 11:52:53
So in order for that to degrade to 1us we're down ~80 dB.
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-21 12:57:06
So in order for that to degrade to 1us we're down ~80 dB.



The way I see it, the perceptual limit is about 40 uSec or 40 x 10E-6 seconds, and the resolution limit is about 1.1 x 10E-10.  so, the resolution limit is 40 x 10-4 below the perceptual limit which is 92 dB down.


When we are talking about timing, there are a lot of ways to quantify it.

For example Clark is talking about trying to synchronize two identical sounds that are operating concurrently. Cancellations which lead to steady-state frequency response differences are a big part of this.

When we are talking about hearing the difference between two different musical sounds, then there are no steady state frequency response differences to listen for, and this is more like me ABXing two files that are not mixed but are displaced in time about a week ago. Its like hearing an echo which puts the perceptual limit up into the range of milliseconds, not microseconds.

Thus we find two false claims that are being used to sell MQA.

One is that the ears are at from 4 to over a thousand times more sensitive than they are to timing differences, and that 4416 is at least 10.000 times worse than it is at transporting timing information.
Title: Audibility of phase shifts and time delays
Post by: 2Bdecided on 2015-04-21 16:54:54
"However from more recent research it turns out that although humans can not hear frequencies above 20kHz, they are sensitive to timing of sounds to about 10 microseconds."
"recent"?!

Klumpp, R. G.; and Eady, H. R. (1956).
Some Measurements of Interaural Time difference Thresholds.
Journal of the Acoustical Society of America, vol. 28, no. 5, Sept., pp. 859-860.

It's the earliest reference I know to 11 us.

Some of it is reproduced on pages 4+5 of this...
http://web.mit.edu/hst.723/www/ThemePapers.../Grantham95.pdf (http://web.mit.edu/hst.723/www/ThemePapers/Binaural/Grantham95.pdf)

Cheers,
David.
Title: Audibility of phase shifts and time delays
Post by: krabapple on 2015-04-21 17:32:54
Are we entering Kunchur-land again here?

Title: Audibility of phase shifts and time delays
Post by: xnor on 2015-04-21 18:28:55
The way I see it, the perceptual limit is about 40 uSec or 40 x 10E-6 seconds, and the resolution limit is about 1.1 x 10E-10.  so, the resolution limit is 40 x 10-4 below the perceptual limit which is 92 dB down.


That's not what I meant. 2^16 values are only available to a tone that reaches full-scale. 80 dB (roughly 13 bits) down we're left with 2^16 * 10^(-80/20) values.
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-21 18:52:28
"However from more recent research it turns out that although humans can not hear frequencies above 20kHz, they are sensitive to timing of sounds to about 10 microseconds."
"recent"?!

Klumpp, R. G.; and Eady, H. R. (1956).
Some Measurements of Interaural Time difference Thresholds.
Journal of the Acoustical Society of America, vol. 28, no. 5, Sept., pp. 859-860.

It's the earliest reference I know to 11 us.

Some of it is reproduced on pages 4+5 of this...
http://web.mit.edu/hst.723/www/ThemePapers.../Grantham95.pdf (http://web.mit.edu/hst.723/www/ThemePapers/Binaural/Grantham95.pdf)



Thanks. Clark's experiment was based on loudspeaker listening, while it appears that Grantham95 is based on headphones. That one would be 4 times the other seems to make sense.

The big mistake in the MQA hype is then that somehow 44/16 can't hack reproducing time delays on the order of a few microseconds. It can hack it, and its resolution goes many orders of magnitude below that.
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-21 19:05:36
Are we entering Kunchur-land again here?



Read here: Kunchur 2007 Temporal Resolution Paper (http://boson.physics.sc.edu/~kunchur/papers/Temporal-resolution-by-bandwidth-restriction--Kunchur.pdf)

His Faux pas about serious problems with the temporal resolution possible with 4416 seem to have been avoided, but then he goes down the Stuart road.

However in the same year, same journal he published this:

Other Kunchur time resolution paper (http://boson.physics.sc.edu/~kunchur/papers/Audibility-of-time-misalignment-of-acoustic-signals---Kunchur.pdf)

"
Every component’s bandwidth limit
(even if it behaves perfectly linearly) causes it to have a
finite relaxation time of ??1/?max; use of digital carriers
limits the shortest resolvable time interval to about half
the sampling interval (which for CD would be 11 ?s);
"
Title: Audibility of phase shifts and time delays
Post by: krabapple on 2015-04-21 19:10:48
Are we entering Kunchur-land again here?



Read here: Kunchur 2007 Temporal Resolution Paper (http://boson.physics.sc.edu/~kunchur/papers/Temporal-resolution-by-bandwidth-restriction--Kunchur.pdf)



I don't need to, again.

What I'm saying is, we have been here before.
Title: Audibility of phase shifts and time delays
Post by: augustine on 2015-04-22 11:25:38
Even if his was audible- I have serious doubts that loudspeakers of the same make an model would be  able to reproduce this timing accurately anyway - people would have to calibrate their speahkers within microseconds so they matched - marketing madness
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-22 15:30:58
The way I see it, the perceptual limit is about 40 uSec or 40 x 10E-6 seconds, and the resolution limit is about 1.1 x 10E-10.  so, the resolution limit is 40 x 10-4 below the perceptual limit which is 92 dB down.


That's not what I meant. 2^16 values are only available to a tone that reaches full-scale. 80 dB (roughly 13 bits) down we're left with 2^16 * 10^(-80/20) values.


Are you suggesting that all future resolution specifications be based on FS = -80 dB?

At that rate 4416 only has 16 dB dynamic range.
Title: Audibility of phase shifts and time delays
Post by: xnor on 2015-04-22 16:12:33
No, but it is good that you bring up SNR because this 'resolution' is proportional to SNR. That's all. I was just curious where this formula would spit out 1us/5us.
Title: Audibility of phase shifts and time delays
Post by: xnor on 2015-04-22 16:19:11
Haven't we also not seen this Yamaha page before: temporal_resolution (http://www.yamahaproaudio.com/global/en/training_support/selftraining/audio_quality/chapter5/09_temporal_resolution/)?

They say not 10us but 6us.
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-23 13:31:57
Haven't we also not seen this Yamaha page before: temporal_resolution (http://www.yamahaproaudio.com/global/en/training_support/selftraining/audio_quality/chapter5/09_temporal_resolution/)?

They say not 10us but 6us.



I wouldn't haggle over 10 uSec as opposed to 5 uSec as being the threshold. However, its impact seems to have been considerably inflated:

"To also accurately reproduce changes in a signal’s frequency spectrum with a temporal resolution down to 6 microseconds, the sampling rate of a digital audio system must operate at a minimum of the reciprocal of 6 microseconds = 166 kHz. Figure 515 presents the sampling of an audio signal that starts at t = 0, and reaches a detectable level at t = 6 microseconds. To capture the onset of the waveform, the sample time must be at least 6 microseconds."

So they are basically saying that 44 KHz can't accurately reproduce audible changes in a signal, and 96 Khz can't accurately reproduce audibly changes in a signal, you have to go to almost twice that, or 166 KHz!  They are basically on the Meridian bandwagon of 192 KHz being the historically-ordained sample rate that is required for sonically accurate reproduction.

Going back in history, Clark's paper that put audibility @ 40 uSec alluded to the fact that he didn't think it was the time delays as such that caused the audible differences, but the changes in frequency response that they necessarily caused. He said this to me in so many words at the time. Later on I did some ABX tests that were positive for audible differences due to the 1/2 sample inter-channel delay in the converters in the CDP 101 combined with a system that had a center channel speaker that received L+R out of an ordinary analog summer.  Same basic explanation.

I think that we are getting treated to the typical placebophile moving goal posts fueled by conflating experimental results with actual operational circumstances that are different.

Title: Audibility of phase shifts and time delays
Post by: 2Bdecided on 2015-04-23 14:33:31
Later on I did some ABX tests that were positive for audible differences due to the 1/2 sample inter-channel delay in the converters in the CDP 101 combined with a system that had a center channel speaker that received L+R out of an ordinary analog summer.  Same basic explanation.
I make that 1dB down at 13.1kHz, 2dB down at 18.4kHz, etc. I'd struggle to detect that these days. With a whole sample delay, you can halve those frequencies (making it easily audible), and get a full 6dB attenuation by 14.7kHz.

But as everyone says, in a proper 44.1kHz 16-bit system there are no such problems.

Cheers,
David.
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-23 14:56:00
Later on I did some ABX tests that were positive for audible differences due to the 1/2 sample inter-channel delay in the converters in the CDP 101 combined with a system that had a center channel speaker that received L+R out of an ordinary analog summer.  Same basic explanation.


I make that 1dB down at 13.1kHz, 2dB down at 18.4kHz, etc. I'd struggle to detect that these days.


I would too. However this was in the late 1980s. The kid was on his game!

Quote
With a whole sample delay, you can halve those frequencies (making it easily audible), and get a full 6dB attenuation by 14.7kHz.


Flip the argument around.  1 sample is 22 uSec,  1/2 sample is 11 uSec, and  1/4 sample is 5.5 uSec.  And there we are in the 5 uSec range.

I don't know if I could have heard the 1/4 sample delay back in the day. I never tried, and I didn't  know how to try as hard then as I learned later.

This summation of delayed and non-delayed signals can happen in wire or it can happen, perhaps with less coherence in the air.

Quote
But as everyone says, in a proper 44.1kHz 16-bit system there are no such problems.


Yes, we have to say that for the placebophiles in attendance. ;-)

Fact of the matter is that even with the 1/2 sample interchannel delay uncorrected, the CDP 101 was challenging to ABX with positive outcome for differences with music and stereo speakers.  Those who ranted and raved about its horrid sound quality were always placebophiles, well-primed by articles from the Golden Eared press.  That hasn't changed.

The point is that the potential audible effects related to 5 uSec delays can be completely explained by their frequency response effects. Then we don't have to throw away what the classic books about psychoacoustics say about the audibility of phase and delay.
Title: Audibility of phase shifts and time delays
Post by: mzil on 2015-04-23 19:27:41
I agree "temporal resolution/smearing" is mumbo jumbo and Bob Stuart woo. What's possibly audible is simply the frequency response alteration due to constructive/destructive interference of the two waveforms. I notice exactly that here in a minor change to the sound between the .5 millisecond spaced clicks and the 1 millisecond spaced clicks:

https://www.youtube.com/watch?v=3zZRy-UArXM (https://www.youtube.com/watch?v=3zZRy-UArXM)

Granted this Youtube test is in milliseconds, not microseconds, but the principal still holds true: the audible detection at this 500 microseconds difference level is due to frequency response, not "timing". I'm new to Audacity, so it would take me forever to construct, but if one of you Audacity veterans would like to make a click track with much shorter than 500 microsecond spacings, those of us with [slightly] younger ears would be more than glad to take the test.
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-23 23:31:42
I agree "temporal resolution/smearing" is mumbo jumbo and Bob Stuart woo. What's possibly audible is simply the frequency response alteration due to constructive/destructive interference of the two waveforms. I notice exactly that here in a minor change to the sound between the .5 millisecond spaced clicks and the 1 millisecond spaced clicks:

https://www.youtube.com/watch?v=3zZRy-UArXM (https://www.youtube.com/watch?v=3zZRy-UArXM)

Granted this Youtube test is in milliseconds, not microseconds, but the principal still holds true: the audible detection at this 500 microseconds difference level is due to frequency response, not "timing". I'm new to Audacity, so it would take me forever to construct, but if one of you Audacity veterans would like to make a click track with much shorter than 500 microsecond spacings, those of us with [slightly] younger ears would be more than glad to take the test.


Here's what Kunchur says about past work trying to determine this experimentally:\

"
In experiments probing temporal resolution, a pair of
stimuli are presented that differ in their temporal structure.
As the temporal difference is progressively reduced,
one finds the threshold for barely being able to discern
a difference. In one experiment by Leshowitz (1971), listeners
were presented with a single pulse or two narrower
pulses (with the same total energy) separated by an interval
?t. The click and click-pair could be distinguished
down to ?t ? 10 ?s.

In this case, the two stimuli have differences
in their amplitude spectra and their discernment
was explained on this basis. Isospectral variants of this experiment
were carried out by Ronken (1970) and later by
Henning and Gaskell (1981) where one stimulus consisted
of a short pulse followed by a taller one separated by an
interval ?t. The second stimulus was a similar pair with
the time order reversed and hence had the same amplitude
spectrum. The shortest ?t for which these stimuli could be
distinguished was about 200 ?s.

Another type of constantamplitude-
spectrum experiment involves the detection of
gaps in noise (Plomp, 1964; Penner, 1977; Eddins et al.,
1992). In these the threshold for gap detection was of the
order of 2 ms. The issue of determining temporal resolution
while avoiding spectral cues was recently tackled (Yost
et al., 1996; Patterson and Datta, 1996; Krumbholz et al.,
2003) through the use of iterated rippled noise.

The experiment of Krumbholz et al. (2003) showed that differences in
delay between a masker and signal could be discerned down
to 12.5 ?s. In their work a masking paradigm was used to
argue that spectral cues did not play a role in the discernment.
Note that in all the previous cited experiments, the
threshold ?t exceeded the nominal 9 ?s.
"

His methodology was as follows:

"The experiment consists of presenting an approximately
square-wave shaped complex tone, with a 7 kHz fundamental,
through earphones with different degrees of low-pass
filtering (i.e., with different time constants ? ) and testing
a listener’s ability to distinguish this filtered tone from the
unfiltered control tone (?=0).
A. Apparatus
A significant potential bottleneck in a temporal resolution
experiment is the temporal-response speed of
the equipment. Typically the apparatus consists of signal
sources, a switching/gating method used to ramp the signals,
an amplifier for driving the tranducer, and the transducer
itself. In the present work, many different approaches
were initally tried and abandoned, including using digital
synthesis (with 24-bit/96-kHz sampling) for the production
and ramping of signals. It was found that such a digital
method had far too inadequate temporal definition for this
purpose. So instead an analog signal generator (model 4001
manufactured by Global Specialties Instruments, Cheshire,
Connecticutt) was used to produce a 7 kHz square waveform
that had 20 ns rise/fall times (a thousand times faster
than the 23 ?s rise/fall times that characterize the 44.1 kHz
sampling rate of the digital compact-disk).

The electronics used in this experiment was designed
and built in-house because the required combination of response
speed, linearity, power-supply stability, and damping
ability (output impedance) was not found in commercial
headphone-amplifiers. The result was an amplifier with
input and output impedances of 1 M? and 50 m? respectively,
a 3-dB power bandwidth of 0–2.2 MHz, a rise/fall
time of 90 ns, and dc offset voltages under 0.6 mV at all
stages. The linearity of the entire signal chain is essentially
perfect (non-linear errors have sound levels of less than 0
dB SPL; please see below).
"

"
The earphones used were a pair of Grado RS1 (Grado
Laboratories, Brooklyn, New York) supra-aural headphones
which have a frequency response of 12 Hz–30 kHz,
an input resistance of 32 ?, and an efficiency of 98 dB/mW.
Identical signals are fed to both left and right ears to provide
a diotic presentation.
"

"In the main experiment, subjects try to discern differences
between a (7 kHz approximately square-wave shaped)
signal with finite low-pass filtering versus a control signal
with no filtering (waveforms depicted in Fig. 3). The control
tone was perceived to have a sharper or brighter timbre
whereas the filtered one had a duller quality (no difference
in loudness was perceived except for the largest setting of
?=30 ?s). In the blind test, the subject tries to judge
whether an unknown sound is the control or filtered tone
for different settings of ? . It was found in preliminary testing
(especially when ? is close to the threshold) that subjects
needed to listen to the tones for several seconds to
form a lasting impression of the sounds; immediately after
switching the subjects had difficulty assessing whether
anything had changed or not. This again confirms that the
gating itself does not provide a cue.
"
Title: Audibility of phase shifts and time delays
Post by: Woodinville on 2015-04-24 05:59:56
Lets not talk about Kuncher, please.

Instead, generate an FIR highpass filter about 64 samples long that rolls off at 2K, so there is effectively no energy below that.

Calculate the roots. Take all of the roots > 1 and invert them (same frequency response, but minimum phase).

Turn the roots back into a filter.

Now you have a minimum-phase click with no energy below 2kHz.

repeat that click every 10 milliseconds. Make a pulse train. Still no content under 2kHz, note...

Now, delay that by 1 sample at 48kHz in one channel vs. the other. Listen to that vs. the two at the same time, i.e. no delay. Use both speakers and headphones.

See what YOU hear.
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-25 00:51:14
For your inspection, comment and listening enjoyment

in the uploads forum:

Uploads forum post Interchannel delay sampler (http://www.hydrogenaud.io/forums/index.php?showtopic=107570&view=findpost&p=896762)

(1) A 0.45 second 24/192 file with a unit impulse in the middle was created
(2) It was appended 3 times to create a file with 4 impulses
(3) A number of copies were created with various delays added to the Left channel. 1,2,3,4,5,7,9,11,15,19,27 samples @ 192 kHz. Units of about 5 uSec
(4) All files were low pass filtered @ 2 KHz with a 10th order minimum phase low pass filter
(5) All files were down sampled to 4416
(6) all files were normalized to 90% (-1 dB FS)
Title: Audibility of phase shifts and time delays
Post by: mzil on 2015-04-25 01:26:16
Oops, I responded in the upload thread so I'll put it here instead.

I may be hearing artifacts in the conversion, not the clicks.

Code: [Select]
foo_abx 2.0 report
foobar2000 v1.3.3
2015-04-24 17:11:12

File A: reference impulses 4 2klp 4416 norm.flac
SHA1: fb1ca13adfb07144138724d31d58dcec72a40a47
File B: impulses 4 2klp 4416 norm shift 9 samples .flac
SHA1: 257245f5370395be10108dd0dc99ca4f2bcc3a84

Output:
DS : Primary Sound Driver
Crossfading: NO

17:11:12 : Test started.
17:14:23 : 01/01
17:15:17 : 02/02
17:15:38 : 03/03
17:15:51 : 04/04
17:16:33 : 04/05
17:16:52 : 05/06
17:17:37 : 06/07
17:17:45 : 07/08
17:17:45 : Test finished.

 ----------
Total: 7/8
Probability that you were guessing: 3.5%

 -- signature --
d98599b14d1dc93a24479b6bbb054825e2329d6a

I'm pretty sure I can ace this test if tried again because the difference became more obvious to me towards the end.
 

 

Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-25 07:32:33
Oops, I responded in the upload thread so I'll put it here instead.

I may be hearing artifacts in the conversion, not the clicks.

Code: [Select]
foo_abx 2.0 report
foobar2000 v1.3.3
2015-04-24 17:11:12

File A: reference impulses 4 2klp 4416 norm.flac
SHA1: fb1ca13adfb07144138724d31d58dcec72a40a47
File B: impulses 4 2klp 4416 norm shift 9 samples .flac
SHA1: 257245f5370395be10108dd0dc99ca4f2bcc3a84

Output:
DS : Primary Sound Driver
Crossfading: NO

17:11:12 : Test started.
17:14:23 : 01/01
17:15:17 : 02/02
17:15:38 : 03/03
17:15:51 : 04/04
17:16:33 : 04/05
17:16:52 : 05/06
17:17:37 : 06/07
17:17:45 : 07/08
17:17:45 : Test finished.

 ----------
Total: 7/8
Probability that you were guessing: 3.5%

 -- signature --
d98599b14d1dc93a24479b6bbb054825e2329d6a

I'm pretty sure I can ace this test if tried again because the difference became more obvious to me towards the end.

This suggests to me that I need to add the files for 2 and 4 sample delays which I have.

How did you do with the file with a 1 sample delay?

If people reliably detect the file with 1 sample delay, then I need to shift the generation process up to a higher sample rate and provide files with less than 5 uSec interchannel delays.
Title: Audibility of phase shifts and time delays
Post by: mzil on 2015-04-25 09:09:19
27 samples shift is so easy I just run the Xs. My perception is of a lateral shift across the horizon of the soundstage, of the target sound:

Code: [Select]

2015-04-25 00:26:33

File A: reference impulses 4 2klp 4416 norm.flac
SHA1: fb1ca13adfb07144138724d31d58dcec72a40a47
File B: impulses 4 2klp 4416 norm shift 27 samples .flac
SHA1: 80197bcba7ed96997778e7752deb1cf16babffa4

Output:
DS : Primary Sound Driver
Crossfading: NO

00:26:33 : Test started.
00:26:39 : 01/01
00:26:44 : 02/02
00:26:48 : 03/03
00:26:53 : 04/04
00:26:58 : 05/05
00:27:02 : 06/06
00:27:05 : 07/07
00:27:09 : 08/08
00:27:09 : Test finished.

 ----------
Total: 8/8
Probability that you were guessing: 0.4%

 -- signature --
a1da43ab4c78a4ecb41725eacf714077512bdee8


9 samples shift is very difficult. I no longer hear a difference in the target sound however from my perception there is an extremely subtle difference in the background hiss. It is so subtle it is one of those things where one can't tell if it is simply a minute level change, like say .2 dB, or if there is actually a change in the spectral balance of the hiss. All I know is there is a tiny difference.

In any event, since these 9 sample shift results aren't in the spirit of what we are truly testing for, I post them simply to expose how we might see some people [like the organic twins] claim to be able to "hear" a difference at your 9 sample shift level even though it might just be some artifact with your conversion process, I'm not sure:

Code: [Select]

foo_abx 2.0 report
foobar2000 v1.3.3
2015-04-24 23:56:52

File A: impulses 4 2klp 4416 norm shift 9 samples .flac
SHA1: 257245f5370395be10108dd0dc99ca4f2bcc3a84
File B: reference impulses 4 2klp 4416 norm.flac
SHA1: fb1ca13adfb07144138724d31d58dcec72a40a47

Output:
DS : Primary Sound Driver
Crossfading: NO

23:56:52 : Test started.
23:57:42 : 01/01
23:57:49 : 02/02
23:57:56 : 03/03
23:58:01 : 04/04
23:58:13 : 05/05
23:58:17 : 06/06
23:58:22 : 07/07
23:58:26 : 08/08
23:58:30 : 09/09
23:58:36 : 10/10
23:58:36 : Test finished.

 ----------
Total: 10/10
Probability that you were guessing: 0.1%

 -- signature --
bbdaa32033488f9bd2442eed5bc22c34151f22cf


I can't hear any difference, of any kind, on the 1 sample shift level.

P.S. What is that oddball file of the five in your folder, recorded at a lower level, also with "9 samples" as part of its name? ... Oh wait. It doesn't say "norm", so that means it hasn't been normalized I guess?... Hmm, maybe your normalizing procedure is what I'm keying on for that change to the hiss at the 9 sample shift level?


Title: Audibility of phase shifts and time delays
Post by: Kees de Visser on 2015-04-25 10:32:26
(4) All files were low pass filtered @ 2 KHz with a 10th order minimum phase low pass filter
Why did you use a 2 kHz low pass filter ? AFAIK Woodinville proposed a 2 kHz high pass filter for the pulse train. Have I missed something ?

See what YOU hear.
JJ: can your pulse train be made with a DAW ? Your method sounds like a Matlab style recipe to me and I don't think I have the tools. Perhaps you can tell us what scientific literature tells us what we are supposed to hear ? I honestly don't care much what "I" can hear.
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-25 11:47:13
(4) All files were low pass filtered @ 2 KHz with a 10th order minimum phase low pass filter
Why did you use a 2 kHz low pass filter ? AFAIK Woodinville proposed a 2 kHz high pass filter for the pulse train. Have I missed something ?


I applied the 2 KHz low pass filter to all the samples, both reference and shifted, not because that seemed to be indicated by JJ's instructions but because it ensured that the listening test was all about timing, and not the least about differences in frequency response.  These test files are IMO not very demanding and they should play reasonably accurately on a wide variety of different kinds of equipment. This makes a good test of listeners, not equipment.
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-25 11:54:35
27 samples shift is so easy I just run the Xs. My perception is of a lateral shift across the horizon of the soundstage, of the target sound:


That was completely intentional - its the confidence building training file.

Quote
P.S. What is that oddball file of the five in your folder, recorded at a lower level, also with "9 samples" as part of its name? ... Oh wait. It doesn't say "norm", so that means it hasn't been normalized I guess?... Hmm, maybe your normalizing procedure is what I'm keying on for that change to the hiss at the 9 sample shift level?


That was a mistake - an intermediate file that crept in and is removed from the new data zip file which contains an additional 2 files for 2 sample and 4 sample (@192 KHz) delay testing:

http://www.hydrogenaud.io/forums/index.php...st&p=896783 (http://www.hydrogenaud.io/forums/index.php?showtopic=107570&view=findpost&p=896783)

Contents:

The reference file. No interchannel delay. It's a downsampled version of the file I used to build the files with the interchannel delays. Compare all the other files to it.

27 sample delay = 140.6 uSec - this is a confidence builder - you should be able to complete it easily and accurately. It is also a test of the suitability of your test environment. Do not proceed to the shorter delays until you can do well with this file.

9 sample delay = 46.85 uSec - This is roughly the result that David L. Clark reported.

4 sample delay = 20.8 uSec

2 sample delay = 10.4 uSec

1 sample delay = 5.2 uSec

If you edit these files and zoom in on one of the impulses and compare the L&R channel you should see the delay which will be less than 1 sample wide for the last 3 files.  This should put to rest any misapprehensions about 44 Khz sampling not being able to encode time differences with high resolution.
Title: Audibility of phase shifts and time delays
Post by: bandpass on 2015-04-25 12:45:18
JJ: can your pulse train be made with a DAW ? Your method sounds like a Matlab style recipe to me and I don't think I have the tools.

You can do it with sox (if I understand JJ's instructions correctly):

Code: [Select]
sox -V -n delay-1.wav synth 1s sq pad 100s 379s sinc -a 82 -t 4k -M 4k repeat 399 channels 2 delay 1s


Assuming that –82 dB is sufficient for the stop-band.  This gives filter order 64; transition-band between 2k & 6k.

Omit the 'delay 1s' for the no-delay version.
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-25 14:02:55
JJ: can your pulse train be made with a DAW ? Your method sounds like a Matlab style recipe to me and I don't think I have the tools.

You can do it with sox (if I understand JJ's instructions correctly):

Code: [Select]
sox -V -n delay-1.wav synth 1s sq pad 100s 379s sinc -a 82 -t 4k -M 4k repeat 399 channels 2 delay 1s


Assuming that –82 dB is sufficient for the stop-band.  This gives filter order 64; transition-band between 2k & 6k.

Omit the 'delay 1s' for the no-delay version.



Seems like its time to grind out a set of samples and put them into the uploads forum.  There may be some wisdom to glean from the comparison of how the two sets of samples work.
Title: Audibility of phase shifts and time delays
Post by: bandpass on 2015-04-25 14:52:51
Seems like its time to grind out a set of samples and put them into the uploads forum.  There may be some wisdom to glean from the comparison of how the two sets of samples work.

Two files here (http://www.hydrogenaud.io/forums/index.php?s=&showtopic=107570&view=findpost&p=896798).
Title: Audibility of phase shifts and time delays
Post by: mzil on 2015-04-25 20:04:29
4 sample delay = 20.8 uSec
 

  Here's my first shot at it and might I add this is before my first cup of coffee for the day. 

Code: [Select]

foo_abx 2.0 report
foobar2000 v1.3.3
2015-04-25 11:19:24

File A: impulses 4 2klp 4416 norm shift 4 samples.flac
SHA1: 301a8107b8d9e8f0a61ac6f722aacfd86a088722
File B: reference impulses 4 2klp 4416 norm.flac
SHA1: fb1ca13adfb07144138724d31d58dcec72a40a47

Output:
DS : Primary Sound Driver
Crossfading: NO

11:19:24 : Test started.
11:21:46 : 00/01
11:22:04 : 01/02
11:22:21 : 02/03
11:22:55 : 03/04
11:23:15 : 04/05
11:23:34 : 05/06
11:24:40 : 06/07
11:24:50 : 07/08
11:24:50 : Test finished.

 ----------
Total: 7/8
Probability that you were guessing: 3.5%

 -- signature --
07a9ed0ac5de3bdebe80d0761ce9e408ee91b465


The good news is here on this test, above,  I'm truly, at least to the best of my ability, relying solely on directional cues of the target sound and ignoring any differences in the background hiss . [But who knows, maybe I am keying on it at a subconscious level]. Would you please speak to why I'm hearing, or thinking I'm hearing, differences in the background hiss on my earlier test. Thanks.

Interestingly my perceived HRTF mental processing for directional cues is claiming not just a lateral shift but also a tiny vertical shift, almost as if you snuck in a small EQ alteration to elicit that, ha-ha.

Tomlinson Holman says we can discern about 10,000+ distinct directions. I wonder if checking his work, and that of others, to determine in what direction and frequency we are most sensitive to tiny differences in directional arc, measured in degrees, and then designing these tests using those frequencies in those directions, makes better sense to maximize sensitivity? For instance, maybe we should be using 3.5 kHz tones at the time spacing and L vs. R balance level which directionally implies slightly to the side of center? Just a guess.

This  4 sample shift wasn't easy but instead of taking 5 minutes to test I suspect a shorter delay of half that, 2 samples, would take an hour of concentration or more, if I can even discern it at all, so I don't think I'll be trying that any time soon. Do you have a 3 sample shift file? That might be doable. Please post it and thanks for these tests. They are fun and educational.
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-25 20:17:13
added file for 3 sample shift:

http://www.hydrogenaud.io/forums/index.php...st&p=896839 (http://www.hydrogenaud.io/forums/index.php?showtopic=107570&view=findpost&p=896839)
Title: Audibility of phase shifts and time delays
Post by: mzil on 2015-04-25 21:22:38
Thanks, I haven't tackled that 3 sample one yet, but here's my second stab at 4 sample shift, after coffee, to verify the first try wasn't just dumb luck. 9/10 score not too shabby:

Code: [Select]

foo_abx 2.0 report
foobar2000 v1.3.3
2015-04-25 12:52:35

File A: impulses 4 2klp 4416 norm shift 4 samples.flac
SHA1: 301a8107b8d9e8f0a61ac6f722aacfd86a088722
File B: reference impulses 4 2klp 4416 norm.flac
SHA1: fb1ca13adfb07144138724d31d58dcec72a40a47

Output:
DS : Primary Sound Driver
Crossfading: NO

12:52:35 : Test started.
12:53:18 : 01/01
12:53:44 : 02/02
12:54:09 : 03/03
12:54:36 : 04/04
12:55:06 : 05/05
12:55:45 : 05/06
12:56:13 : 06/07
12:56:47 : 07/08
12:58:17 : 08/09
12:58:51 : 09/10
12:58:51 : Test finished.

 ----------
Total: 9/10
Probability that you were guessing: 1.1%

 -- signature --
695a168ab31fd7a7d6d6e73a44b7c27772740f85


--

WOW! They've added keyboard shortcuts so A, B, and X can be selected by their keyboard letters! Woo-hoo. News to me. That really does reduce cognitive load and greatly adds to my sensitivity. Has that always been there but I didn't notice it?

Previously, these were the mental steps:

Thought: "I want to hear the alternate source B now"

A) Stop listening to music

B) open eyes

C) activate visual cortex

D) locate mouse cursor on screen

E) place hand on mouse

F) move hand in analogous fashion to desired cursor motion

G) make minor tracking adjustments as cursor moves across screen

H) stop on "B" box

I) correct for overshoot if applicable

J) click "B" box

K) disengage visual processing

L) close eyes

M) listen to newly selected B source

N) compare to "A" sound in mind's memory, from a couple of seconds ago, prior to the gymnastics to get to "B", and pretend that there was no loss of concentration from having to go through that multi-step process, steps A through M above, which involved switching to an alternate sense, sight, and then returning back to hearing.

NOW, thanks to the keyboard shortcuts:

Thought: "I want to hear the alternate source B now"

A) press downward on right hand  finger already in position, resting on keyboard letter "B"

B) listen to newly selected B source, INSTANTLY, to see if the transition was detectable. "Memory" not really being used, in my opinion.

Hallelujah!
Title: Audibility of phase shifts and time delays
Post by: mzil on 2015-04-26 00:36:05
I tried and I am unable to hear 3 samples shift. Looks like under these conditions my best is 4 sample shift.
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-26 03:01:48
I tried and I am unable to hear 3 samples shift. Looks like under these conditions my best is 4 sample shift.



Take heart! Clark's test had results that showed the threshold of detection as being about twice that. I ascribe the difference to our casting it as an ABX test which was not nearly as easy to work out way back when.

Hopefully others will try it and report their results so we are not reliant on the results of just one person. I tried it and my old ears barely make it through the first test with 95% confidence.
Title: Audibility of phase shifts and time delays
Post by: bandpass on 2015-04-27 09:21:01
Attached files generated by:

Code: [Select]
for d in 0 1; do sox -V -n delay-${d}.flac synth 1s sq pad 100s 379s sinc -a 82 -t 4k -M 4k repeat 399 channels 2 delay ${d}s norm -1 trim 0 4; done


My analysis is that the file with the delay adds a 1 sample (22 uSec) delay to the left channel where the pulse rate is 100 Hz.

Good, I believe that's what JJ intended.

Can this methodology produce delays that are other than 1 sample?

Yes, any positive integer can be substituted.

Can this methodology produce delays that are less than 1 sample?

I think a linear-phase sub-sample delay requires a rate change (but that's easy enough to add).

Of course, as shorter delay could be introduced simply by running the chain at a higher rate (but bear in mind possible resampling in the OS or hardware).

What are the typical ABX results?

Don't know if I'm typical, but I couldn't ABX the two files.
Title: Audibility of phase shifts and time delays
Post by: Arnold B. Krueger on 2015-04-27 13:59:24
Attached files generated by:

Code: [Select]
for d in 0 1; do sox -V -n delay-${d}.flac synth 1s sq pad 100s 379s sinc -a 82 -t 4k -M 4k repeat 399 channels 2 delay ${d}s norm -1 trim 0 4; done


My analysis is that the file with the delay adds a 1 sample (22 uSec) delay to the left channel where the pulse rate is 100 Hz.

Good, I believe that's what JJ intended.

Can this methodology produce delays that are other than 1 sample?

Yes, any positive integer can be substituted.

Can this methodology produce delays that are less than 1 sample?

I think a linear-phase sub-sample delay requires a rate change (but that's easy enough to add).

Of course, as shorter delay could be introduced simply by running the chain at a higher rate (but bear in mind possible resampling in the OS or hardware).

What are the typical ABX results?

Don't know if I'm typical, but I couldn't ABX the two files.



FWIW (not much given the age-attenuated state of my hearing) neither can I. 

However, the one value of delay we have at hand is one that has been only slightly bettered by other means, and is in the range where audibility may be elusive for many listeners.

The next logical step would appear to be to produce a family of files similar to those already posted on the downloads forum for the other methodology, covering a range of delays, starting with one that is so large as to be unmistakable (e.g. about 140 uSec).
Title: Audibility of phase shifts and time delays
Post by: bandpass on 2015-08-26 08:10:44
In short, 10 microseconds is nowhere near the actual abilities of 4416 to convey timing differences among
 signals is:  1/(44100*65536) =

3.460042871315193e-10  seconds. Less than a nanosecond.
Let's see how this pans out in a simple experiment:
Does ‘degrading’ to redbook preserve the 1ns difference?

Results of three runs (showing sample values around the peak):
Code: [Select]
            Time    Input 1         Input 2         Output 1        Output 2
          -----    -------------  ------------    -------------  ------------
              0    0.99999992736  0.99999989569  0.9953168775    0.99532626430
          1e-09    0.99999995297  0.99999992736  0.99531690404  0.99532629550
          2e-09    0.99999997392  0.99999995297  0.99531692453  0.99532632064
          3e-09    0.99999998789  0.99999997392  0.99531693989  0.99532634020
          4e-09    0.99999999674  0.99999998789  0.99531694921  0.99532635370
          5e-09    0.99999999953*  0.99999999674  0.99531695293*  0.99532636208
          6e-09    0.99999999674  0.99999999953*  0.99531695107  0.99532636441*
          7e-09    0.99999998789  0.99999999674  0.99531694362  0.99532636115
          8e-09    0.99999997392  0.99999998789  0.99531693012  0.99532635184
          9e-09    0.99999995297  0.99999997392  0.99531691102  0.99532633740
          1e-08    0.99999992736  0.99999995297  0.99531688634  0.99532631692

              0    0.99999992736  0.99999989569  0.99531617435  0.99532265775
          1e-09    0.99999995297  0.99999992736  0.99531620182  0.99532268988
          2e-09    0.99999997392  0.99999995297  0.99531622371  0.99532271642
          3e-09    0.99999998789  0.99999997392  0.99531624001  0.99532273691
          4e-09    0.99999999674  0.99999998789  0.99531625025  0.99532275181
          5e-09    0.99999999953*  0.99999999674  0.99531625491*  0.99532276113
          6e-09    0.99999999674  0.99999999953*  0.99531625398  0.99532276485*
          7e-09    0.99999998789  0.99999999674  0.99531624746  0.99532276252
          8e-09    0.99999997392  0.99999998789  0.99531623488  0.99532275461
          9e-09    0.99999995297  0.99999997392  0.99531621672  0.99532274110
          1e-08    0.99999992736  0.99999995297  0.99531619297  0.99532272201

              0    0.99999992736  0.99999989569  0.99532299675  0.99532591924
          1e-09    0.99999995297  0.99999992736  0.99532302469  0.99532594904
          2e-09    0.99999997392  0.99999995297  0.99532304704  0.99532597326
          3e-09    0.99999998789  0.99999997392  0.99532306381  0.99532599188
          4e-09    0.99999999674  0.99999998789  0.99532307498  0.99532600446
          5e-09    0.99999999953*  0.99999999674  0.99532308057*  0.99532601144
          6e-09    0.99999999674  0.99999999953*  0.99532308010  0.99532601284*
          7e-09    0.99999998789  0.99999999674  0.99532307405  0.99532600865
          8e-09    0.99999997392  0.99999998789  0.99532306241  0.99532599840
          9e-09    0.99999995297  0.99999997392  0.99532304518  0.99532598257
          1e-08    0.99999992736  0.99999995297  0.99532302190  0.99532596115
Input 1 shows the peak at 5ns in.  As expected, input 2 shows exactly the same samples as input 1 but delayed 1ns.

The outputs vary between runs and are noisy (as expected, due to the random TPDF dither applied at the redbook conversion stage) but look good: in each case, the peak value time matches that of the respective input (peak values have been highlighted with asterisks).

The slight attenuation in the outputs may be due to headroom being applied at the dither stage and/or HF roll-off during resampling but in any case, it's timing we're interested in here.

So the conclusion is yes, redbook does preserve timing at least down to 1 ns.

Code used (N.B. sox v14.4.2):
Code: [Select]
sox -r 1000000k -n in1.wav synth 1s sq pad .1 .1 rate -vtf 44100 norm rate -vtf 1000000k norm
sox in1.wav in2.wav delay 1s
sox in1.wav -b 16 tmp.wav rate -vtf 44100
sox tmp.wav -b 32 out1.wav rate -vtf 1000000k
sox in2.wav -b 16 tmp.wav rate -vtf 44100
sox tmp.wav -b 32 out2.wav rate -vtf 1000000k
sox -M in1.wav in2.wav out1.wav out2.wav -t dat - trim .099999995 11s
Title: Audibility of phase shifts and time delays
Post by: Green Marker on 2015-08-26 09:27:31
Let's see how this pans out in a simple experiment:


I like.

So are you saying that with a simple experiment you have overturned a major tenet on which MQA and other 'high res' systems claim superiority?
Title: Audibility of phase shifts and time delays
Post by: bandpass on 2015-08-26 10:30:39
Yes, please point  ‘temporal smearing’ proponents in the direction of my post.

I can't claim to be the first to have done something along these lines (axon described a similar experiment back in 2006), but I've done it in a way that is hopefully accessible to all and easily repeatable by anyone who wants to do so.
Title: Audibility of phase shifts and time delays
Post by: Woodinville on 2015-09-05 01:26:44
I forget which is which, but one of Proakis' books explains lots of ways to do pretty much infinitely variable sub-sample shift via interpolation by about 10 plus a 6 tap fixed-formula filter operating on the 10x filter response.
Title: Re: Audibility of phase shifts and time delays
Post by: danadam on 2017-11-22 02:44:18
I downloaded those files already a year ago, so I can't promise I'll try 3 samples delay any time soon. Probably in another year ;-)
Code: [Select]
foo_abx 2.0.2 report
foobar2000 v1.3.10
2017-11-22 02:35:21

File A: Impulses shift 0 samples 2klp norm 4416 .flac
SHA1: 8fc00a4bb6a1bb0a66ec5c83cfaa36f9d8fddd13
File B: Impulses shift 4 samples 2klp norm 4416 .flac
SHA1: 6133aaa124c97a3f768f3d9216af2eb07b7c0bf3

Output:
ASIO : ASIO4ALL v2
Crossfading: NO

02:35:21 : Test started.
02:38:48 : 01/01
02:39:48 : 02/02
02:40:31 : 03/03
02:42:31 : 04/04
02:44:00 : 05/05
02:44:54 : 06/06
02:47:26 : 07/07
02:49:35 : 08/08
02:56:36 : 09/09
02:57:49 : 10/10
02:59:46 : 11/11
03:01:23 : 12/12
03:03:16 : 13/13
03:04:04 : 14/14
03:07:30 : 15/15
03:07:30 : Test finished.

 ----------
Total: 15/15
Probability that you were guessing: 0.0%

 -- signature --
b76d1c64aff74736f104c5cd98c431b6c12b6dc8
Title: Re: Audibility of phase shifts and time delays
Post by: danadam on 2017-11-22 03:00:04
And while I revived that old thread... the formula from the first page:
Code: [Select]
1/ ( 2 pi bandwidth number_of_levels)
I wanted to make sure I got it right. It is a bandwidth, not a sampling rate? And a number of levels of concrete signal, meaning that timing precision of a signal depends on its level? So for full scale 44/16 signal it is:
Code: [Select]
1 / ( 2 * pi * 22050 * 2^16)
Is that correct?

Also, is there any "citable" online source for this? (not that I don't believe it :-) )
Title: Re: Audibility of phase shifts and time delays
Post by: splice on 2017-11-23 01:22:16
I believe it's sample rate.
"The time resolution of a 16 bit, 44.1khz PCM channel is not limited to the 22.7µs time difference between samples. The actual minimum time resolution is equivalent to 1/(2pi * quantization levels * sample rate). For 16/44.1, that is 1/(2pi * 65536 * 44100), which is about 55 picoseconds. To put that in perspective, light travels less than an inch in that time. "
Title: Re: Audibility of phase shifts and time delays
Post by: saratoga on 2017-11-23 15:04:38
I believe it is sample rate, although the difference is only a factor of 2.

The ability to introduce subsample shifts in noiseless waveforms is not very interesting, but volumes have been written on estimating shifts in noisy waveforms for radar, sonar, ultrasound, astronomy, interferometry and so on. There is probably a derivation of the noiseless case somewhere.
Title: Re: Audibility of phase shifts and time delays
Post by: jsdyson on 2017-11-23 15:35:34
I have an anecdote -- not well tested, but something that SEEMS to be real to me:   when I was running tests on some software, I had the historical habit of using a low pass filter like this on SOX --- seemed to make the highs sound a little tighter but without significantly boosting the average HF content:  'lowpass -2 9k 1q lowpass -2 18k 1q'.   Sometimes, I'd use 15k instead of 18k depending upon material.   The slight tightening of the transients before loosing the HF further up seemed to help... but... I also noticed that some of the spatial perception was lost.   The rolloff or very slight 'ringing' -- actually manifest as a near flat response to the cutoff instead of some dB of loss -- seemed to have caused a SLIGHT loss of locality.  I am not one of the 'wine-taster' audio people -- and if I hadn't noticed this 'out of the blue' several times, I wouldn't be mentioning it.   So, I no longer use a gratuitous 'slightly peaky' rolloff to get a sharper sound without a very strong justification.  Now, if I want to SLIGHTLY brighten or smooth out some harsh HF edge, I might use such a filter, but not without carefully considered reason.
Title: Re: Audibility of phase shifts and time delays
Post by: Funkstar De Luxe on 2017-11-23 15:43:16
I have a license for iZotope RX which can shift channels by 0.1 increments of a sample. More than willing to use it to do some listening tests if anyone is interested.

The tool is called Azimuth adjustment.
Title: Re: Audibility of phase shifts and time delays
Post by: bandpass on 2017-11-24 06:19:25
And while I revived that old thread... the formula from the first page:
Code: [Select]
1/ ( 2 pi bandwidth number_of_levels)
I wanted to make sure I got it right.
There’s a derivation of the formula here (https://www.computeraudiophile.com/forums/topic/30381-mqa-is-vaporware/?do=findComment&comment=725238) (with refinement in the following few posts).