Measuring \"payload\" skew across network...

[Don Y]

I have an application that distributes audio among multiple, distant
nodes (emitters/speakers) over an ethernet. I control the relative
timing (phasing) of the signals from each emitter -- either in synchronism
or leading/lagging by specific dynamic amounts (which can be on the
order of many ms).

To date, I\'ve been measuring this skew in the time domain by emitting
a plosive burst whose start time can readily be identified (on a \'scope)
relative to the time it appears from another emitter. This is the
direct analog of a \"clap board\" in a film production.

Is there a cleverer way I can do this -- likely in the frequency
domain -- by deploying a signal of particular \"characteristics\"
that would let me \"miss\" the start time of the emission and still
yield the same results? (it\'s just annoyingly inconvenient to
have to set up each experiment with an introductory delay just so
I can be prepared for the output(s). A \"steady state\" sort of
approach would be nicer -- set up a stimulus, then probe each
emitter to check results for that emitter!)

[I.e., you need to be able to differentiate between THIS point on
the time series on THIS device and THE SAME point on another device.
You can\'t risk that point being on a subsequent cycle of a periodic
waveform!]

 

[server]

On Thu, 24 Sep 2020 07:51:10 -0700, Don Y
<blockedofcourse@foo.invalid> wrote:

I have an application that distributes audio among multiple, distant
nodes (emitters/speakers) over an ethernet. I control the relative
timing (phasing) of the signals from each emitter -- either in synchronism
or leading/lagging by specific dynamic amounts (which can be on the
order of many ms).

To date, I\'ve been measuring this skew in the time domain by emitting
a plosive burst whose start time can readily be identified (on a \'scope)
relative to the time it appears from another emitter. This is the
direct analog of a \"clap board\" in a film production.

Is there a cleverer way I can do this -- likely in the frequency
domain -- by deploying a signal of particular \"characteristics\"
that would let me \"miss\" the start time of the emission and still
yield the same results? (it\'s just annoyingly inconvenient to
have to set up each experiment with an introductory delay just so
I can be prepared for the output(s). A \"steady state\" sort of
approach would be nicer -- set up a stimulus, then probe each
emitter to check results for that emitter!)

[I.e., you need to be able to differentiate between THIS point on
the time series on THIS device and THE SAME point on another device.
You can\'t risk that point being on a subsequent cycle of a periodic
waveform!]

Send something aperiodic and wideband - a long chirp or noise or
something - and cross-correlate. I assume the emitters are identical,
which simplifies the situation.

In theory, that could be done continuously at inaudible levels.



--

John Larkin Highland Technology, Inc

Science teaches us to doubt.

Claude Bernard

 

[Martin Brown]

On 24/09/2020 15:51, Don Y wrote:

I have an application that distributes audio among multiple, distant
nodes (emitters/speakers) over an ethernet.  I control the relative
timing (phasing) of the signals from each emitter -- either in synchronism
or leading/lagging by specific dynamic amounts (which can be on the
order of many ms).

To date, I\'ve been measuring this skew in the time domain by emitting
a plosive burst whose start time can readily be identified (on a \'scope)
relative to the time it appears from another emitter.  This is the
direct analog of a \"clap board\" in a film production.

Is there a cleverer way I can do this -- likely in the frequency
domain -- by deploying a signal of particular \"characteristics\"
that would let me \"miss\" the start time of the emission and still
yield the same results?  (it\'s just annoyingly inconvenient to
have to set up each experiment with an introductory delay just so
I can be prepared for the output(s).  A \"steady state\" sort of
approach would be nicer -- set up a stimulus, then probe each
emitter to check results for that emitter!)

Chirp might well do it. Same technology as bats use but with a fairly
high computational load reducing the data.

You can even modulate each emitter with a characteristic start
frequency, sweep rate and duration to be able to tell them apart.

The other choice might be periodic Walsh functions on a length 2^N which
are an orthogonal basis set on binary sequences.

https://mathworld.wolfram.com/WalshFunction.html

IOW You can uniquely determine the delay from each emitter in one go if
you can sample the resulting waveform fast enough.

--
Regards,
Martin Brown

 

[Lasse Langwadt Christensen]

torsdag den 24. september 2020 kl. 17.29.36 UTC+2 skrev Martin Brown:

On 24/09/2020 15:51, Don Y wrote:
I have an application that distributes audio among multiple, distant
nodes (emitters/speakers) over an ethernet.  I control the relative
timing (phasing) of the signals from each emitter -- either in synchronism
or leading/lagging by specific dynamic amounts (which can be on the
order of many ms).

To date, I\'ve been measuring this skew in the time domain by emitting
a plosive burst whose start time can readily be identified (on a \'scope)
relative to the time it appears from another emitter.  This is the
direct analog of a \"clap board\" in a film production.

Is there a cleverer way I can do this -- likely in the frequency
domain -- by deploying a signal of particular \"characteristics\"
that would let me \"miss\" the start time of the emission and still
yield the same results?  (it\'s just annoyingly inconvenient to
have to set up each experiment with an introductory delay just so
I can be prepared for the output(s).  A \"steady state\" sort of
approach would be nicer -- set up a stimulus, then probe each
emitter to check results for that emitter!)

Chirp might well do it. Same technology as bats use but with a fairly
high computational load reducing the data.

You can even modulate each emitter with a characteristic start
frequency, sweep rate and duration to be able to tell them apart.

The other choice might be periodic Walsh functions on a length 2^N which
are an orthogonal basis set on binary sequences.

https://mathworld.wolfram.com/WalshFunction.html

IOW You can uniquely determine the delay from each emitter in one go if
you can sample the resulting waveform fast enough.

walsh, maximum length or gold codes and correlation, basically GPS or CDMA

 

[Don Y]

On 9/24/2020 8:29 AM, Martin Brown wrote:

On 24/09/2020 15:51, Don Y wrote:
I have an application that distributes audio among multiple, distant
nodes (emitters/speakers) over an ethernet. I control the relative
timing (phasing) of the signals from each emitter -- either in synchronism
or leading/lagging by specific dynamic amounts (which can be on the
order of many ms).

To date, I\'ve been measuring this skew in the time domain by emitting
a plosive burst whose start time can readily be identified (on a \'scope)
relative to the time it appears from another emitter. This is the
direct analog of a \"clap board\" in a film production.

Is there a cleverer way I can do this -- likely in the frequency
domain -- by deploying a signal of particular \"characteristics\"
that would let me \"miss\" the start time of the emission and still
yield the same results? (it\'s just annoyingly inconvenient to
have to set up each experiment with an introductory delay just so
I can be prepared for the output(s). A \"steady state\" sort of
approach would be nicer -- set up a stimulus, then probe each
emitter to check results for that emitter!)

Chirp might well do it. Same technology as bats use but with a fairly high
computational load reducing the data.

You can even modulate each emitter with a characteristic start frequency, sweep
rate and duration to be able to tell them apart.

The other choice might be periodic Walsh functions on a length 2^N which are an
orthogonal basis set on binary sequences.

https://mathworld.wolfram.com/WalshFunction.html

IOW You can uniquely determine the delay from each emitter in one go if you can
sample the resulting waveform fast enough.

Can I get an *instrument* to do this for me? E.g., I\'ve been using a DSO
and just letting it measure the delay between two signals.

 

[Jeff Liebermann]

On Thu, 24 Sep 2020 07:51:10 -0700, Don Y
<blockedofcourse@foo.invalid> wrote:

Is there a cleverer way I can do this -- likely in the frequency
domain -- by deploying a signal of particular \"characteristics\"
that would let me \"miss\" the start time of the emission and still
yield the same results?

Look into various automated room equalization schemes used in higher
end hi-fi speaker systems. For example, from 2013:
<https://hometheaterreview.com/automated-room-correction-explained/>
I\'m sure there are patents available for dissection.
<https://www.google.com/search?tbm=pts&q=room+equalization>

My Harman Kardon AVR 254 amp has such a feature. I place a plug in
microphone near where I expect to be sitting when listening to the
hi-fi, punch the \"setup\" function, and the amp adjusts the delays and
frequency response to each of the 7 speakers (channels). The noises
emitted during the setup process sound like a mixture of clicks (time
of flight delay?) and hiss (broadband noise). Actually, I\'m not sure
if it can deal with variations in speaker distance and delays from the
listener. The diagram in the manual showing recommended placement for
the 7 speakers is a perfect circle, which suggests that it can\'t
handle random speaker placement. Still, I think it might be worth
investigating.


--
Jeff Liebermann jeffl@cruzio.com
150 Felker St #D http://www.LearnByDestroying.com
Santa Cruz CA 95060 http://802.11junk.com
Skype: JeffLiebermann AE6KS 831-336-2558

 

[Chris]

On 09/24/20 15:51, Don Y wrote:

I have an application that distributes audio among multiple, distant
nodes (emitters/speakers) over an ethernet. I control the relative
timing (phasing) of the signals from each emitter -- either in synchronism
or leading/lagging by specific dynamic amounts (which can be on the
order of many ms).

To date, I\'ve been measuring this skew in the time domain by emitting
a plosive burst whose start time can readily be identified (on a \'scope)
relative to the time it appears from another emitter. This is the
direct analog of a \"clap board\" in a film production.

Is there a cleverer way I can do this -- likely in the frequency
domain -- by deploying a signal of particular \"characteristics\"
that would let me \"miss\" the start time of the emission and still
yield the same results? (it\'s just annoyingly inconvenient to
have to set up each experiment with an introductory delay just so
I can be prepared for the output(s). A \"steady state\" sort of
approach would be nicer -- set up a stimulus, then probe each
emitter to check results for that emitter!)

[I.e., you need to be able to differentiate between THIS point on
the time series on THIS device and THE SAME point on another device.
You can\'t risk that point being on a subsequent cycle of a periodic
waveform!]

If yu are using ethernet, transit time across the network can depend
on network load, retries etc and is not guaranteed. For that sort of
work, a private subnet or a different protocol may be more appropriate.

Have you modelled worst case transit times ?...

Chris

 

[Robert Baer]

jlarkin@highlandsniptechnology.com wrote:

On Thu, 24 Sep 2020 07:51:10 -0700, Don Y
blockedofcourse@foo.invalid> wrote:

I have an application that distributes audio among multiple, distant
nodes (emitters/speakers) over an ethernet. I control the relative
timing (phasing) of the signals from each emitter -- either in synchronism
or leading/lagging by specific dynamic amounts (which can be on the
order of many ms).

To date, I\'ve been measuring this skew in the time domain by emitting
a plosive burst whose start time can readily be identified (on a \'scope)
relative to the time it appears from another emitter. This is the
direct analog of a \"clap board\" in a film production.

Is there a cleverer way I can do this -- likely in the frequency
domain -- by deploying a signal of particular \"characteristics\"
that would let me \"miss\" the start time of the emission and still
yield the same results? (it\'s just annoyingly inconvenient to
have to set up each experiment with an introductory delay just so
I can be prepared for the output(s). A \"steady state\" sort of
approach would be nicer -- set up a stimulus, then probe each
emitter to check results for that emitter!)

[I.e., you need to be able to differentiate between THIS point on
the time series on THIS device and THE SAME point on another device.
You can\'t risk that point being on a subsequent cycle of a periodic
waveform!]

Send something aperiodic and wideband - a long chirp or noise or
something - and cross-correlate. I assume the emitters are identical,
which simplifies the situation.

In theory, that could be done continuously at inaudible levels.

As in using a burst of white noise from a given speaker for timing
purposes.
Could be less noisy than the plosive burst.

 

[Don Y]

On 9/24/2020 9:53 AM, Chris wrote:

On 09/24/20 15:51, Don Y wrote:
If yu are using ethernet, transit time across the network can depend
on network load, retries etc and is not guaranteed. For that sort of
work, a private subnet or a different protocol may be more appropriate.

Have you modelled worst case transit times ?...

Of course! Every device on the switch is of my own design -- I know
exactly what it sends/receives and when. The switch participates in the
clock synchronization algorithm (similar to PTP).

So, I distribute the \"content\" slightly ahead of time. Then, tell
each node when to emit it, relative to that synchronized \"clock\". The
issue then distills to ensuring I can keep ahead of the \"consumers\"
(cuz I can\'t deliver a complete \"audio snipet\")

 

[Don Y]

On 9/24/2020 9:51 AM, Jeff Liebermann wrote:

On Thu, 24 Sep 2020 07:51:10 -0700, Don Y
blockedofcourse@foo.invalid> wrote:

Is there a cleverer way I can do this -- likely in the frequency
domain -- by deploying a signal of particular \"characteristics\"
that would let me \"miss\" the start time of the emission and still
yield the same results?

Look into various automated room equalization schemes used in higher
end hi-fi speaker systems. For example, from 2013:
https://hometheaterreview.com/automated-room-correction-explained/
I\'m sure there are patents available for dissection.
https://www.google.com/search?tbm=pts&q=room+equalization

Thanks!

My Harman Kardon AVR 254 amp has such a feature. I place a plug in
microphone near where I expect to be sitting when listening to the
hi-fi, punch the \"setup\" function, and the amp adjusts the delays and
frequency response to each of the 7 speakers (channels). The noises
emitted during the setup process sound like a mixture of clicks (time
of flight delay?) and hiss (broadband noise). Actually, I\'m not sure
if it can deal with variations in speaker distance and delays from the
listener. The diagram in the manual showing recommended placement for
the 7 speakers is a perfect circle, which suggests that it can\'t
handle random speaker placement. Still, I think it might be worth
investigating.

OK. That\'s a different goal than what I\'m pursuing but I can
possibly borrow some ideas/techniques from it.

In my case, the individual \"nodes\" are located over reasonably large
distances (50 ft) and not \"line-of-hearing\" (concocted term as a play
on \"line-of-sight\").

I also want to be able to numerically demonstrate the individual delays
(hence the DSO time-domain approach... no \"magic\" involved in convincing
yourself that a SPECIFIC delay has been achieved)

(later, I\'ll need to do the same with video; and the audio wrt the video)

 

[Don Y]

On 9/24/2020 8:06 AM, jlarkin@highlandsniptechnology.com wrote:

On Thu, 24 Sep 2020 07:51:10 -0700, Don Y
blockedofcourse@foo.invalid> wrote:

I have an application that distributes audio among multiple, distant
nodes (emitters/speakers) over an ethernet. I control the relative
timing (phasing) of the signals from each emitter -- either in synchronism
or leading/lagging by specific dynamic amounts (which can be on the
order of many ms).

To date, I\'ve been measuring this skew in the time domain by emitting
a plosive burst whose start time can readily be identified (on a \'scope)
relative to the time it appears from another emitter. This is the
direct analog of a \"clap board\" in a film production.

Is there a cleverer way I can do this -- likely in the frequency
domain -- by deploying a signal of particular \"characteristics\"
that would let me \"miss\" the start time of the emission and still
yield the same results? (it\'s just annoyingly inconvenient to
have to set up each experiment with an introductory delay just so
I can be prepared for the output(s). A \"steady state\" sort of
approach would be nicer -- set up a stimulus, then probe each
emitter to check results for that emitter!)

[I.e., you need to be able to differentiate between THIS point on
the time series on THIS device and THE SAME point on another device.
You can\'t risk that point being on a subsequent cycle of a periodic
waveform!]

Send something aperiodic and wideband - a long chirp or noise or
something - and cross-correlate. I assume the emitters are identical,
which simplifies the situation.

Actually, they aren\'t. And, driven class D which adds some
\"switching variation\" to the timing.

> In theory, that could be done continuously at inaudible levels.

Hmmm... I\'m not sure that would buy me anything. I\'d need to
have measurement kit in place \"all the time\" that could \"hear\"
(or otherwise observe) the signals emitted from other nodes...?

Originally, when designing/demonstrating the clock synchronization
algorithm, I placed a dozen \"nodes\" on a bench, each tethered to
the switch via a \"random\" amount of CAT5e. I modified the code
to generate an \"impulse\" on a digital I/O -- on each node -- that
represented its idea of when the clock was occurring.

I cabled these to 12 channels of a logic analyzer and arbitrarily
triggered off of one to observe the others in relationship to it.

As measuring EACH delay -- and the jitter it exhibited -- would be
tedious (the analyzer won\'t automate that process), I tweeked
the code to generate a pulse of width X, instead of just an impulse.
I adjusted for the computed instantaneous CPU clock frequency
(cuz there\'s no guarantee that they are identical).

Setting X to a large value, I reconfigured the analyzer to trigger on
\"all inputs hi\". This will only occur if all of the pulses overlap,
in time. So, a successful trigger means the clocks are synchronized to
within X of each other. (You can let this run \"forever\" to see if
they ever exceed this value)

Then, I reduced X until I couldn\'t reliably get a trigger.

The advantage of that approach was that there\'s no hand-waving going
on; folks can see that the pulses each appear to be \"about X width\".
And, can understand that a trigger means \"at least ALL partially
overlapping\".

But, I have since deployed the nodes. They are now 50+ feet apart
over a few thousand square feet. \"Wire distance\" between them
(if I have to connect something to each and fish it around walls,
through doors, etc.) is considerably longer.

Rather than using a similar DIGITAL \"impulse\" technique (running a
low level signal through gobs of wire). So, I decided to use the
amplifier as a buffer -- gain and lower impedance drive. But,
a narrow pulse won\'t make it through the amplifier with the same
clean edges that a digital I/O can be emitted.

So, synthesize an audio signal that has a nice recognizable shape
and let it travel down the wire to the DSO (add identical lengths
of wire to each node to eliminate transit delay differences)

[DSO can only do a few channels at a time so this means setting up
an experiment for nodes 1 & 2, then 1 & 3, then 1 & 4, etc. It
takes a lot longer to demonstrate this -- and more patience on the
part of the observers.]

 

[Jeff Liebermann]

On Thu, 24 Sep 2020 11:29:44 -0700, Don Y
<blockedofcourse@foo.invalid> wrote:

In my case, the individual \"nodes\" are located over reasonably large
distances (50 ft) and not \"line-of-hearing\" (concocted term as a play
on \"line-of-sight\").

Depending on the size of the loudspeakers and SPL (sound pressure
level) generated, 50 ft seems like it would work with any commodity
electret microphone. This is what HK EZset mic looks like:
<https://www.google.com/search?q=harman+kardon+EZset+microphone&tbm=isch>
If 50 ft is too far, or a speaker is hiding behind some obstruction,
just add some more cable to the microphone and move it around as
needed. Since sound travels at about 1 ft/millisec through the air,
and about 0.7 ft/nanosecond through the audio cable, you can ignore
the mic cable delay.

I also want to be able to numerically demonstrate the individual delays
(hence the DSO time-domain approach... no \"magic\" involved in convincing
yourself that a SPECIFIC delay has been achieved)

(later, I\'ll need to do the same with video; and the audio wrt the video)

What you\'re doing kinda sounds like Art-Net, which uses DMX to control
the lighting. I suppose it could be adapted to control the delays at
each loudspeaker:
<https://art-net.org.uk>
<https://art-net.org.uk/structure/time-keeping-triggering/arttimecode/>
<https://art-net.org.uk/resources/dmx-workshop/>
<https://www.google.com/search?hl=en&tbm=isch&q=art-net>
Other than some light reading and random Googling, I know next to
nothing about Art-Net technology. As before, it might be worth
looking into what it does and what\'s available.


--
Jeff Liebermann jeffl@cruzio.com
150 Felker St #D http://www.LearnByDestroying.com
Santa Cruz CA 95060 http://802.11junk.com
Skype: JeffLiebermann AE6KS 831-336-2558

 

[Don Y]

On 9/24/2020 2:10 PM, Jeff Liebermann wrote:

On Thu, 24 Sep 2020 11:29:44 -0700, Don Y
blockedofcourse@foo.invalid> wrote:

In my case, the individual \"nodes\" are located over reasonably large
distances (50 ft) and not \"line-of-hearing\" (concocted term as a play
on \"line-of-sight\").

Depending on the size of the loudspeakers and SPL (sound pressure
level) generated, 50 ft seems like it would work with any commodity
electret microphone. This is what HK EZset mic looks like:
https://www.google.com/search?q=harman+kardon+EZset+microphone&tbm=isch
If 50 ft is too far, or a speaker is hiding behind some obstruction,
just add some more cable to the microphone and move it around as
needed. Since sound travels at about 1 ft/millisec through the air,
and about 0.7 ft/nanosecond through the audio cable, you can ignore
the mic cable delay.

It would be \"challenging\" to try to handle all of the emitters in
this way. I\'d have to concentrate on groups that are \"within earshot\"
of each other. Then, move onto the next (ovelapping!) group, etc.

I suppose I could drive a fixed signal from each and tweek the delays
until the signal was maximized at the microphone\'s location. Then,
physically measure it\'s location, etc. And, work backwards from that
to get to the values that I seek.

I also want to be able to numerically demonstrate the individual delays
(hence the DSO time-domain approach... no \"magic\" involved in convincing
yourself that a SPECIFIC delay has been achieved)

(later, I\'ll need to do the same with video; and the audio wrt the video)

What you\'re doing kinda sounds like Art-Net, which uses DMX to control
the lighting. I suppose it could be adapted to control the delays at
each loudspeaker:
https://art-net.org.uk
https://art-net.org.uk/structure/time-keeping-triggering/arttimecode/
https://art-net.org.uk/resources/dmx-workshop/
https://www.google.com/search?hl=en&tbm=isch&q=art-net
Other than some light reading and random Googling, I know next to
nothing about Art-Net technology. As before, it might be worth
looking into what it does and what\'s available.

I already have the ability to shift the emitted signals in time. What
I need is a way to show (graphically, numerically, acoustically)
the temporal differences between these emitted signals.

The demo I\'m building uses the user\'s present location to determine
which emitters should be powered up, which powered down, and how to
adjust the delay and SPL of each driver to \"accompany\" the user
as he walks around, relative to their locations.

For example, if you\'re listening to emitter A producing some particular
content -- but walking towards emitter B -- I want to bring B on-line
at a level and delay that \"walks alongside you\" as you leave A\'s influence
and approach B\'s.

I can\'t just make B \"really loud\" (to compensate for the increased distance
relative to A) because there will also be a transport delay (sound through
air) for the wavefront to reach the user. Depending on the source material,
any delay exceeding the echo threshold will want to be perceived as a
second sound.

Instead, I want to adjust the level AND delay so that you perceive the sound
as coming from A more rapidly than if you just approached it as it grew
louder.

[Think about what it sounds like when you walk past one speaker and towards
another; you feel like you are PASSING one sound source and APPROACHING
another. I want the sound source to be perceived as ACCOMPANYING you.]

By reducing the delay at B, I can ensure its wavefront reaches you before
that of A (Haas\' effect). And, let B\'s output dominate your approach
as B is attenuated by your passing AND my deliberate attempts to
minimize it\'s perceptual influence. As you approach B, it\'s SPL is reduced
so it\'s comparable to what A was like when you were in its proximity.

[of course, there is some doppler involved but I think, at walking speed,
you\'d not hear the pitch changes -- even though I\'m manipulating those
pitches by dynamically twiddling with delays]

In certain areas, I can surround the user (demonstrate-ee?) and interactively
twiddle with the delays to draw attention to individual emitters without
altering SPLs. But, that\'s a less impressive demo.

 

[Martin Brown]

On 24/09/2020 17:41, Don Y wrote:

On 9/24/2020 8:29 AM, Martin Brown wrote:
On 24/09/2020 15:51, Don Y wrote:
I have an application that distributes audio among multiple, distant
nodes (emitters/speakers) over an ethernet.  I control the relative
timing (phasing) of the signals from each emitter -- either in
synchronism
or leading/lagging by specific dynamic amounts (which can be on the
order of many ms).

To date, I\'ve been measuring this skew in the time domain by emitting
a plosive burst whose start time can readily be identified (on a \'scope)
relative to the time it appears from another emitter.  This is the
direct analog of a \"clap board\" in a film production.

Is there a cleverer way I can do this -- likely in the frequency
domain -- by deploying a signal of particular \"characteristics\"
that would let me \"miss\" the start time of the emission and still
yield the same results?  (it\'s just annoyingly inconvenient to
have to set up each experiment with an introductory delay just so
I can be prepared for the output(s).  A \"steady state\" sort of
approach would be nicer -- set up a stimulus, then probe each
emitter to check results for that emitter!)

Chirp might well do it. Same technology as bats use but with a fairly
high computational load reducing the data.

You can even modulate each emitter with a characteristic start
frequency, sweep rate and duration to be able to tell them apart.

The other choice might be periodic Walsh functions on a length 2^N
which are an orthogonal basis set on binary sequences.

https://mathworld.wolfram.com/WalshFunction.html

IOW You can uniquely determine the delay from each emitter in one go
if you can sample the resulting waveform fast enough.

Can I get an *instrument* to do this for me?  E.g., I\'ve been using a DSO
and just letting it measure the delay between two signals.

I would not have thought so - although I could be wrong. What I have
described is similar to some of the tricks used in phased array radar.

You could probably roll your own with a soundcard and some software.

--
Regards,
Martin Brown

 

[Don Y]

On 9/25/2020 1:28 AM, Martin Brown wrote:

On 24/09/2020 17:41, Don Y wrote:
On 9/24/2020 8:29 AM, Martin Brown wrote:
On 24/09/2020 15:51, Don Y wrote:
I have an application that distributes audio among multiple, distant
nodes (emitters/speakers) over an ethernet. I control the relative
timing (phasing) of the signals from each emitter -- either in synchronism
or leading/lagging by specific dynamic amounts (which can be on the
order of many ms).

To date, I\'ve been measuring this skew in the time domain by emitting
a plosive burst whose start time can readily be identified (on a \'scope)
relative to the time it appears from another emitter. This is the
direct analog of a \"clap board\" in a film production.

Chirp might well do it. Same technology as bats use but with a fairly high
computational load reducing the data.

You can even modulate each emitter with a characteristic start frequency,
sweep rate and duration to be able to tell them apart.

The other choice might be periodic Walsh functions on a length 2^N which are
an orthogonal basis set on binary sequences.

https://mathworld.wolfram.com/WalshFunction.html

IOW You can uniquely determine the delay from each emitter in one go if you
can sample the resulting waveform fast enough.

Can I get an *instrument* to do this for me? E.g., I\'ve been using a DSO
and just letting it measure the delay between two signals.

I would not have thought so - although I could be wrong. What I have described
is similar to some of the tricks used in phased array radar.

<frown> What I need is something like a LORAN receiver! :<

> You could probably roll your own with a soundcard and some software.

I was hoping to avoid \"yet another development project\". And, more
significantly, forcing folks to rely on my implementation of the
tool that validates my \"claims\".

E.g., by emitting plosive bursts, I can let an observer watch the traces on
a DSO and simultaneously *listen* to the pops coming from the speakers.
Set the relative delay to a second or so and the sweep to half a second per
division. So, they can hear the pop for the first speaker and simulataneously
see a blip on the top trace of the DSO. Then, a second later, hear the pop
from the second speaker and simultaneously see the blip on the *bottom*
trace. I.e., they can grok that the DSO is showing the events that they
are hearing in a visual manner.

Then, tweek the delay down to half a second and have them see the blips
come closer together. Adjust the sweep to 0.1 sec per division and they
can internalize the fact that I\'m still showing them the same events...
just with time magnified.

[Of course, I can design an instrument that does this. But, then they
have to take it on faith that my instrument is actually performing
the analysis depicted (no sleight of hand). The difference between
telling the auditing accountant that your inventory is full of ready-to-sell
product and PROVING that it is!]

In this way, I can keep shortening the delay and expanding the sweep
until the blips hit the echo threshold. Then, beyond to almost coincide.

So, the \"follow me\" demo can be experienced aurally as well as visually;
they can \"see\" what is happening to explain what they are hearing.

For a more technical audience, they can see the actual timescales involved
to create the effects cuz the idea of moving the plosive pops in time
is something they will readily grasp but they might not be aware of the
psychoacoustics involved in their perception. I.e., by adjusting the
timing of a particular wavefront so that it arrives at the listener\'s
location BEFORE the wavefront from a closer -- and LOUDER! -- source,
that I can trick the brain into thinking those almost coincident sounds
are actually coming from the farther -- and less intense -- emitter.

[Of course, they\'ll likely want to tweek the delays themselves to get a better
feel for where the effect manifest and fades]

 

[Joe Gwinn]

On Thu, 24 Sep 2020 11:21:32 -0700, Don Y
<blockedofcourse@foo.invalid> wrote:

On 9/24/2020 9:53 AM, Chris wrote:
On 09/24/20 15:51, Don Y wrote:
If yu are using ethernet, transit time across the network can depend
on network load, retries etc and is not guaranteed. For that sort of
work, a private subnet or a different protocol may be more appropriate.

Have you modelled worst case transit times ?...

Of course! Every device on the switch is of my own design -- I know
exactly what it sends/receives and when. The switch participates in the
clock synchronization algorithm (similar to PTP).

So, I distribute the \"content\" slightly ahead of time. Then, tell
each node when to emit it, relative to that synchronized \"clock\". The
issue then distills to ensuring I can keep ahead of the \"consumers\"
(cuz I can\'t deliver a complete \"audio snipet\")

How accurately must the delay be measured? As a practical matter, it
will be say one eighth of the wavelength at about 3 KHz, the peak of
ear sensitivity. In any event, it cannot be better than a wavelength
at the highest frequency that will pass through the speakers, the mic
usually being far wider band.

My first thought was to use a multi-sine test waveform, which works
well for such things as determining the electrical length of a long
transmission line, but ...

Practical rooms have a lot of reverb and multi-path, so I\'d hazard
that only wideband signals with unambiguous correlation peaks are
going to work. One can limit the time spent searching for peaks in
delay space by a coarse-fine strategy, starting with band-limited
signals.

Joe Gwinn

 

[Don Y]

On 9/25/2020 9:51 AM, Joe Gwinn wrote:

On Thu, 24 Sep 2020 11:21:32 -0700, Don Y
blockedofcourse@foo.invalid> wrote:

On 9/24/2020 9:53 AM, Chris wrote:
On 09/24/20 15:51, Don Y wrote:
If yu are using ethernet, transit time across the network can depend
on network load, retries etc and is not guaranteed. For that sort of
work, a private subnet or a different protocol may be more appropriate.

Have you modelled worst case transit times ?...

Of course! Every device on the switch is of my own design -- I know
exactly what it sends/receives and when. The switch participates in the
clock synchronization algorithm (similar to PTP).

So, I distribute the \"content\" slightly ahead of time. Then, tell
each node when to emit it, relative to that synchronized \"clock\". The
issue then distills to ensuring I can keep ahead of the \"consumers\"
(cuz I can\'t deliver a complete \"audio snipet\")

How accurately must the delay be measured? As a practical matter, it
will be say one eighth of the wavelength at about 3 KHz, the peak of
ear sensitivity. In any event, it cannot be better than a wavelength
at the highest frequency that will pass through the speakers, the mic
usually being far wider band.

That\'s a tricky question. Ideally, I\'d like to use whatever \"scheme\"
to also measure the skew between \"system clocks\" (described in my
response to Larkin\'s reply) -- because \"deinstalling\" devices to
measure the skew as I had done originally is not an option. In
that case, I was seeing skews on the order of 1us. That will improve
tenfold in the next version -- so, ~100ns. (this is why I would
like to make the measurements electrically and not acoustically)

[ultra fine-grained synchronization is important so that \"stuff\"
happening -- sensed or actuated -- on node X can be temporally ordered
wrt \"stuff\" happening on node Y]

My first thought was to use a multi-sine test waveform, which works
well for such things as determining the electrical length of a long
transmission line, but ...

Practical rooms have a lot of reverb and multi-path, so I\'d hazard
that only wideband signals with unambiguous correlation peaks are
going to work. One can limit the time spent searching for peaks in
delay space by a coarse-fine strategy, starting with band-limited
signals.

Yeah, I think going with an acoustic solution just replaces one
(relatively simple) problem with another (more complex) one.

My original approach may prove to be the simplest and most realistic;
though I may have to replace the \"audio amp\" with a more general
\"buffer\". Or, perhaps a laser module that I can arrange to pulse
and route down a potentially long piece of fiber to a remote
measurement station to avoid dealing with long transmission lines.

[That would also have a decent chance of scaling to even longer
\"node separations\" than I\'m presently addressing]

Hmmm... let me look into what\'s inside these SFPs and see if I can
repurpose them for the task...

 

[Don Y]

On 9/25/2020 3:11 PM, Don Y wrote:

My original approach may prove to be the simplest and most realistic;
though I may have to replace the \"audio amp\" with a more general
\"buffer\". Or, perhaps a laser module that I can arrange to pulse
and route down a potentially long piece of fiber to a remote
measurement station to avoid dealing with long transmission lines.

[That would also have a decent chance of scaling to even longer
\"node separations\" than I\'m presently addressing]

!!!

Hmmm... let me look into what\'s inside these SFPs and see if I can
repurpose them for the task...

Yes, this is the way to do it! I\'ll see if I can free up
a pair of GPIOs on each module -- PPSout & PPSin. Then, I
can use some bolt-on module (like an SFP + drive *or*
even a GPS!) to allow each pair of nodes to calibrate their
own clock synchronization algorithms.