Topic: Matrix Time Drift -- How Much Is Too Much

[mattmiller]

I'm processing my first two-recorder matrix recording and am following the procedure outlined for Cool Edit Pro.  Namely, I've converted my time scale into samples, noted the sample numbers of an event at the beginning and end of both sources, subtracted them to get the number of samples between events for each source, and calculated the percentage by which I need to shrink the longer source.  The result:  Over approximately an hour, my two sources (HD-P2 MIC, R-09HR SBD) differ by about 1000 samples, or about 0.02 seconds at 48k per second.  The factor by which I need to multiply the longer (HD-P2) source is 0.99999392265.

1.  Is this even worth correcting?  I haven't tried yet, but maybe a factor can't even be defined to that many decimal places.  And, it seems the error in visually lining up the sources is going to dwarf the error in the two clocks.
2.  Is this likely to be consistent on other recordings using these two units -- i.e., do I need do the math every time?  I believe I've read that once you know your factor, you just apply it every time.

[faninor]

I'm no expert but I've played around with mixing my own two-recorder experiments.

1.
1000 samples seems a rather large amount to be off (so is 500 on either end if you line the recordings up near the middle). I would expect some real bad phasing unless you're applying a low pass filter to one source. If you don't correct it somehow it's probably not worth doing a matrix.

I think you're right that Cool Edit Pro doesn't give you nearly that many decimal places though. I've always been curious myself if, when the drift is just 1000 samples over 1 hour, would it be cleaner to use some tool that would simply remove 1000 of the samples evenly spaced throughout the slower source so that they line up without putting the recording through any complicated processing to interpolate new samples. I wonder if such a tool exists?

Also I've found Sony Vegas to be a very quick and easy tool compared to Cool Edit Pro for getting rid of drift between two sources.

2.
In my experience the amount of drift might vary. At minimum confirm that they really are matching up if you don't redo the math.

[page]

I'm processing my first two-recorder matrix recording and am following the procedure outlined for Cool Edit Pro.  Namely, I've converted my time scale into samples, noted the sample numbers of an event at the beginning and end of both sources, subtracted them to get the number of samples between events for each source, and calculated the percentage by which I need to shrink the longer source.  The result:  Over approximately an hour, my two sources (HD-P2 MIC, R-09HR SBD) differ by about 1000 samples, or about 0.02 seconds at 48k per second.  The factor by which I need to multiply the longer (HD-P2) source is 0.99999392265.

1.  Is this even worth correcting?  I haven't tried yet, but maybe a factor can't even be defined to that many decimal places.  And, it seems the error in visually lining up the sources is going to dwarf the error in the two clocks.
2.  Is this likely to be consistent on other recordings using these two units -- i.e., do I need do the math every time?  I believe I've read that once you know your factor, you just apply it every time.

1) Yes. I've recorded 2 hrs on units before with differences that amount to about that and I could hear a difference in drum beats at the end. Started to notice it around 48 or 49 minutes in that something was strange. Skipped to the end and laughed.

2) Yes, those two specific units should always be off by that amount pending something going wrong with the clock. Replace it with the same model but different physical unit and you get a different time, but those two specific units shouldn't change.

I'm no expert but I've played around with mixing my own two-recorder experiments.

1.
1000 samples seems a rather large amount to be off (so is 500 on either end if you line the recordings up near the middle). I would expect some real bad phasing unless you're applying a low pass filter to one source. If you don't correct it somehow it's probably not worth doing a matrix.

I think you're right that Cool Edit Pro doesn't give you nearly that many decimal places though. I've always been curious myself if, when the drift is just 1000 samples over 1 hour, would it be cleaner to use some tool that would simply remove 1000 of the samples evenly spaced throughout the slower source so that they line up without putting the recording through any complicated processing to interpolate new samples. I wonder if such a tool exists?

I think 1000 samples per hour is average to low. I've seen as much as 10k an hr and as little as about 750. Each units clock is different, which is why syncing during the recording is the best practice if thats an option, baring that, doing the speed math is the next best thing. Audacity handles 7 or 8 decimal places if you have trouble getting CEP to work.

Always do the math. I agree, if you don't correct it, it's not worth doing.

(edit for second reply)

[mattmiller]

I'm doing something backwards here.  Or at least the results are backwards.  Here's my math:

Mic source point A = sample #17,835,732
Mic source point B = sample #189,621,177
Span = 171,785,445 samples

SBD source point A = sample #15,387,235
SBD source point B = sample #187,171,636
Span = 171,784,401 samples

Mic source therefore spans 1,044 more samples and needs to shrink to match the SBD source.  171,784,401 divided by 171,785,445 is 0.99999392265.  So I opened the Mic source in Cool Edit Pro and used Effects > Time/Pitch > Stretch, put in 99.999392265 as the "ratio" (which sounds more like a percentage to me, but its default "ratio" is 100 for "no change").  I made sure to check the Time Stretch (Preserve Pitch) mode and left all the other defaults.  The result is that the file has GROWN in size by approximately the number of samples that I was trying to shrink it.  I think the actual number was in the 1100s, and then when I went back and looked at the settings in the Stretch window it had chopped the "ratio" down to just 99.999 instead of the full 9 decimal places.  I'm assuming that's the source of the difference in the magnitude of the stretch, but why did it expand instead of shrink?  And is 99.999 good enough?  If the end result is that I'm off around 100 samples over an hour is that something to worry about?

[page]

I'm doing something backwards here.  Or at least the results are backwards.  Here's my math:

Mic source point A = sample #17,835,732
Mic source point B = sample #189,621,177
Span = 171,785,445 samples

SBD source point A = sample #15,387,235
SBD source point B = sample #187,171,636
Span = 171,784,401 samples

Mic source therefore spans 1,044 more samples and needs to shrink to match the SBD source.  171,784,401 divided by 171,785,445 is 0.99999392265.  So I opened the Mic source in Cool Edit Pro and used Effects > Time/Pitch > Stretch, put in 99.999392265 as the "ratio" (which sounds more like a percentage to me, but its default "ratio" is 100 for "no change").  I made sure to check the Time Stretch (Preserve Pitch) mode and left all the other defaults.  The result is that the file has GROWN in size by approximately the number of samples that I was trying to shrink it.  I think the actual number was in the 1100s, and then when I went back and looked at the settings in the Stretch window it had chopped the "ratio" down to just 99.999 instead of the full 9 decimal places.  I'm assuming that's the source of the difference in the magnitude of the stretch, but why did it expand instead of shrink?  And is 99.999 good enough?  If the end result is that I'm off around 100 samples over an hour is that something to worry about?

Yeah, don't do 99, do 1.0000whatever for shrinking (99 would be used to grow the shorter source). you want to speed it up by X percent, where 1 is no change and anything above that is "faster".

When you do it, listen very very carefully two the sources seperately at the very end, like the last 30 seconds, and then listen to the mixed source. If it's off, you'll know.

[runonce]

Is the preserve pitch setting necessary...?

My impression is that the sources AREN'T truly the same pitch...just close enough to make it hard to hear... at least to the average listener.

The playback clock is the constant...the recording's rate is the variable...

Isnt this situation an extreme example of the "wrong sample rate in the header"...?

Playing a 48 at 96 sounds like chipmunks...playing a 48 at 48.002 might not be very noticeable, until an hour has passed...and you are out of sync...

Making any sense here? (sorry not truly addressing the post)

[Shadow_7]

If the clocks differ, I would think that preserving the pitch would  be counter productive.  Not that the adjustment is large enough to change the pitch (significantly).

My math is a bit different, but I've done it by a lot of trial and error (and minimal math).  By comparing the wave shapes in audacity and zooming in a lot.  My Korg MR-1000 seems to differ between left and right channels.  Or at least the converted results via audiogate.  It's the same device so it always starts in sync at 0 to 0.  Basically 0.0006 seconds per hour difference between channels.  Roughly.  Trial and error says speed 1.00000013 compensates for it mostly to 30 minutes.  A hair high, but short of adding another digit, close enough.  As I wonder what the FPU accuracy is limited to on a 32 bit CPU and with sox? 

It does make a difference, maybe not on short clips aligned on a per clip basis.  But for anything extracted from say 20 minutes into an half hour session and lasting longer than five minutes, fairly significant.  Mainly when mixed.  Which will happen at the point of reproduction if you're not using headphones.  Low stuff will drift out of phase and cancel itself out.  So bass, drums, timpani, tuba, will seem less significant in the mix.  And everything will sound a bit muddy even if the source wasn't.  It doesn't sound like that on headphones, unless you hit the mono button on a headphone preamp.  Higher frequencies seem to be relatively immune for some reason, probably because I didn't take a pure science approach to the math.  And the mic elements were spaced. 

I do wonder if the drift rate drifts depending on battery strength.  But I would assume that it's mostly fixed / device dependent.

[mattmiller]

Yeah, don't do 99, do 1.0000whatever for shrinking (99 would be used to grow the shorter source). you want to speed it up by X percent, where 1 is no change and anything above that is "faster".

That's what I figured.  I was thinking I was compressing the waveform to some some smaller percentage, when in fact I'm "speeding up" the sample frequency, resulting in the desired shrinkage.

Is the preserve pitch setting necessary...?  My impression is that the sources AREN'T truly the same pitch...just close enough to make it hard to hear... at least to the average listener.

If the clocks differ, I would think that preserving the pitch would  be counter productive.  Not that the adjustment is large enough to change the pitch (significantly).

I did this because it was recommended to do so in the tutorial I was following (in another thread).  I think it was describing the process for Audition, which I was able to follow pretty much exactly for my old version of Cool Edit Pro.  The choices are Time Stretch (preserves pitch), Pitch Shift (preserves tempo), and Resample (preserves neither).  So the "Resample" is the appropriate choice here?

[Yane]

My experience with the Cool Edit and later Adobe Audition time stretching algorithm was negative, it simply sounded bad. In a situation like yours, I ended up correcting the drift for each song -- like you, I figured out the relative clock speeds, then I  computed the  offset needed at the middle of each tune and applied it to that segment of the audio, easy to to non-destructively in CE. The amount of drift in a few minutes of audio wasn't noticeable. As to how much drift is acceptable, it's not a problem unless it's a problem: a few milliseconds of offset will result in comb filtering -- try introducing offsets and listen to the results to learn what it sounds like.   I assume you allowed for the offset resulting from the different distances to the microphones?

[ghellquist]

There are som bad items of news here.

-- the ear can hear a difference of a few milliseconds. The effect is a smearing of the stereo image. Running at 44100 samples per second, a few milliseconds might be perhaps 100 samples. 1000 is definitely enough to be a problem in my world. Remember that in a stereo array the two mics may be less than one millisecond apart and still give an acceptable stereo image.  ( 340 meters / second means 1millisecobnd equals 0,34 meters or slightly above a foot for the non-SI people).

-- the drift between two non-synched recorders will have a varying drift over time and between days. Most important aspect is probably temperature.  And this will change over time, as will the drift. And both boxes will most probalby not have exactly the same temperature change around the crystal oscillator where it matters.

// Gunnar

[stevetoney]

All math aside, what people have not mentioned in this post is that once any software correction is done, the question of how much is too much shift gets answered when you listen to the end result.  If the sources aren't matched well and they drift apart enough that you hear a reverb-y like sound in the mixdown, then the drift is too much.  If you don't hear it, then it's OK. 

That said, my experience was that 1 millisecond of difference between the sources was on the upper end of my personal threshold of what was acceptable...2 milliseconds was WAY noticeable.

[page]

the drift between two non-synched recorders will have a varying drift over time and between days. Most important aspect is probably temperature.  And this will change over time, as will the drift. And both boxes will most probalby not have exactly the same temperature change around the crystal oscillator where it matters.

Interesting, thanks for the heads up, didn't know that.

All math aside, what people have not mentioned in this post is that once any software correction is done, the question of how much is too much shift gets answered when you listen to the end result.  If the sources aren't matched well and they drift apart enough that you hear a reverb-y like sound in the mixdown, then the drift is too much.  If you don't hear it, then it's OK. 

Agreed, I noticed it first in a shift in stereo image (due to phase cancelation), and then finally reverb.

[Shadow_7]

In sox it's speed, tempo, and pitch.  speed compresses or expands the whole.  pitch shifts the frequencies.  tempo adds or subtracts content to make it longer or shorter.  Not that I'm up on the particulars of the inner workings.

I already do a 1.00011 speed adjustment to match the recording rate of my camcorder.  So it looks like I might be doing a 1.00011013 left channel and 1.00011003 right channel adjustment on my Korg MR-1000.  I couldn't get close enough by adjusting the speed of only one channel.  And I didn't really want to add a time shift in trimming of the content to compensate.  And it does look like that's the max number of digits that actually have an affect.

It does seem to vary for me.  1.00000011-ish for Tuba Christmas last year.  1.00000013-ish for Memorial day this year.  I noticed it mostly in a non-existent low end.  And others comments that my recordings were dull / nothing special.  I've been researching and drooling over a couple dozen pairs of microphones trying to resolve my low end issue.  Turns out that it's more of a sync / speed issue.  Although there are mics with better low end than my STO-2's.

[junkyardt]

spend your time churning through complicated math to do it in CEP if you insist, but there are other programs that make this process much more intuitive and easy, with no math necessary. i churn out near-flawless matrixes in Sony Vegas in 10-15 minutes or less now that i know how to do it right.

[mattmiller]

spend your time churning through complicated math to do it in CEP if you insist, but there are other programs that make this process much more intuitive and easy, with no math necessary. i churn out near-flawless matrixes in Sony Vegas in 10-15 minutes or less now that i know how to do it right.

The math isn't complicated.  Recording A is x minutes long and recording B is y minutes long.  Adjust one so that it matches the other in length, and then line them up.  Among the questions being raised here is whether the clock differences are constant throughout the recording or from one night's recording to another.  Vegas doesn't address that any better than CEP.

On a related topic that I don't think deserves it's own thread (correct me if I'm wrong)...

How does the addition work with the digital scale?   (I searched but couldn't find anything using my chosen keywords).  I normalized each of my sources to -6 dB.  If I mix them without touching either of them, what will the peak amplitude of the resulting file be (assuming two -6 dB peaks overlap)?  And if I increase one of them to -3 dB and reduce one of them to -8 dB?