Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Yannick Willox]

Hello all,

what I did during comparison of 6 AD convertors (Stagetec Truematch, Digital Audio Denmark 2408R, Soundscape IO896 (Apogee designed), Soundscape Ibox48, the old Soundscape Ibox24 (the only 48K max device in this test)) included a remastering test : I created a 20 min test CD, which was played through a Benchmark DAC1 and into all AD convertors - line level. I did the same with my Gracedesign 801R preamp inserted.

One test track on the CD was just plain pink noise.
This is where it gets worrying : ALL convertors experienced a severe rolloff starting from 10KHz and going down to -8 to -10 dB at 20K. Only one AD convertor had NO rolloff whatsoever on this track.
Listening to the pink noise, this last convertor was also the only one that sounded identical to the original files.
This convertor is the one with 24ch AD and 24chDA, and is about 50 euro per analogue IO ...

[Yannick Willox]

The answer comes from the other ADDA test thread :

Joe Crawford wrote on Sat, 05 March 2005 21:58

I need to spend some time investigating my 1 sample offset between the left and right stereo channels, but I’ll still try to get back and try the digital loop-back sometime this week.   Joe Crawford

The answer to my test is that all the expensive convertors had the very same 1 sample offset (R ch 1 sample late), resulting in the filtering in the L+R spectral analyzer. Switching this to L or R only gave correct results. Also nudging all R tracks of the offending convertors did correct the result.

Dan, what is happening here ? Do all highend AD convertors have a timing problem between L and R channels ?

I need to redo a week of listening to those ADs, because now it seems I have to shift half of the tracks...

Very worrying  

[bobkatz]

Yannick Willox wrote on Sun, 06 March 2005 06:00

This is where it gets worrying : ALL convertors experienced a severe rolloff starting from 10KHz and going down to -8 to -10 dB at 20K. Only one AD convertor had NO rolloff whatsoever on this track.

This sounds very suspicious. I have to question the test. Even the funkiest, worst-made converters are relatively flat, at the worst case, -1 dB at 20 kHz is the worst I've seen. The filters and how they are implemented can cause some strange tiny bumps but these are often exhibited as tiny frequency response ripples throughout the passband, or small rolloffs above 18K, and so on, in a 44.1kHz piece.

Please elaborate on your test procedures and how you reached the conclusion that all but one converter was -8 to -10 dB at 20K!

[Joe Crawford]

Bob (and Yannik)

I’m trying to follow this as well.  As near as I can tell a 1 sample difference between the left & right channels (if it occurs in the D/A side of the test - during normal playback) is like a ¼ inch difference in speaker spacing during that playback.  This, while it will cause comb filtering at the ultra-high frequencies, is probably not even noticeable during mastering.  You can’t position your head that close.

If it occurs during the A/D side of the DA-AD loop-back I don’t see how it could cause a major problem.  I don’t know of a procedure that I use in the studio where I loop-back a track (i.e., DA-AD) without going through some outboard gear (which either de-correlates it or delays it some) and then keep both the original and the loop-back track in the mix.  It’s a hell of a lot easier to just dup the track if you want a copy.

But Yannik, it looks like if you do this (the loop-back), and you have the 1 sample offset, and you then run an FFT on the re-recorded data in A+B, you will get results like what you see in the A+B FFT.  At least, 22.050 kHz should totally cancel.  I just don’t see where it will impact us.

Joe Crawford

[Yannick Willox]

Joe, that is exactly what is happening.

In the meantime I got informed that this is accepted practice in the ADDA world, shifting of 1 sample to and fro is 'unhearable'.

Well, I can hear it. It sounds awful. It makes a system unreliable for someone (me) who shifts some mic capsules 1 or 3 samples to get better stereo (MS matrixes).

A remastered recording through an AD like this, with a strong centre image (say a big orchestra cymbal) sounds REALLY different from the original.

Why would you design a highend AD - that will be used for remastering - that imparts this sonic character, if it is so easy NOT to shift the R channel ?

Why would any highend AD designer, making convertors that have (almost)linear phase up to 20K, impart a 90 degree (!) phase shift between L and R at 11K on their designs ???

I lost faith. I will buy the 50 dollar/ch ADDA convertor, because it sounds better. It isn't buggy.

[zmix]

I use my racks of A/D  D/A converters to insert analog gear across tracks in Logic. I use stereo inserts and mono and I compensate the latency of the insert down to the sample. I have NEVER encountered a disparity between any of my converters, either in stereo or across several.  
This sounds like a driver problem. Contact the software manufacturer. The last thing we want to worry about is our gear intervening.

[Joe Crawford]

Yannik

I still don’t understand why this causes a problem.  I recreated your test.  But, the only way I could duplicate the conditions you described was to:

.  1) DAAD (loopback) a monaural track of white noise out of Nuendo
.  2)  recording it back as a stereo track to get the 1 sample offset
.  3)  export the stereo track
.  4)  read it into Wavelab
.  5)  convert (i.e. render) it back to monaural (by adding the two channels together)
.  6)  ran the FFT analyzer on it.

And, as you described, the frequency spectrum is relative flat up to approximately  7k or 8k Hz where it then drops (inverse exponential?) by about -10 dB at 20k Hz.

In other words, I had to digitally combine (add together) two copies of the same track with one of the copies shifted by 1 sample time.  Could you please explain the procedure you are using and what you are trying to accomplish.  I’m sure someone on the forum has another way of accomplishing the same without all the negative effects.

Joe Crawford

[Yannick Willox]

Zmix, there are no drivers involved.
Standard AES/EBU and tdif interfacing here.

Joe,

I can see the reasoning if someone says it's trivial, like moving a mic 0.7 cm or shifting your head slightly. But what happens if you sum to mono, with material that does not originate from a mic, but eg from an analogue MS matrix, or a finished mix on tape ?

I do not believe there are NO scenarios where this is an issue.

What will happen with a Soundfield mic if their AD convertors have the same issue ? two of the four capsules would be in the wrong virtual position, and their B format becomes a joke.

My Pearl DS60 stereo mic has 4 outputs, front and back capsules are available for mixing separately. Just imagine what happens when front mid goes in ch1, front back goes in ch2, then there is the 1 sample shift, then I mix them both equally to obtain a omni mid mic.
Hocus pocus, I now have a 'truly great' omni mic with the backplate 0.72 cm in front of the front plate. I wonder how that will sound off axis ?

Believe, it IS an issue. Any AD manufacturer thinking it is not, will not convince me. Neither will I buy this (these, or most of the) brand. Does it take weeks of extra development time just to get both channels in sync ?

Why NOT get both channels in sync, even if one believes it is not important ?

I call this bad engineering.

[Ralf Kleemann]

Yannick Willox wrote on Tue, 08 March 2005 10:04

Why NOT get both channels in sync, even if one believes it is not important ?

We had a similar discussion on a mailing list for the Metric Halo converter boxes, http://www.mhlabs.com/, with the general consensus that even if it has no directly audible impact, it is very nice to know about this phenomenon and to be able to work around it. BTW, the guys from Metric Halo were completely transparent about this and provided very helpful advice.

To compensate this, just add a 1 sample delay to the other channel. Logic, for instance, ships with a plug-in called Sample Delay, which also works for stereo tracks with separate R/L settings.
A nice side effect of this is that with the sample delay, you can compensate for microphone positions far more precisely than would be possible with the "sorta like 0.1 millisecond" resolution in Logic's timeline.

The delay also depends on the sampe rate. 1 sample at 44.1 kHz covers 0.023 ms, or 23 microseconds. Translated into distance (speed of sound at sea level, 15

[Yannick Willox]

My consensus is that it can have serious impact.
I also have been using sample delay for about 10 years now.
Our system has no stereo sample delays with separate L and R settings, so I would need to split some channels I do not really want to.

I maintain my view that
a. the 1 sample delay should not be there
b. at least it should be mentioned VERY clearly in the manual, and not be left for the end user to discover.

[Ralf Kleemann]

Yannick Willox wrote on Tue, 08 March 2005 12:00

b. at least it should be mentioned VERY clearly in the manual, and not be left for the end user to discover.

Yes, there is that... Still, I wouldn't panic about this unless I was very clear about the practical implications of this problem (I know that you know).

It's not an easy job to determine which ones of your 8, 16, or 36 channels sport a 1 sample delay, and if you're not absolutely clear about which channels to compensate, it's easy to introduce a 2 sample delay instead...

Best regards, Ralf

[Joe Crawford]

Yannik
I see your point on the sum to mono, but only if the sample delay is in the A/D side of the DAAD loop-back.  I have not yet proven where it’s coming from.  I don’t know whether it’s in the D/A side, the A/D side, or a software bug.

A test I ran yesterday has me wondering about whether it might not be a bug in Nuendo.  I played a 4 kHz stereo sine wave and, using the RME DSP mixer routed it (at 0dBFS) to all 8 D/As on the ADI8 DS.  I then patched the 8 D/As to the 8 A/Ds and recorded back to 4 stereo tracks.  All 8 of the channels/tracks lined up perfectly.

Apparently it is not consistent and may even be configuration dependant.  I guess right now it could just as easily be a scheduling bug in Windows (XP SP1 here) as anything else.

And Ralf, your right.  It's going to be hard (more like impossible) to compensate for it until I figure when/where/how it occurs.  Right now I’m starting to run out of time for more testing but I’ll try to get back to it when I can fit it in.

Joe Crawford

[PookyNMR]

Check out Nika Aldrich's paper on digital fillters.  A single sample delay creates a low pass filter.  That is the nature of this phenomenon.

[Joe Crawford]

Thanks Nathan – I sort of proved that for myself with the testing.  In fact at 44.1kHz it’s a low pass filter with the -3dB point somewhere around 9k or 10k Hz and dropping by about 10db per octave.

I’m just still trying to find out what is causing it.  I have it isolated down to the A/D side of the equation (at least on my system), but I still can’t find out who is causing it (i.e. Nuendo, the ASIO drivers, or even Windows XP).  It looks even worst at 88.2 kHz (described in my post on the other ADDA thread).

Joe Crawford

[danlavry]

Ralf Kleemann wrote on Tue, 08 March 2005 13:16

Yannick Willox wrote on Tue, 08 March 2005 12:00 b. at least it should be mentioned VERY clearly in the manual, and not be left for the end user to discover. Yes, there is that... Still, I wouldn't panic about this unless I was very clear about the practical implications of this problem (I know that you know). It's not an easy job to determine which ones of your 8, 16, or 36 channels sport a 1 sample delay, and if you're not absolutely clear about which channels to compensate, it's easy to introduce a 2 sample delay instead... Best regards, Ralf

Here is a copy (cut and paste) of the first massage of a thead I satrted on this forum about time delays. Adding 2 Ch with  one sample delay effects that mix, especialy the high frequencies. But one does not get 10dB at 10KHz...

Copy of the previous text:

Time delays, when is it real?

Some times delay are important to watch for. Others are of academic interest or of little practical use. Let us examine some cases:

An important, and often overlooked case is when mixing (adding) a signal that appears on more than one track. Perhaps the simplest example is a stereo recording, when some portion of the sound arrives at both the L and R channel. The common practice in stereo is to use a stereo converter with equal delay on both channels. Yet, any additional processing done to one channel but not the other may make the delays unequal.

Of course, the same situation applies to multi channel recording. Not unlike stereo, it is best (from time matching stand point) to use a multi channel AD utilizing a common clock. Mixing AD’s made by different manufacturers is likely to introduce time delays between channels. Again, keeping the portion of the sound (signals) shared by more than one channel at the equal delay is a good idea. The equal delay concept all the way to the mix can prevent problems.

What are the problems?

Say you wish to add 2 simple signals. Both are a equal 1KHz sine wave tone. The expected result is to double the amplitude. But if one tone is delayed by say 500uSec both signals are out of phase and the addition will yield a total cancellation.

Reducing the delay to less than 500usec will cause a partial cancellation. The concept of cancellation or partial cancellation (attenuation) does not require equal amplitude waves, or even equal waves. Such signal attenuation due to time delay happens to the portion of the sound wave that is shared by the channels being added (mixed).

A lot of delay is required to cause attenuation of very low frequency energy. But higher frequencies are much more susceptible to such a mix. For example, a 20KHz signal cycle lasts 50usec. Half a cycle is 25usec, therefore 25usec is a point of maximum attenuation. The same 25usec inter channel delay will have little effect on an 100Hz tone, where a cycle lasts 10000usec.

How good of a time match?

Of course, the answer depends on how much delay is acceptable and at what frequency.
Below is some reference data I computed for those interested:

25usec delay at 1KHz attenuates by -.027dB
25usec delay at 5KHz attenuates by -.688dB
25usec delay at 10KHz attenuates by -.3.01dB
25usec delay at 15KHz attenuates by -8.343dB
25usec delay at 20KHz attenuates completely (no signal)

10usec delay at 1KHz attenuates by -.004dB
10usec delay at 5KHz attenuates by -.108dB
10usec delay at 10KHz attenuates by -.436dB
10usec delay at 15KHz attenuates by -1.002dB
10usec delay at 20KHz attenuates by 1.841dB

5usec delay at 1KHz attenuates by -.001dB
5usec delay at 5KHz attenuates by -.027dB
5usec delay at 10KHz attenuates by -.108dB
5usec delay at 15KHz attenuates by -.243dB
5usec delay at 20KHz attenuates by .436dB

1usec delay at 1KHz attenuates by -.0004dB
1usec delay at 5KHz attenuates by -.0001dB
1usec delay at 10KHz attenuates by -.004dB
1usec delay at 15KHz attenuates by -.009dB
1usec delay at 20KHz attenuates by .017dB

The data above shows is a good indicator for the amount of attenuation when mixing 1KHz, 5KHz, 10KHz 15KHz and 20KHz tone due to some delay (25, 10, 5 or 1usec).

This is one case when delay can make a big difference. Note that I am talking about ELECTRIC SIGNALS DELAY, not acoustic delay of sound in the air. It is difficult (if not impossible) to control the acoustic delay to say 1usec. Yet, keeping the AD conversion and processing delay EQUAL will guard from such cancellation. I AM NOT TALKING ABOUT AN ACOUSTIC ISSUE SUCH AS MIC PLACMENT. I AM TAKING ABOUT AN ELECTRICAL SIGNAL HANDLING ISSUE.

To be continued...

Br
Dan Lavry