Thread: discussion reliability DAEquality test

Questions has arisen about the quality/reliability of the DAEquality test.

The test is clearly able to show which drive has a better error-correction mechanism. But how does the test show anything about error-conceilment techniques, like interpolation and muting.

As you know some drives are very good in interpolation to hide an encountered error while another drive might not be able to correct or conceil the error. For the first drive the error might not be audible while for the other drive a 'klick' might be audible.

For example, Plextor is very good in ripping CDS200 (audio protection with intentional incorrectable C2 errors). Apparently this is because Plextor can interpolate the errors very good while other drives with a less good interpolation techniques choke on the disc and have audible 'pops' and 'klicks'.

In what way is the behaviour like described above taken into account for in the testing procedure?

Hopefully somebody can give me some insights!

For this the dB(A) graph can tell you about error hiding... It is the loudness for the signals which couldn't have been corrected (YAxis). The error correction capabilities are the ones how late the first non-100% correctable errors occur... (XAxis)

cu, Andre

Thanx for your answer.

Would an audible click/pop be visual as a large spike in the dB(A) graph?

Can the effect of interpolation or muting be recognized in the graph?

Can muting result in a click? Maybe if the number of muted samples is high enough?

The number errors measured in Loudness is the number of FFT analysis done. For that the average and maximum value tells the power difference of the two signals. No accoustic model is used for this calculation, so the audible error may differ from these values.

'No accoustic model is used'. Is it possible that when an acoustic model is used that the output of several graphs would be completely different? (That is: would it be possible that my Plextor shows good results when an acoustic model is taken into consideration?)

Yes, pops and clicks are visible. Depending on the loudness of the surrounding music, it must be higher or lower... E.g. on a -6 dB music position, an error of -40 dB is probably not audible. But on a position where the music is only -18 dB, -40 dB will be quite audible.
The errors themselves are either intepolation or muting. So these are the effects in the graph. Muting itself can create differences up to the music level...
Yes, theoretically muting can result in a click, depending on the algorithm used.
The click does not occur by the length of the mute, but by a quick change of sound pressure (high frequency). So this usually happens at the beginning or the end of a mute.
Yes, using a psychoaccoustic model, it is possible that the graphs are completly different and that some drives would be better than another one, even it is now told the other way... (but not that likely)

cu, Andre

Originally posted by Andre Wiethoff
The errors themselves are either intepolation or muting.

That's not possible : without the help of C2, EAC can only detect errors that are not interpolated nor muted, and it detects most of them.

????

Please give an example of an error that is _not_ interpolated or muted!? Of course you can say that the value is just random, but anyway this is a kind of interpolation/replacement... (Not a good one though)...

The basic of EACs secure mode without C2 is that on errors the probabilty of twice an extraction is the same is low (having twice exactly the same errors)..


cu, Andre

Originally posted by Pio2001
That's not possible : without the help of C2, EAC can only detect errors that are not interpolated nor muted, and it detects most of them.

I always thought that the concept behind non-C2 secure mode was that errors were not repeatable, something that my testing has unfortunately proven to not be true in all cases, that is, I have shown repeatable error patterns are possible. Nevertheless, if you are relying on non repeatable errors, then interpolation would not effect it that much.

True, EAC would not be able to detect that a specific errored sample had different bit errors on different reads but it would still be able to detect that there were differences based on which samples were interpolated in one read but not the next.

The Ricoh interpolates errors yet it is quite capable of detecting errors in non C2 secure mode. When I compare rips that are different using EAC, I am usually able to tell which waveform is the more accurate by zooming in on the leading edge of the differences. Almost invariably, one of the waveforms shows a linear interpolation in the first sample.

As for muting, I would suspect that damage severe enough to cause muting would have been detected by EAC well before muting due to the massive error rate in surrounding samples.

Regards,
Bob

( technical background for this discussion = http://www.ee.washington.edu/consele...udio2/95x7.htm )

I always though that there was no interpolation nor muting, and that the noise that was decoded as a symbol (thus leading to error detection) was outputted raw.

And this idea was backuped by the fact that my Yamaha standalone player always plays perfectly without any audible clicks some CDRs that are so damaged that CD ROM drives are muting big parts of the waveform.

My undestanding was that muting always leading to a null sample, reading twice would output two null samples, therefore no error detected, and that interpolation occuring necessarily between two valid samples, and valid samples being error free, interpolating twice between the same samples would lead to twice the same output, so no error detected either.

Originally posted by Andre Wiethoff
The basic of EACs secure mode without C2 is that on errors the probabilty of twice an extraction is the same is low (having twice exactly the same errors)...

I never though about it.
But after all why not...
So instead of saying "an error never returns twice the same data", we say, "any error always returns the same data, but they never occur twice at the same place" right ?

If we look at how the data is descrambled :



Having a C2 error means having at least 4 or 5 bad symbols in a C2 block, so reading problems must run for 13 or 17 CD blocks at least.
But having isolated clicks in the digital output means to have a bad C2 block and a good one two blocks later. Since the data two blocks later comes from CD blocks in the same aera as the ones of the bad C2 block (the blank ones between, in the picture), we can assume that the aera is not completely damaged, but that the eror rate is just becoming rather high in this aera, with random read errors here and there, leading from time to time to 4 or 5 bad symbols in a C2 block.

EAC checks for differences in 27 sectors each time, thus 27*588/6=2646 blocks at once.

This seems plausible. It must be efficient even if an error always return the same data (interpolation or muting)

This way, is the secure mode more secure than the paranoid mode ? The probability to get twice the same errors in a sector (98 blocks) must be higher than in 2646 blocks, isn't it ?

I'm running a test with a special CD, that should settle the interpolation question for my drives, I've still to record twice or thrice the whole CD realtime, then edit the wav, then run the analyse.exe on it. Interpolation should drastically reduce the difference level on this CD, that is a old dying CDR. Leave me some days to go on

My undestanding was that muting always leading to a null sample

No, in fact usually the drives repeat the last valid sample (otherwise there would be always a click, as there would be a change of sound pressure)...
When valid samples occur afterwards, usually a linear interpolation is done from that sample to the valid sample...

Scratches and other errors are usually optical problems (not artificial CIRC errors like copy protections). It is very easy that the optics read one time a '1' and the other time a '0' on a very defective part. It is then the error correction doint something from there...

This way, is the secure mode more secure than the paranoid mode ?

Yes and no. It works better for also recovering errors 1-2 blocks before the actual error, but this is usually also done in the paranoid mode. So both nearly have the same probability to find errors (as the events are independent)...

cu, Andre

Originally posted by Andre Wiethoff
No, in fact usually the drives repeat the last valid sample (otherwise there would be always a click, as there would be a change of sound pressure)...
When valid samples occur afterwards, usually a linear interpolation is done from that sample to the valid sample...

To clarify, there are two 'standard' patterns for interpolation:
Linear, and
hold.

For linear, the assumption is that there is one errored sample with both the preceeding and following samples being correct. In this case the errored value is replaced by a linear interpolation, that is, the value of the preceeeding sample added to following sample divided by 2 (and rounded down if memory serves). This gives the characteristic straight line of three samples.

In this example the samples are 12, 34, 45, 08, 57, 63 where the sample of value 08 is bad, all other samples are good. The interpolated output will be 12, 34, 45, 51, 57, 63.

The hold interpolation is where there are 2 or more consecutive errored samples. In this case the last good value is held in subsequent samples up till and including the second last bad sample, the value for the last bad sample is the linear interpolation between the held value and first good value.

In the following example the sample values are 12, 14 ,16,34,76, 01, 20, 22, where the values 34, 76 and 01 are bad samples. In this case the interpolated output would be
12, 14, 16, 16, 16, 18, 20, 22.

Actually the linear interpolation is a just a variation of the hold but it is easier to treat them separately. Phillips tried very hard to keep everything to a linear interpolation (much of the interleaving is arranged to increase the linear interpolation region as much as possible - to 48 frames +2symbols to be precise) as the hold interpolation produces considerable high frequency hash ( a sort of zinging sound). Most good high fidelity players actually reinterpolate hold interpolations using much more complex schemes to reduce the hash. CDROMs appear to stick to the simple hold scheme so they are much noisier on damaged disks.

Muting occurs where the error rate gets so high that interpolation is considered not good enough. Exactly what this threshold is I have no idea but I expect it is up to the manufacturer to decide. Mutes are usually measures in 10ths of a second so they are massive. Also, to stop click caused by dropping the output from whjere it is to zero, the players implement an interpolation shceme. In early Phillips players, the interpolation to a mute was 5 frames or 30 samples fro nt elast good value to zero. Although my Ricoh does interpolate data at the ATAPI interface it does not seem to mute - most strange.


One interesting result is to look at what happens with linear interpolation using non C2 detection. A simple example with a couple of big assumptions. Lets have a rip returning a single isolated bad sample. Assumption 1, lets assume that the drive returns the correct value more often than the incorrect ones. for example, the correct value is say 26, for 8 reads, it returns the following values:
26, 28, 45, 42, 26, 78, 26, 41.

Assumption2 - we have two choices, the drive interpolates bad samples or it returns the bad samples unchanged. For illustration based on the surrounding good samples, the interpolated value would be 32.

So if the drive returns interpolated data to the ATAPI interface, the actual output will be:
26, 32, 32, 32, 26, 32, 26, 32.

Assumption3 - let us use a simple majority voting logic to determine which value is correct. If the drive does not interpolate, the results would be 3 votes for 26 and 1 vote for everything else, thus the winner is 26 - the correct value. If the drive interpolates it will be 3 votes for 26 and 5 votes for 32. 32 wins - the wrong value.

Most interesting

If the drive interpolates, you could use the following, if the drive returns differing values for the same sample, determine if it returns only 2 values, if so determine if one and only one value is a linear interpolation, if it is the other value is correct. In reality it is not that easy, there are lots of scenarios that would have to be covered.

Of course with accurate C2 there is no problem, it can be used to validate each returned sample. I assume this is how Plextools gets its results.

Regards,
Bob

@BobHere

I assume you are talking about E12 and E22 errors, because an E32 errors flags the whole frame as erroneous.

Originally posted by Upp3rd0G
I assume you are talking about E12 and E22 errors, because an E32 errors flags the whole frame as erroneous.

At least Plextools use C2 info at the sample level. I don't remember if it's a special function for Plextools.

Originally posted by BobHere
The hold interpolation is where there are 2 or more consecutive errored samples. In this case the last good value is held in subsequent samples up till and including the second last bad sample, the value for the last bad sample is the linear interpolation between the held value and first good value.

Example with the Yamaha CRW3200 EWK reading a damaged version (green) of a CDR (red) :



Zooming into the blue mark :

Originally posted by BobHere
Most good high fidelity players actually reinterpolate hold interpolations using much more complex schemes to reduce the hash.

The Yamaha CDX860 (450 €, 1991) doesn't seem too complex :



Zooming



It is the same as the CD ROM drive, but the linear interpolation runs for 9 samples instead of 2.

I looked for a muted part in my current test files, but I didn't find any.

Originally posted by Upp3rd0G
Assumption3 - let us use a simple majority voting logic to determine which value is correct. If the drive does not interpolate, the results would be 3 votes for 26 and 1 vote for everything else, thus the winner is 26 - the correct value. If the drive interpolates it will be 3 votes for 26 and 5 votes for 32. 32 wins - the wrong value.

That's what I wanted to say from the beginning, but I stand corrected.
Andre explained that such a case is very rare. In reality, the comparison are made in 15876 samples (27 sectors). In this range, when an error occur, it is caused by a place difficult to read. There will be likely many wrong samples because of this. The idea is that when the same range is re-read, there will be about the same amount of wrong samples, but not necessarily exactly at the same place. Some that were wrong the first time will be right, and some that were right will be wrong.
Re-reading, EAC will not keep the sample coming more than 8 times, but the 15876 samples coming more than 8 times. This will only occur if the range is error free, exept in the rare case where only one or two samples are wrong.

That's why I think that the more is compared, the most secure is it.
Example : reading the whole track twice and comparing CRCs is a bit more secure that reading once in secure mode, because, everything else staying the same (no C2), more than 27 sectors are compared (the whole track indeed).

PIO ... did you get that waveform overlay in Soundforge by merging to images with a picture editor or how did you do that ?

I'm truly impressed !

That's why I think that the more is compared, the most secure is it.

Theoretically yes, but this is not how EAC works. It reads large blocks only for error detection. Once an error was detected there, it will read smaller blocks around that position (8 tries etc.)
So only the first error occurance of the large blocks is looked at. The probability of an error at that position reading 27 blocks or reading 27*x blocks is the same. Thus detecting the error has still the same probability...
(If I am correct)

cu, Andre