In article , Meindert
Sprang wrote:
"Jim Lesurf" wrote in message
...
I've now done this and the results can be seen at
http://www.audiomisc.co.uk/HFN/OTTre...d/results.html
My thanks to Keith Howard for kindly agreeing to let me publish his
measured results. Note they were made in 2007 so only covered the
players and DACs he had to hand at that time. But - rather
depressingly - all but one of the nine he tried showed they had
problems coping with the waveform that has peaks at +3dBFS.
FWIW personally I'd love to see all reviews on DACs or player use the
'waveform from hell' I devised for the original article (a link to
that is on the above page) as a test of how they cope - or not! :-)
I have the uncomfortable feeling that whilst reviews continue to
overlook this area, problems will continue to afflict some new designs
without anyone knowing.
This test is flawed.
Depends what you mean by "flawed". cf below...
The sample points on the right represent a sinewave of +3dBFS.
Correct. That is what sampling theory, etc, indicate the samples should
mean when reconstructed if we assume sampling/filtering at ADC was done
using standard-theory time-symmetric sinc-like filters.
If you would sample a sinewave on the tops, you would indeed
get the picture as shown on the left. If you move the sample points to
the positions as shown in the picture on the right, the sample points
would be 3dB down and they would still represent a sinewave of 0dBFS.
Following this reasoning, one could say that if a DAW normalizes by
simply measuring the highest sample and scaling all others accordingly,
that DAW is flawed too.
I tend to agree. However this once again depends on how you choose to
define "flawed". Information theory is quite happy to accept the sample
values as representing a waveform that peaks at 3dBFS. And the critical
point is that some CDs, etc, do carry sample values that will require
excursions 0dBFS.
So I don't think the *test* is "flawed" given that players may often be
presented with series of samples that require peaks above 0dBFS to
reconstruct as part of the music. Saying the *discs* are "flawed" might be
a better argument. But in theory they should be fine *provided* the
players/DAC designer realised the implications of sampling theory, etc, and
made a player/DAC that copes as theory indicates.
It should at least "reconstruct" the whole waveform to be able to
determine the *real* maximum amplitude.
I agree. However the reality is that some Audio CDs, etc, *do* give series
of sample values that reconstruct to have excursions above 0dBFS. Hence the
test exposes that some players may have a problem reconstructing such
waveforms.
The problem them becomes a real one for those who buy CDs/files. Should
their player/DAC cope with such material or not? My view is that it should.
This is for two reasons.
1) Such material exists, so a player/DAC should reconstruct it in an
orderly way in accord with the sampling theorem, etc.
2) It is simple enough these days for the designer to arrange (1) if they
have a clue. So why should they refuse give the existence of such material?
The snag being, how to tell if no-one checks using such a test or examines
the CDs to see if any require this? The test exposes a problem area.
In an ideal world, though, I'd certainly *much* prefer those who release
CDs and files (and stream) to avoid any intersample peaks reaching 0dBFS or
more. Alas, that world isn't the one we live in. :-/
So the question is:
do we know how a DAW measures the maximum amplitude?
Nielsen and Lund have looked at such issues IIRC in AES papers. However I
assume that it would depend on the details of the specific cases and
equipment. The actual max possible peak intersample values will depend on
the bandwidth/shape/etc of the source material being processed as well as
the sample rates, etc. So the complication is that a precise figure will
depend on the material as well as the details of the system used.
Overall, though, I'd expect allowing around 3dB to be generally 'safe' cf
below.
One problem, alas, is the obsession some CD/file 'mastering' people have
with LOUDNESS. To deal with avoiding intersample peak problems that would
have to be tackled, I think. TBH I wonder if some of those 'mastering' have
a clue about any of these questions, or give a hoot... :-/
FWIW if you look at my website and many other places you will find plenty
of examples of commercial CDs 'mastered' with plenty of samples in the
range above -1dBFS.
My experience is that 'classical' types of music avoid these problems
because those who make the CDs take more care and aren't obsessed with
loudness, and perhaps because those involved take care. But pop/rock
material is often LOUD and the assumption seems to be "that is what sells".
Is it documented in the manual/specs? If not, using -3dBFS is always
safe because this is the worst case we could encounter as seen in the
test.
The peak excursions can potentially be much higher than 3dB. e.g. if you
look at
http://www.audiomisc.co.uk/HFN/OverTheTop/OTT.html
you can see where I got over +5dBFS for a deliberately chosen 'waveform
from hell'. 8-]
So if you wish to be *really* careful you'd need to avoid samples going
above -5dBFS. But I would agree that such extreme examples would be
unlikely in real practice with acoustic music. (Not so sure for
'electronic' creations for pop/rock!)
In reality, a piece of audio with thousands of samples would probably
have a few that are nearly on the top of a loudest sinewave,
statistically speaking, so -1dBFS would *probably* do as well as someone
else already mentioned.
FWIW my own opinion is that for material at rates like 48k or 44.1k it is
wisest to ensure no samples go above about -3dBFS. Although more like -1 or
-2 may often be OK - depending on the material. That was why I raised this
issue when people were discussing 'normalisation'. The bottom line is that
simply normalising to the biggest sample being at 0dBFS is unwise.
IIRC If you look at BBC Radio 3 material, they generally try to keep peak
samples below about -4dBFS. (This is even more important for methods that
encode using methods like AAC which are 'lossy' so may alter the size of
peaks.) So BBC R3 caution puts them in between the 3dB value from the
simple sinewave example Keith used, and my 'waveform from hell'.
BTW if anyone wants to experiment with the 'waveform from hell' it is still
available as a zipped wave file from
http://jcgl.orpheusweb.co.uk/temp/WaveFromHell.zip
(Actually a pair of wave files with different sign relationships
left/right.)
Use with care though, not sure how happy some amps or speakers would be
with that waveform. :-)
Slainte,
Jim
--
Electronics
http://www.st-and.ac.uk/~www_pa/Scot...o/electron.htm
Armstrong Audio
http://www.audiomisc.co.uk/Armstrong/armstrong.html
Audio Misc
http://www.audiomisc.co.uk/index.html