Hi,
I came across this test when a member in a different forum (Pianoteq)
mentioned it. They also claimed to be able to pick out the 24bit samples
without guessing with 100% certainty, and this made me curious. So I
acquired the test data and had a look. After doing my own (failed)
listening test, I could however not just disregard the claim of the
other forum member, as they absolutely ruled out any guessing was
involved. So... I looked closer.
To cut a long story short, I found certain flaws in the test data,
likely due to a problem in processing. Because of this, I would not be
surprised if a statistically significant number of listeners were able
to hear differences between the files. For example, I can discern very
clearly between the files when looking at their spectra even after
downsampling to 44/16. And this is something I should not be able to
do.
You can try this for yourself. Take the Vivaldi sample (the real 24bit
one), and export it to 16bit using Audacity. Be sure to use the
high-quality settings and e.g. the triangle dither. Then re-import it
into Audacity. Now pick e.g. a two-second window and run a FFT analysis.
You can see the differences directly; however, the dithering noise is
very low and only really visible in the high frequency range (>20kHz),
where the original file has virtually no spectral power. This is to be
expected.
Now comes the trick. Convert both samples (the original and the 16bit
conversion) to 44.1kHz and look at them again. When I do this here
(since Audacity -does- the conversion correctly, at least in its
highest-quality mode), all differences between the spectra more or less
vanish completely, especially when looking at frequencies of 16kHz and
below; this is also to be expected. Thus when I do these steps in
Audacity, the dithering indeed only affects the highest frequencies,
since they are the weakest.
Soo... and now we have a look at that 'B' sample. Again choose an
analysis window, e.g. 1:23 to 1:25. Take care to shift it 1ms to the
left to compensate for the 1ms cut as documented. Look at it in 96kHz.
Already -lots- of differences between A and B, even in low frequencies
and at relatively high levels. Convert it down to 44.1kHz. Be amazed at
how different the spectra still look. The differences are subtle, but
they -are- present, and as I mentioned they cover the whole spectral
band, including the lower frequencies. Although -my- ears might be
rubbish, I am absolutely not surprised that someone with good hearing
can hear those differences.
The problem I have with this is that when I accept the sampling theorem,
which I do (both as a physicist and a programmer), then any
'improvement' a 96/24 or 192/24 recording offers has to happen either at
very low levels or at high frequencies, since -everything- from -96 to
0dB and from 0 to 22kHz is encoded losslessly. So it is indeed a very
important baseline test that both the original 24bit and the dithered
files in a comparison test yield a (virtually) identical result when
converting down to 44/16. Audacity manages this quite well. The used
version of Adobe Audition does not, it seems.
------------------------------------------------------------------------
kalessin's Profile: http://forums.slimdevices.com/member.php?userid=63059
View this thread: http://forums.slimdevices.com/showthread.php?t=101742