Walt,
I do understand your point of view (because you are using MD
extensively and it works really fine for low-frequency frog calls).
But I'm still very sure, that I'm not wrong with my findings.
> think we should all note that Raimund does not own MD and has no
>experience recording with MD. His "tests" of ATRAC did not test
ATRAC,
>but another entirely different compression type. In other words
he's a
>complete novice in what MD does.
It is correct that I first used MP3 for my compression tests. I'm
very grateful to Evert Veldhuis and Dan Dugan who repeated the test
with their own MiniDisk machines. I have added these results to my
site (the ATRAC sample is the third color spectrogram):
http://www.avisoft-saslab.com/compression/compression.htm
It can be seen, that ATRAC behaves even worse than MP3 at the same
compression ratio of 5:1.
>This is a deliberately oversimplified description of what ATRAC
>compression does, to the point of being highly misleading. I will
not
>attempt here to describe all it does, but just attempt a little
>correction. I do not claim to know all the details of it's internal
>operation, I expect no one except the experts at Sony really know
all it
>does and why.
Yes, ATRAC is a very complex system and it is difficult to
understand what's really going on. However, if you have some
knowledge and experience in digital signal processing in general
(which I have), it is possible to understand its principle and the
artifacts that it can cause. As you mentioned earlier, this is not a
scientific seminar. Therefore, I rejected the complex details...
I will not continue with a detailed description of ATRAC (I already
tried that in April). But I should mention, that the bit-depth
reduction takes place in the frequency domain (the time domain data
is first converted into the frequency domain (using DCT, something
like FFT) and the coefficients representing the signal in the
frequency domain are stored at lower resolution (depending of the
available bits) onto the disk...
The bit-depth can be reduced dramatically (down to 4 or even 0
bits). This can be seen, when you convert Jeremy's recording
MDtest2.wav into an 8-bit file and compare it with the DAT
recording. The artifact at 6.3 (the quantization noise) is still
visible at spectrograms generated from that 8 bit file. This means,
that the bit-depth must have been reduced significantly below 8 bits
at that section!
ATRAC is an adaptive system. This is the reason, why it does not
behave the same in subsequent trials. The algorithm cuts the
incoming data into short frames (each a few milliseconds long). So,
even if you fed the same (digital) signal into an ATRAC encoder
several times, you will get different results (caused by varying
phase relations between the waveform and the single frames =96 this
can also be seen in my tests on synthetic signals). If you were able
to synchronize the frames to the waveform, you would get consistent
results in subsequent trials. So, there is no serious argument
against the scientific value of these tests.
>It is well to remember that the analysis software being used does
much
>the same thing as compression, in fact it's really extremely severe
and
>crude compression of the sample then displayed as a visual image.
It
>produces it's own artifacts, some of which Raimond misinterprets as
>being in the sample. It takes time and experience to learn to
interpret
>the fine details of sonograms.
The artifacts on my spectrograms can not be attributed to the
limitations of the spectrogram display! Why is the noise not visible
on the DAT recording? Here is the comparison once again:
DAT:
http://www.avisoft-saslab.com/compression/MDtest2DAT.gif
MD:
http://www.avisoft-saslab.com/compression/MDtest2MD.gif
To my mind, this comparison of natural sounds is not inappropriate
this time. I can recognize the difference by ear too (this would be
another proof in case I were blind).
The differences between various A/D converters or pre-amplifiers can
surely be neglected compared to the effects sometimes introduced by
ATRAC. Believe me or not ;-)
Best regards,
Raimund
________________________________________________________________________
________________________________________________________________________
|
