Simon and Matt (and others),

Simon Barnes <[EMAIL PROTECTED]> writes:

> Matt Wall wrote:
>     > there would be a loss of quality, however in that "perfect"
>     > world where = it was all digital and no attenuation or any
>     > other interference at all, = going from your recorder/player
>     > to your other identical md = recorder/player since the ATRAC
>     > systems should be identical and should = attempt to compress
>     > the audio the same in all the same places the = recording's
>     > should not have a generational loss.
> 
> In the digital domain, it IS possible to make a perfect copy, but
> this is not the issue. When the original audio PCM data is ATRAC'd,
> 80% of the data is discarded, so the decompressed audio PCM coming
> out can be quite different to what went in, but will sound quite
> similar. If this data is then ATRAC'd again, the hardware does not
> *know* that it is dealing with an already processed data stream, and
> tries to reproduce the data as if it were an original full-spectrum
> signal, and this is the reason for the generational loss.

Basically yes, but see below.

> It could be possible to design an algorithm which always
> reconsituted a compressed data set the same way, preventing
> generational loss, but my guess is that it would sound worse than
> ATRAC, and most of us would prefer to have better sounding first
> generation copies at the expense of worse multi-generations.

This is a topic that someone has done some research in. Have a look at
Frank Kurth's paper (http://www.minidisc.org/multigen_lowloss.pdf):
"An Audio Codec for Multiple Generations Compression without Loss of
Perceptual Quality".

The abstract:

        We describe a generic audio codec allowing for multiple, i.e.,
        cascaded, lossy compression without loss of perceptual quality
        as compared to the first generation of compressed audio. For
        this sake we transfer encoding information to all subsequent
        codecs in a cascade. The supplemental information is embedded
        in the decoded audio signal without causing degradations. The
        new method is applicable to a wide range of current audio
        codecs as documented by our MPEG-1 implementation.

The idea is that they hide and pass along compression parameters in
the audio that can be picked up again by the next generation
compression stage.  What they pass is: Number of channels, bit rate,
window boundaries, and subband bit allocation and scale factors. They
say this data is inaudible (this data hiding is called
"steganography", it's the same technique used for watermarking).

And, the finding that addresses your comments:
        
        It becomes clear that lossy coding changes the signals
        spectral content in a way that does not allow subsequent
        encoders to perform a suitable psychoacoustic analysis.
        This leads to a degeneration of coding parameters and, finally,
        of overall sound quality. We further illustrate this ...

It's an excellent paper and gives a good explanation of the whole
problem of generational loss. I recommend giving it a shot even if you
think it might be above your head. There are several good diagrams and
plots.

Rick

p.s. this paper is now linked from the "Generational Loss Tests" page:
http://www.minidisc.org/generations.html
-----------------------------------------------------------------
To stop getting this list send a message containing just the word
"unsubscribe" to [EMAIL PROTECTED]

Reply via email to