Le Numérique Est-il Fidèle ?, Illusion ou réalité ? Options

[wfplb]

Finalement... le numérique n'est-il pas un leurre, une invention des marchands ?

Overload in Signal Conversion

Is There a Future for High-Rez?

[Miss Kiki]

super article, j'ai fait un petit digest des passages les plus parlants :

The loudness war, as we shall see, has now
become so furious that equipment downstream
of mastering is not able to pass or
reproduce the audio material without adding
signicant distortion to it. It appears to be
a contradiction that the professional audio
industry promotes better standards such as
96 kHz sampling, 24 bit resolution, DSD, exotic
dither schemes etc. and the hi- world
is becoming conscious about jitter-rejection
etc. – while the current main delivery format,
16 bit/44.1 kHz, is becoming more and
more abused. Most pop releases now get so
much distortion added during reproduction
that 16 bit resolution is completely overkill.

The distortion problems arise because of
more or less conscious use of clipping in the
digital domain during production and mastering.
Digital processors may have familiar
names like "Limiter" or "Compressor", but
there is no guarantee they will have much
in common with their analog counterparts,
and not even that they obey the fundamental
sampling criteria. Clipping can also be
hidden inside computer processors, DAWs or
digital mixers, making later recovery difcult
or in many cases impossible.

While frequent clipping in the digital domain
can cause listening fatigue, things are
made worse during mastering when clipped
material is normalized to digital full scale.
Such material will later be subject to additional
distortion during data-reduction, sample
rate conversion and digital to analog conversion.

Both narrow-band and broadband (e.g.
clipped) signals can have a theoretical peak
value larger than the actual representation
in the digital format. The simplest example
is a high frequency sine wave. Depending on
frequency and phase the peak value of the
digital sine wave representation may be up
to 3 dB lower than the theoretical, or analog,
peak value. If this headroom is used for
increasing the digital gain, the peak value of
the reconstructed analog signal will be higher
than the DC value representing digital full
scale, 0 dBFS.

One way of avoiding the problem of clipping
is to reduce the overall peak level of the
recordings. As this also reduces the average
level it will probably be considered unacceptable
by many sound producing professionals,
so another solution is called for.
An alternative is to to reduce the digital level
in the playback environment prior to digital
to analog conversion. This would reduce
distortion only in future products, however,
leaving hundreds of millions of playback
devices with potential problems. The
price for this constructional measure is a reduced
signal-to-noise ratio in the D-to-A conversion
circuit.

Using an inter-sample precision digital limiter
with the threshold value set to -0.3 dBFS,
that is, basically just to reduce overshoots,
the distortion of the DACs disappeared almost
completely. The overall loudness of the
signal was not reduced, however, so for many
applications this seems to be the better option.

In conclusion, we have demonstrated that
the problem of overload in digital to analog
conversion, sample rate conversion and
in connection with data-reduction systems is
denitely present in practice.
Measuring and listening to in vivo signals
provide a strong evidence that headroom in
the digital production chain has been exhausted,
and that audio quality currently
may be limited for this reason rather than
a basic lack of resolution or a high enough
sample rate.


Because the audio industry is currently
stuck with non-intelligent level restrictions,
such as counting consecutive samples at full
scale, mastering engineers should start using
oversampled meters and limiters, or at
least normalize against e.g. -2 or -3 dBFS
rather than 0 dBFS. It should also be noted
that some data reduction codecs are even
more sensitive to excessively hot level than
a linear signalpath.