Long lengthy and probably off topic explanation of normalization. Hit
'D' now if you want to get back to discussing music ;-)
On Thu, 2 Nov 2000, Chris Fahey wrote:
quoted 6 lines Kent says:> > Kent says:
> > As far as mis-preparation goes, normalizing to 0dB (full code) is the
> > quickest way to make the encoder screw up.
>
> I don't understand the point of normalizing at all. You're just adding
> another generation or processing to the sound.
Not exactly. When you normalize, you find the peak level in a sound file,
and compute the scaling factor X such that the peak * X == full code. Then
you multiply ever sample in the file by that factor. Mathematically this
is a reversable operation; according to information theory, nothing is lost.
Thus normalization doesn't affect sound quality in any way except to make
it louder.
When you normalize, you use the maximum number of bits to represent your
sound. The quieter a sound is the fewer bits of available resolution are used
to represent it. If you record something at a peak level of -21db, it's
only using 12 bits of resolution. You can easily prove this yourself by
recording something (i.e. your favorite record) first with the gain set so
that the signal peaks below -21db, and then record it so that it's as loud
as possible without going over 0dB. Then normalize the two recordings to
the same level and A/B compare them. The quiet recording will sound noticably
coarser.
The real subjective loudness of a signal, however is not a function of the peak
level in the signal. The more important measurement is the RMS power of
the music, which represents an overall average loudness of the track. You
make a track sound louder by compressing the dynamic range -- pulling down
peaks, and raising the volume of soft passages. Modern CDs in the
'pop' genres (i.e. everything but acoustic classical recordings) are usually
mastered to have an RMS level in the -12 to -15 dB range. That's about
as loud as you can go without making the recording sound really squashed
and nasty.
But to get back to MP3 encoders: They do their best work on compressed,
normalized signals, because a properly compressed and normalized signal on
average uses the maximum number of bits of resolution, as an average
over time. The catch is that you don't want to normalize to full code --
you want to have your signal maximum peak some small amount below -- most
sources recommend some fraction of a dB below full code.
Why is this? I don't know precisely, but I suspect it's because full code
signals present a pathological input to the encoder, which then introduces
distortion and artifacts. I know through practical experience that if you
go to full code normalization, your MP3 files will sound like ass.
---------------------------------------------------------------------
To unsubscribe, e-mail: idm-unsubscribe@hyperreal.org
For additional commands, e-mail: idm-help@hyperreal.org