> > I just want to know how can I normalize, and I didn't ask how about expand
> > LAME to support normalization.
>
> But wouldn't a lame-integrated normalization (i.e. scaling of the sample
> values) improve quality over a seperately done normalization, because the
> intermediate results would be floating-point instead of 16bit (or lesser)
> integer? I think that a simple scaling option in lame would be useful (i.e.
> something like --pre-scale <scalefactor>).
>
> As an alternative, are there any sound formats that use floating-point
> values for samples? If so, maybe lame could be made to be able to read such
> a format. If not, it should be quite trivial to include a raw floating-point
> format for use by custom-written preprocessing tools.
>
> Or am I just too paranoid about normalizing to 16bit integer values?
>
> -- Niklas
I also worry about this, but usually there are much bigger problems
to worry about :-)
Right now, the resampling and the stereo->mono averaging in LAME is
done with short ints because the encoding routines want short int's
for the input. This does result in some loss of precision, so it
might be worth trying to convert "mfbuf" from short int to float. If
the speed impact is negligible, I think all internal processing should
be done in floating point.
Mark
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
