At 09:24 28/05/00 +0200, you wrote:
>Normalization consists of determining the output leven of the audio data,
>and then multiplying the audiodata with a certain factor so that the highest
>value becomes 95 or 99% of the absolute maximum (plus or minus 32768).
>The trick is twofold: First of all, determining the output level is not that
>obvious. Many programs just count the highest peak in the data, but one peak
>can thus influence the rest of the data considerably. Other programs
>calculate the maximum and the desired level with more sophisticated
>algorithms (root mean squared, which aproaches the human hearing better than
>a peak scan)
>You could consider an adaptive algorithm, which would divide the data in
>blocks, determine the highest peak in each block, and would then use a
>gradually sliding factor to multiply the samples with.
Hello,
Yes, but if you use RMS you must need compression to avoid clips, and
compression means a reduction in the dynamic range, and less dynamic range
reduces the "quality" but it sounds stronger. Recent CDs are RMS normalized
(Usually around -8 db) and they sound really horrible (IMO). I think that
MP3 coding for a -8 db RMSed song need more bits and is more difficult to
detected masked freqs. Am I wrong ?
I think that kind of normalization should be done in the decompression
process (it means information destruction) like a radio station compressor
does.
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
