What does ''normalize'' really do?

Do you loose quality if you reconvert an mp3 back to Audio.

Those are two seperate questions.

Normalizing is used to bring the highest peak to a certain level, often 0dBfs.

I assume you mean converting an mp3 to wav? The wav will sound exactly like the mp3.

DATA (MP3) > AUDIO CD I meant.


What do you mean by 0dbfs?

The tracks on the audio cd will be identical in quality to the mp3.

0dBfs is 0 dB full scale, or the highest sample a 16/44.1 wav can have without clipping.

The wav will sound exactly like the mp3. In the other words Audio CD sounds exactly like the MP3 file...
But beware transcoding, making a MP3 from that Audio CD!

I assume he meant to normalize to 100% - to full digital scale.

Edit: Sorry, dreamliner77 was faster...

But beware transcoding, making a MP3 from that Audio CD!

Why?


Where could clipping occur?

Because the file would then be encoded twice as MP3, being distorted both times. This may or may not be ok, but is never ideal. Instead, use the origonal source when compressing new files.



If you try and go above the full scale volume in a sample. In 16 bit digital audio, volume is between 0 and 255 (or -128 and 127 if its a signed int). If you try and raise a sample above this value, it will be clipped.

This can happen by accident when decoding a file. MP3s are just approximations, and when they're decoded its possible a sample that was ok when encoded could be decoded at 256 or higher. For this reason many people use Replaygain, a standard for audio gain adjustment that avoids clipping during decodeing of compressed audio.

How would clipping effect the mp3?

http://www.hydrogenaudio.org/index.php/Clipping

Actually 16bit integer audio samples are always stored as signed integers; and values range from -32768 to 32767.

Yes ! 32768 = 2^15, the last bit for the sign (?)
What about the maximum 32767 ? I've already seen audio CDs with maximum samples at 32768 (not -32768). CEP can't create 32768 samples in 16bit.
Be careful when normalizing to exactly 0dBFS in 16bit, the loudest or a few (1, 2) more samples might get clipped. This is not to say that that's audible.

The "sign bit" counts for -32768 in 16 bit twos-complement arithmetic. If you have seen a value of +32768 from 16 bit signed value, it's probably just inverted.

There are other ways of doing signed arithmetic, but they are pretty much obsolete.