VileRancour wrote:
Ooops 😀 here it is http://www.youtube.com/watch?v=7XaRds4e3eo .
Anyways, there is another funny thing about digital sound. The success of mp3 has more to do with just how flawed our ears are than with the quality of it. I am not speaking about the psychological "eh, 128 kbps is enough" thing, but about the fact that the signal to noise ratio of mp3 is about the same as the SNR of 2-4 bit ADPCM. Similiar to ADPCM, it has a giant dynamic range, yet small signal to noise ratio (ADPCM usually has a DR of 96 dB and SNR of around 24-40 dB while mp3 has a DR of around 125 dB and a SNR of 25 dB). What makes mp3 and other psychoacoustic formats different is that they mask the noise using psychoacoustic methods. The mathematical "bit depth" of 128 kbps mp3 is just under 3 bits and that of 320 kbps mp3 is just over 7 bits. A really bad encoder makes the noise just above the hearing threhold, while a good encoder like LAME masks it so well that it is not just hidden, but literally impossible to hear at high bitrates by human ears.
So yeah, mp3 is not actually "16-bit". The internal data is 32-bit floating point, but the number of bits per sample is anywhere between 1 to 7. Many people think ADPCM is some kind of an atrocity because they see "4-bit" and think "zomg Atari 2600 sound". Not many people know that their 128 kbps "CD quality" mp3s have a bit depth of 2.9 bit/sample.
ADPCM is usually inferior to mp3 because it has no psychoacoustic masking to hide the noise. There are however many exceptions. There are problem samples like the infamous castanets.wv or the beginning of Fatboy Slim's song Kalifornia that can be easily ABXed from the original even at 320 kbps. ADPCM adds a little noise in those cases, but otherwise handles them correctly because it is a purely "mathematic" compression system, while the psychoacoustic model of mp3 or aac just collapses on these problem samples.
EDIT - I forgot to say, those bittdepths are for mono mp3s. In stereo mp3s they are half that.