Click to view attachment
-V 2 --vbr-new

Click to view attachment
-V 5 --vbr-new

Click to view attachment
Original wave file

And suddenly, we've got 60dB *theoretical* SnR.... not taking into account audibility. What's the noisefloor in a usual room again?

- Lyx

edit: just did a quick google search - apparently, the noisefloor in a *very* quiet room is around 40-50dB.

Please, the discussion is senseless because it ignores basic psychoacoustic principles.

You do not "hear" SNR's because of masking. You hear SMR (Signal To Masker) ratio's. The worst case SMR that your hearing needs (for real-life signals) is about 20 - 25dB. For many signals, you can't even hear anything wrong with a SMR of less than 10dB.

This is well-established theory that is veryfied by experimentation.

PS. "Removing things you can't hear" is in fact nothing else than the codec "adding noise" by storing things less accurately. It's exactly the same thing. As long as the SMR is maintained this noise is inaudible, and this is how all the codecs work.

Hmm, lets see if i understand this right: so, the allowed SnR is basically adjusted dynamically depending on how much other frequencies in the signal mask the noise? (i asume this is not just done in a "flat" way but instead frequency-dependend, so that on certain frequencies more/less noise may be allowed, depending on masking). Did i get this right?

In sum, S/N Ratio is being adjusted via noise being added, perceptually (like image dithering), according to psychoacoustics (just look up any ATH Curve)
This noise is not randomly generated (or not quite so), but is the difference between the original signal and the sum of sine waves (fourier series or whatever) used to recode the sound signal.

Since math formulas are easier to compress than "bitmap data" (again, in comparison to images, MP3s can be seen as a type of "vector format" in which mathematical coordinates are given, and rebuilding is made afterwards, versus PCM audio being sample by sample, bit-by-bit "bitmap" data)

At least that's what I got out of it.

And, referring to my samples, you can see that the mp3 encoding actually REMOVED noise from it, whereas that other guy's encoding added noise.. His sound also had a much lower volume level to start with.

@Shade
Thanks for the explanation. That makes way more sense.


Well, the mistake i did was probably that i mistook the "frequency-smearing" beginning at around -60dB (something which should be impossible to hear) as noise.

edit: or i did it right, and at that point, inaccuracies come into play in your example.

- Lyx

Although it is frequency dependant, and loudness dependant, those only matter to a small degree.

The most important thing is the structure: tonal or noisy.

Correct, except for the broken analogies: Not like image dithering, not so related to ATH.



Yes (somewhat simplified). You can reconstruct the signal perfectly if you'd keep all sine waves (Fourier theory). If you only store some or store them inaccurately, the errors show up as noise (quantization noise).



No, no, no, no. In fact video compression works nearly the same, except that "which stuff to keep" is different (and things like tonality or SMR and so don't apply). The Fourier maths work just as well for 2D as they do for 1D.