I have experimented a bit with settings that produce about 96 kbps bitrates and tested them with foo_abx. Previously, with MP3 I would have not used such a low bitrate even on a portable, but now my listening tests with the aoTuV beta v. 4.5 are very promising. Many samples are almost transparent.
However, the transparency is not my topic this time. I have noticed that the Track Peak values change considerably when files are encoded at about 96 kbps. At first I noticed this with aoTuV beta 4.5, but I encoded a few test samples also with the official Vorbis v. 1.1.1 and as a reference with LAME v. 3.97 b1 -V8 --vbr-new. All resulting files show big changes in the Track Peak values.
I used foobar2000 v. 0.8.3 for replay gain scanning. Here are the values with my six diverse test samples:
(http://kotisivu.mtv3.fi/alexb/ha/trackpeak.png)
I highlighted the original lossless files. The used encoder version and settings are in the Album field. The biggest change is with the sample named "Get Your Way". The Track Peak value rises from 0.99 to 1.50. Only the sample named "Stay" shows small changes.
Is this something to be worried about?
Edit: Actually, I didn't know that peak values over 1 (=100%) are even possible.
IMHO there can be two explanations:
1. I doubt you can hear difference of even 1 dB between two LOSELES files. May be encoder "thinks" that way.
2. Since encodings are lossy, one file can be subjectively louder than the other while both of them have the same loudness. So psychoacoustic model tries to compensate it.
As I said the files sound fine. I just don't understand the Track Peak values.
I analyzed the same files with J. River Media Center v.11.0. All displayed Peak Level values are limited to 100% (similar with the Track Peak value 1.0 on foobar), except the values of the "Stay" sample, which shows identical values with foobar (87-92%).
Quantization will introduce some changes in frequency samples. Once converted back to time samples, it is likely that some values have been increased compared to the initial ones (some other are decreased).
This is a quantization error. If some initial values were near full scale, the new values can go over full scale (ie more than 1.0).
The higher the quantization the more likely it is to occur.
Edit: Actually, I didn't know that peak values over 1 (=100%) are even possible.
Depends on the storage format. 8-24 bit LPCM can only store samples within the range [-1 , +1) (-1 inclusive, +1 exclueive). 32 bit floating point is another story. Also MP3, AAC, Vorbis, MP2 and the like are not restricted to the [-1, 1) range.
IMHO there can be two explanations:
1. I doubt you can hear difference of even 1 dB between two LOSELES files. May be encoder "thinks" that way.
So, your theory is that lossy encoders attenuate the signal a bit ? I don't think they do -- at least most of'em.
2. Since encodings are lossy, one file can be subjectively louder than the other while both of them have the same loudness. So psychoacoustic model tries to compensate it.
[a href="index.php?act=findpost&pid=342255"][{POST_SNAPBACK}][/a]
Please explain what the objective loudness is which you refer to by "loudness".
Sebi