In lossy enconding, the quality evolves, from the strongest compression to the weakest, in a rather strange way.
At very low bitrates, that is below 96 kbps, the more it is compressed, the more it is distorded, and treble-rolled off. The treble loss is done on purpose, because treble are the heaviest frequencies to encode. They are plainly removed before the encoding. Otherwise, the encoder would use 50% of the file for frequencies, if not inaudible, at least very little important in the orchestra, and the remaining 95% of music would have to deal with the 50% remaining room in the file.
Between 96 and 140 kbps, according to the hardware and the listening conditions (laptop speakers especially), some people find that the sound is the same as the original. But on a hi-fi system, this is not the case. Distortion is present.
At 140 kbps, there is a change. The sound becomes perfect for a large part of the instruments, or even for entire tracks. Degradations can still be audible, but they do not affect all sounds. Only given frequecies or instruments. Before the 2000', castanets were well-known for being unencodable into mp3. Since then, encoders have been adapted.
From 200 kbps upwards, succeeding an ABX test becomes a real challenge. But it is still possible as long as we purposely choose the musical sample, among tens of others, that gets down the encoder. With mp3 especially, that remains true at high bitrate : everything's perfect, except for a small glitch or click, just in one place, just where it is annoying when we know where it is. The kind of things that we'd better not know if we want to listen in peace, otherwise, we hear is it everytime.
But mp3 is hindered by too restrictive technical specifications, like the impossibility for VBR to go above 320 kbps, even for a small time, just where it glitches and where it is needed, or for the impossibility to manage the last frequency band scalefactor independantly from the others, which requires to cut treble a bit agressively in order to avoid filesize bloating when there is just a small sound just in this band that requires all the possible precision.
Mp3 has thus been surpassed in quality and reliability by other lossy codecs. Musepack (mpc files. To my knowledge, only Cog can play them on MacIntosh) have been the pioneer in audiophile lossy codecs. The sweet spot is around 170 kbps. A nearly always transparent sound, with still some small glitches, but less than in mp3.
It is now supplanted by AAC (iTunes or Nero codec, .aac or .mp4 extention), proprietary, and Ogg Vorbis, free.
ABC/HR tests (large scale ABX tests that allow to rank encoders by quality in a double blind process) on average bitrate, near 140 kbps, have shown that Windows Media Audio Pro was very good also. But, not satisfied with being proprietary, it is also rare, and often confused with WMA standard. Contrary to WMA standard, WMA pro is unplayable on most mp3 players.
WMA standard is not very good. It has been ranked below mp3 in Roberto Amorim's 2004 128 kbps multiformat test.
Thus we have, by ascending quality order, around 140 kbps
-WMA standard
-Lame mp3
-AAC, WMA pro, Musepack and Ogg Vorbis, all tied.
According to Guruboolez' personal ABX tests (a golden ear having performed hundreds of ABX on hundreds of musical samples from all musical styles), Vorbis would be slightly superior to the others since the aoTuV R1 version (from 2005).
These codecs acheive a quite good transparency since 140 kbps in high fidelity listening. It is still possible to trick them at these bitrates, but it is necessary to know them well. This is less easy than with mp3, where everything that sounds like a pulse train and that "buzzes" is badly rendered, for exemple.
From 180 kbps, one has to be an expert in order to ABX them, and choose the music in killer sample directories.
Beyond, some killer samples still exist, but quality keeps going up with the bitrate. To find a new killer sample is an exploit, appreciated by developers, who are never far away from Hydrogenaudio forums.
Musepack, with quality 10, for example (slightly below lossless) have been ABXed only once in the world, to my knowledge, and by one listener. When we have tried to repeat the exploit with the same sample, we failed.
To my knowledge, nobody has ever ABXed Vorbis at maximum quality.
However, this is not very interesting because between 190 and 400 kbps, all we gain is the elimination of a glitch or a click that affect 1 second among 30 CDs, and not far above, there are lossless codecs.
Personnaly, I use Vorbis quality 6 (around 180 kbps). I find Vorbis distortion, when it is audible, nicer to the ear that mp3's, or Musepack's. Mp3 and Musepack gurgle. Vorbis only seems to add hiss, so that subjectively, the sound itself, that is behind, sounds unaffected.
Here are my listening test results. Anyway, the whole topic is interesting. Everyone contributes with his own ABX : ABX Just Destroyed My Ego - Hydrogenaudio Forums
Many blind tests are listed in Hydrogenaudio's wiki : Hydrogenaudio Listening Tests - Hydrogenaudio Knowledgebase
Big ABC/HR public listening tests are given at the bottom, in "external tests", by Roberto Amorim and Sebastian Mares.
I recall that comparisons based on spectrgrams, cancellation, or artifact amplification (SoundExpert tests) are not valid for assessing psychoacoustic lossy encoders quality. |
