Many indie musicians use 128K bit rate for their mp3s. There are several reasons for this- most directly, some sites (such as mp3.com) require it. Additionally, musicians often want to sell their CDs to people who have downloaded the mp3s- one way to encourage this is to make the mp3s somewhat sonically compromised- not outright bad-sounding, but erring on the side of file smallness rather than giving people, for free, something they can convince themselves is as good as having the CD.   128K bit rate is also economical of space- frequently if some device is boasting '10 hours of solid music!' from playing mp3s, it is probably assuming 128K mp3s.   Comparing 128K mp3 encoders is somewhat tougher- we may be flipping back and forth comparing small details of sonograms or charts. They look more similar than the obvious distortions of 32 and 64K- but we will still find striking differences, that can be quantified and analysed.   Amadeus128 (a LAME encoder) leads off the pack, with a very strong showing. It does quite well in purity to frequency response, and has a particular characteristic to its pre-echo and over-ring characteristics: it resists pre-echo and over-ring at the central 1K frequency where there is a strong tone. It discards >16K treble as every 128K encoder but Blade does, and it controls the over-ring after the first noise band pretty well. There is also another characteristic- Amadeus LAME and its fellow LAME encoders also seem to resist pre-echo and over-ring in the low bass. If you look you'll see it, visible only in the rightmost chart of each encoder's results. This characteristic maps quite directly to ability to control the bass- it equates to bass authority.   Blade128 is, as always, taking its own path- given a mere 128K Blade goes a huge amount of the way towards accurately reproducing the ultra-high frequencies, 16K and up. It does this at a cost that's also very typical of Blade- look at the huge amount of over-ring present after that initial noise burst, and the substantial over-ring after the 1K tone. Look at the amount of pre-echo immediately following this- if you flip between this and the next sonogram (CokaCoda 128) they will seem like entirely different sounds. CokaCoda is producing distortion on the 'impossible transient', but Blade absolutely turns it to mush. And yet nothing even approaches its lack of harmonic distortion in the extreme highs- and it's very competitive with all of the encoders over the full range of frequencies, with a visible 'notch' in the chart that corresponds to 1 kilohertz and indicates that Blade is reproducing that critical, pure, 1K tone with even more attention to its accuracy. If you are looking for great tonal purity at 128K and the presence of extreme highs, there's no other choice- but you must sacrifice all transient accuracy and impact to get it. Some may find that too high a price to pay, but there are types of music that would suit this type of sound. One example might be chamber music- such as a string quartet.   CC128, CokaCoda constant bit rate of 128K, falls solidly into the LAME camp with regard to its sonic characteristics. Compared to Amadeus's implementation of the LAME encoder, CokaCoda handles the over-ring of the noise band a touch more poorly- handles the impossible transients a tiny bit worse- gets slightly more disheveled dealing with the tone wavelets at the end. These are, however, extremely tiny differences- indicating that for CokaCoda, the VBR option takes on additional importance. For straight 128 encoding, Amadeus's version of LAME is slightly but clearly preferable.   SWA128 offers several points of great interest to an audiophile- it puts up a very strong showing when you simply look at the charts, but there are certain details in the sonogram that are extraordinarily significant. First- look at the first noise burst, the dark band directly below the white area (that all but Blade throw away). Note that it is visibly 'cleaner' than any of the other encoders- but! That is not telling the whole story- look closely at the area to the immediate right, where over-ring happens! See the little white horizontal line? For many high end audiophiles, particularly audio designers, alarm bells are going off loudly upon seeing that. It is a picture of sheer over-ring: of a very objectionable kind of over-ring that will be heard as an ear-fatiguing haze, initally zippy and exciting-sounding, but rapidly turning to fatigue and dislike. Fraunhofer were brilliant with the low bit rate algorithms- but their evident decision to push the accuracy of 12-16K at the cost of introducing a _severe_ ringing at the cutoff point is a very bad decision, sonically. Compare it to the much gentler, more diffuse over-ring from Amadeus LAME. (My understanding is that Xing encoders make this error even more severely- for now I am looking at LAME, Blade and Fraunhofer, because with Xing I would have to pay for an encoder that is widely considered inferior.)   If you cannot hear high treble, enjoy 'etched' and hazed treble, or simply consider it a small price to pay for what is otherwise a very strong and clean encoder, you might prefer Fraunhofer. There is, however, one final caveat. Have a look at the 'wavelet' area to the far right of the sonogram. Most encoders evenly produce distortion components here- but not Fraunhofer! Instead, it is like the little girl with the curl in the middle of her forehead- for the first bunch of wavelets it puts on a bravado performance of startling accuracy and then BAM! Suddenly there is severe artifacting, right in the most objectionable area (upper-midrange) and then it's gone again. Fraunhofer is not _even_ in its error content. Its sonic characteristic is that it attempts to put on a Hi-Fi Stereo Spectacular, complete with highs that will cut glass whether the music likes it or not- and periodically blows a seam, throwing in brief but severe artifacts when sufficiently challenged. That describes a clear but brittle sound- and I can't overemphasise the high frequency ringing, which is comparable to even Blade's degree of blur except that Blade is a blur, and Fraunhofer is an outright _ring_ at a sharply resonating cutoff frequency. For this reason I can't recommend Fraunhofer to audiophiles- the sonic signature is very badly designed at 128K and up.   Moving on, VBR5 is CokaCoda variable bit rate encoding at 128K bit rate, maximum of 256K, and a quality level of 5. (With CokaCoda, increasing numbers decrease the quality.) This equates to about 140K in practice- it's not quite 128K anymore though it's close. CokaCoda VBR5 is in many ways the quality winner- in some ways it compares to Fraunhofer, in that there is the hint of a HF ring just like Fraunhofer's, but it is not as severe and particularly at the 5 setting it is markedly more diffuse. The wavelet section is handled more consistently, and the initial wavelets are markedly cleaner than even Fraunhofer- but there's one more thing that doesn't show directly in the sonogram- you have to compare the pre-echo charts to see it. LAME VBR is substantially the winner in lack of low-bass pre-echo and over-ring. All those charts are on the same (arbitrary, but consistent) axes- and the amount of frequency distortion in the bass matches or beats every other encoder, including Blade- and the amount of pre-echo and over-ring in the bass is _half_ that of Blade or Fraunhofer (and slightly beats CBR LAME). This characteristic translates directly to bass authority- the funny thing is, this is in spite of the presence of that single large artifact on the overly-long squarewave! Is LAME VBR stealing information from the subsonic ranges to add muscle and control to... well, the subsonic ranges? The area on the charts we're talking about clearly corresponds to 15 hz and below. This may be a behavior similar to Fraunhofer in the high frequencies- LAME VBR can make the bass as solid as a rock, but the cost of it may be the occasional artifact- the artifact itself being subsonic. To some extent that is more permissible- stronger subsonic artifacts are caused by trucks passing by in the street.   VBR7 is almost perfectly identical with VBR5, except that it is slightly smaller and the ringing at the high frequency cutoff is a bit less diffuse, more focussed (a bad thing- though it's half what the Fraunhofer ringing is). Bizarrely, a look at the charts will show that in the 1 kilohertz region, VBR7 has slightly less pre-echo and over-ring than VBR5 had! For the most part, however, the encodings are sonically identical. Until you get to..   VBR8, where it breaks down. VBR8 has twice the frequency distortion, particularly in the low end- and three times the pre-echo and over-ring, substantially more than CBR LAME encoding at 128K. Oddly, VBR8 is _less_ than 128K: the total file is 66K where the constant bit rate files are at least 68K and at most 71K. However, the severe rise in distortion makes the saving in file size a bad move- in some frequency ranges the over-ring is actually worse than Blade (and that takes some doing!) without any corresponding benefit in high frequency extension or response evenness.   My recommendation would be as follows- CokaCoda LAME VBR 128/256/7 for most rock and electronic music or anything that needs a lot of guts and glory to it- Fraunhofer for anyone who wants to be hit with a lot of simulated super-fidelity, or to show off stereos with (not for long term use), Amadeus LAME export if you need to put out exactly 128K CBR, and Blade if you must have 16-22K represented at 128K bit rate, or if you don't need any transients but must have sweetness and tonal purity, even at 128K.   Next, we will finish up by looking at the high bit rates! The trends we're seeing will continue, sometimes in a shocking way... Go to the next page

page created Mon, Oct 23, 2000 last modified Fri, Oct 27, 2000 Send Email back to Encoders Analysis of encoders using sonogram plotting Amadeus128.mp3 69 K MPEG Audio Blade128.mp3 68 K MPEG Audio CC128.mp3 69 K MPEG Audio SWA128.mp3 71 K MPEG Audio VBR5.mp3 111 K MPEG Audio VBR7.mp3 106 K MPEG Audio VBR8.mp3 66 K MPEG Audio