I call this picture, "Death Of Audio". If you understand what it is saying, it's downright shocking, and makes a compelling case that Fraunhofer is unacceptable _above_ 128K bit rate. SWA160 is the name, the HF cutoff is now exactly 20Khz, and get a load of the _savage_ white artifacts top right. Look at the pre-echo/over-ring chart- the huge ragged spike is all pure artifacts, generated by the sharp edges that start and stop the 'wavelets' (they are _not_ the literal, audio form of wavelets, they are simply tiny bits of waves specifically chosen to give encoders fits.) Plainly, Fraunhofer is having fits, being driven into extraordinarily intense artifacting. Other than this, it does okay- but what a 'but!' I'd have to mark this 'Not Acceptable'. Also, no other >160K encoder cuts off lower than 22 kilohertz- in fact BladeEnc can deliver more extended highs at 128K than Fraunhofer is willing to do at 160K. It's the same problem as before- too much effort to keep clarity up to the cutoff frequency, and nasty artifacts when you hit that frequency and bounce off it. This is not an audiophile approach- it's a mid-fi approach. It'll sound worse the better your equipment is. However, there _is_ music for which this approach will work- for instance, Steely Dan is very controlled, disciplined music in which all the tonalities are very carefully sculpted and controlled. If you used Fraunhofer on Steely Dan, it would sound more like Steely Dan. That's not wrong. Using it on Fairport Convention with Sandy Denny- now that would be just wrong :)   Blade192 is typically interesting- again, even at this high bit rate, Blade has no compunction about throwing massive over-ring at the problem. The over-ring after the first noise band is as intense, and almost as sustained, as Fraunhofer 160's artifacts. Blade's really quite shameless about this- it absolutely glories in using over-ring. However there is quite a surprise in store- look at the pre-echo/over-ring charts for this and the next Blade, 256K.   As you get to Blade256, the whole bass and midrange over-ring tightens up dramatically. Bass over-ring is less than a third of what it was, the whole midrange is twice as tight as before- anyone using BladeEnc and considering higher bit rates should seriously consider 256K if they were even _thinking_ about using 192, as the improvement is all out of proportion to the added file size. In the deepest bass, Blade goes from the _worst_ over-ring to almost the best. That best is, surprisingly...   Amadeus256 (a LAME encoder) is the surprise winner here, on several levels. Its frequency balance rivals any other encoder at any bit rate- but the real startler is the pre-echo/over-ring chart. Amadeus's LAME export at 256K not only meets or beats any other encoder across the board for pre-echo/over-ring, it also substantially beats itself at 320K (!) and, at the important areas of extreme low frequency authority and ability to produce that 1K tone and then make it instantly stop, Amadeus's LAME 256K export is simply the best performance you can get at any bit rate (except of course, not even mp3ing in the first place).   When you've been looking at all those different sonograms and charts it's startling to be confronted with the Amadeus256 chart- what does it mean to the sound, to have the over-ring dropping off the chart like that, vanishing at some points? Essentially, this becomes a question of coherency- with an ability to damp over-ring like that, sounds and tones can come out of a background of silence and not blur against each other. There's a sense of spaciousness that is not simply brightness, an ability for the sound to startle and then be instantly gone again. The sound gets out of the way of the music more easily. Amadeus256 doesn't compare with the finest in audio equipment in this- it's simply the best out there for mp3s. However, within the confines of mp3s, this version of LAME is plainly a high point.   CC256 is quite similar- which is hardly suprising, since it is _also_ LAME CBR at a 256K bit rate. It doesn't quite hit the peaks of Amadeus's version of LAME- its over-ring is a little less diffuse, its damping of over-ring at 1K and in the low bass is not as profound. These are very minor differences, though- in essence it is very much the same encoder. The lesson to be learned here is 'LAME 256K is some kind of amazing optimisation- it peaks out at that bit rate and beats any other encoder at any other bit rate, including itself at higher bit rates'. Strange but true.   Blade320 finally, amazingly begins to actually try and suppress over-ring and pre-echo :) as we've seen at 256K it began to seriously tighten up the bass, and at 320 what happens is that Blade begins to control over-ring in a very unusual and interesting way that no other encoder approaches- rather than just cut it out, Blade continues to do over-ring, but attenuates it in a manner very reminiscent of natural reverberation, only a lot faster. Those stubbornly persistent high frequency over-rings, always nicely diffuse, have taken on a 'fading away' quality- and this is Blade's triumph- that even to the last extremes of bit rate it will always find something better to do with the bits than mere (faugh!) over-ring.   To be specific, what we are seeing with Blade is an interesting optical effect that no other encoder's sonogram quite matches- if you look at the noise block, particularly the first one, you will see that it looks quite _dark_ compared to any of the others. There are a few light spots and speckles, but the overall impression is noticably darker than any other sonogram at any bit rate. This, of course indicates accuracy in reproduction, as a perfect reproduction would be a uniform black across the entire sonogram- but the significance of the 'dark-seeming' noise block of Blade 320K is this: it's still not fantastic at transient attacks, but there is nothing to rival it in tonal purity. Blade is taking all the extra bits and putting them directly to the extreme highs without being overly fussed about pre-echo: the overpowering necessity here is to reproduce any tonality, no matter how demanding or subtle. Anyone who takes acoustic music seriously would see the applications of this- if you had to pick one encoder to get across the colors and tonalities of classical music, it would have to be Blade. Even at 320K it won't come up with the energy of a LAME encoder for something like rock or electronic music- but if you're going for subtlety, no other encoder is willing to go that far with it. Blade's final appearance in this study exemplifies the performance it's been sticking to the entire time.   Amadeus320 is peculiar in the extreme- unless my testing has gone horribly wrong for some reason, Amadeus LAME at 320K is significantly worse than at 256K. (Isn't that the sort of information you love to find out about?) To be specific, total frequency response error is about twice what it should be and the pre-echo and over-ring are substantially worse than at 256K. What seems to be happening (and it's hard to tell) is that it is trying to do what Blade does- attain great purity in those noise blocks, which by definition can't be reproduced in the given bit rate) and is losing control of the mids, the bass, pretty much everything else. Whoops. This is in fact borne out by the charts- of all the charts, Amadeus320 has the evenest frequency response across the frequency chart. At all points it's beaten by Blade320- it's not the best- but for the degree of error, the error is the most consistent across frequency, where all the other encoders and rates tend to produce increasing amounts of error in the highs. Also, it rises to an unacceptable degree of over-ring in the upper mids- and then falls back again, more than any other encoder. It really looks like at 320K, this LAME encoder is trying terribly hard to produce _even_ amounts of error across the frequency domain while still controlling pre-echo and over-ring. Unfortunately, in doing this it loses the balance and gives up the huge advantages Amadeus LAME 256K had, in its effort to do what Blade does. Stick with the 256K- if you want what Blade does, use Blade for that encoding :)   CC320 is the last encoder being covered (barring possible additions later). It doesn't bring anything new, but neither does it commit the errors of Amadeus LAME 320K: CokaCoda LAME 320K manages a nice balancing act that avoids obvious problems. It doesn't beat Amadeus LAME 256K in pre-echo but it comes very close, almost perfectly mimicking that pre-echo chart. It doesn't beat Blade in evenness of frequency response, but again it comes very close, almost perfectly mimicking _that_ chart. It doesn't reach Blade's quality of amazingly accurate extreme highs, but it comes very close. It's very consistent. No other encoder handled the demanding 'wavelet' section of the test wave quite so adeptly. Several other encoders will beat it in isolated areas, but if you can spare the space for 320K and need a versatile encoder that strikes a good balance among several different virtues, CokaCoda LAME 320K looks very promising.   So, to sum up- in the high bit rate area it comes down to LAME and BladeEnc. LAME (represented by several different encoders) tends to hit performance peaks- Amadeus hit a peak beyond all others at 256K, CokaCoda put in a performance nearly as excellent and more well-rounded at 320K. BladeEnc defined an area all its own- extreme tonal purity and sweetness at the expense of transient definition and punch- and just kept getting better until at 320K it was doing things (the 'reverb-like' over-ring) that no other encoder pretended to do. Fraunhofer disgraced itself with HF artifacts- but let's not forget the stunning, virtuosic performance _it_ put it at a lowly 32K. The message to learn here is simply that each encoder, whether by design, optimisation or mere coincidence, has developed a sonic personality that is completely distinct and centers on different values. Fraunhofer is all about low bit rates- and disgraces itself when it tries to act High End. LAME seems to be about energy and high performance, and sure enough, it hits some peaks to exceed all the other encoders in just those areas. And Blade seems to place tonal purity above all else, especially transient clarity- and tonal purity is exactly where it cannot be surpassed- if you're willing to pay the price. If you push each of these designs to its greatest peak, they do not become like each other- they all become more like themselves. Hopefully, with a clear enough understanding of what that means, it will be obvious which one to use for a given mp3 encoding task. Click here to return to the start page.

page created Mon, Oct 23, 2000 last modified Fri, Oct 27, 2000 Send Email back to Encoders Analysis of encoders using sonogram plotting Amadeus256.mp3 137 K MPEG Audio Amadeus320.mp3 172 K MPEG Audio Blade192.mp3 103 K MPEG Audio Blade256.mp3 137 K MPEG Audio Blade320.mp3 172 K MPEG Audio CC256.mp3 138 K MPEG Audio CC320.mp3 173 K MPEG Audio SWA160.mp3 88 K MPEG Audio