MP3 and other HiRes formats

1. Introduction

Hey babes... today I´ll be talking about high resolution formats / codecs. I will show that even lossy codecs like MP3, AAC or WMA are perfectly able of encoding music in high resolution with spectacular results. I will show you listening tests, graphs and measurement results. I´ll also talk about the necessary requirements so that all of these codecs perform at their very best. I´ll tell you how these codec really sound (minus the usual bullshit). And finally, I´ll give advice how to unleash the best possible quality while using lossy codecs. I won´t talk about FLAC, APE, TAK, ALAC or WavPack... those are all lossless, a.k.a. encode music without taking something away. This article is divided into seperate paragraphs which I´ll mention now so that you might jump to the one interesting you the most:

1. Introduction
2. A bit of lossy history
3. How lossy codecs work
4. Transparency
5. Fairytales - or why MP3 & Co. are HiRes capable
6. Seeing is believing
7. Listening to differences
8. And now... the problem
9. The solution
10. How to make the best MP3s (& Co.) ever!
11. This really is high resolution? Some disadvantages
12. REAL "lossyless" high resolution
13. The sound of noise
14. Transcoding horror
15. Conclusion

2. A bit of lossy history

Fig. I: Responsible for all the 'mess': Karlheinz Brandenburg (copyright: Wikipedia)

The guy above did it all: Karlheinz Brandenburg. He was part of a group of scientists at the University of Erlangen-Nuremberg and in 1989 he described in his dissertation several techniques necessary for lossy codecs. The principles he wrote about are the foundations for any lossy codec and they have been in use ever since. Herr Brandenburg is called by many people the 'father' of MP3. In fact, after his dissertation he further developed this codec in cooperation with other scientists at the Fraunhofer Society. Since then, MP3 has become the most dominant codec to store music with. Around the same time (1992) Sony unleashed ATRAC necessary for the MiniDisc, by that catapulting lossy coding into mainstream consciousness. In 1999, Microsoft released WMA in order to have its own MP3-alternative which then could be licensed to partners for loads of money (or so they wished). In the same year, the Fraunhofer Society standardized what was to be the successor to MP3: AAC or MP4 and just like WMA and MP3 the use of this codec costs money. The Xiph.Org foundation followed suit in 2000 with a completely open and free alternative called Vorbis, commonly referred to as OGG. But as I said above, MP3 still is the most used lossy codec even though it´s old and technically inferior to the other codecs (ATRAC is worse).

3. How lossy codecs work

To make it short: they make music smaller while trying to keep sound quality on par with the original they were encoded from. They remove what our ears cannot hear... well, that´s not very precise. They remove parts of the music our ears AND our brain are unable to perceive. 'Perceive' is important as lossy codecs remove information our brain would ignore anyway. Because of this ear/brain combination those codecs are called 'psychoacoustic'; us humans never listen with our ears only, our brain is indeed the biggest part of our hearing. I won´t go into detail describing how lossy codecs are able to shrink filesizes... but have you ever thought about the description 'they remove parts of the music'? In fact, they don´t literally remove those parts; what they do is dynamically decreasing bit depth for certain parts, frequencies or information they deem to be inaudible. For example: if a louder part masks a softer part, lossy codecs decrease bit depth for the soft part to, say, 1 bit (see Fig. II). When you decrease bit depth you create a noisy residue, called quantization noise. The ability of a codec to hide this noise partly determines how transparent it sounds to us.

Fig. II: audio masking and subsequent bit depth decreasing (copyright: Wikipedia)

4. Transparency

If MP3 & Co. are only removing what we cannot perceive anyway, why do we sometimes hear more or less horrible compression artifacts? Well, the sound of a lossy codec basically depends on how efficiently it encodes, on the available bitrate (measured in kBit/s), encoding speed (fast isn´t always best) and if the codec is maintained well (a.k.a. continuously developed). MP3 for example has been in constant development for 20 years, first and foremost in its LAME variant. It is now so good that it reaches transparency for many people with a bitrate as low as 128 kBit/s. 'Transparency' means that the lossy encoding cannot be distinguished from the lossless original it was derived from. In other words, for many people a 128 kBit/s MP3 sounds the same as the original, 11 times bigger 1.411 kBit/s (bitrate of the CD) lossless source. In my opinion that fact not only proves how un-trained the ears (and brains) of casual listeners are, it also serves to point out how good lossy codecs have become. Wouldn´t it be transparent, you´d hear artifacts like flanging, pre-echo, smeared transients, problems with virtual stage, distortions (quantization noise), etc. These artifacts are a problem... for decades now MP3 has "enjoyed" a reputation as bad as the always-bad-stepmother in fairytales. Audiophiles and casual listeners alike constantly claim that MP3 (or any other lossy codec of their choice) sounds cold, lifeless, digital and... yes, compressed; lossy codecs are too often accused of "dumbing down" the sound.

5. Fairytales - or why MP3 & Co. are HiRes-capable

You see, all of this is was the truth. 15-20 years ago. In 2013 it´s indeed like the bad stepmother; a fairytale. Most lossy codecs are far, far better than their reputation. To you this may be surprising, dear constant reader, but they are indeed able to encode high resolution material like 24/44.1, 24/48 or even 24/96.
...
...
...
...
(Pause for dramatic effect)
...
...
...
...
Don´t believe it? Here´s why... all lossy codecs share a common trait: they encode any audio material not with static integer but with floating point values. You may remember that I talked for hours and hours about the ability of the MiniDisc to encode audio signals with a quality surpassing that of CDs. It´s exactly the same with MP3 & Co.: since they all employ floating point precision, the bit depth fed to the encoder's input is irrelevant, they will encode anything, whether it´s 16 bit, 24 bit or 32 bit.

What´s the difference between integer and floating point? Integer handles values like this: 23, 45, 156, 001, etc. Floating point is like this: 7,89654367, 674,342167, 55,236548955214587, etc.... you get the idea. It is able to work with higher precision since it allows for much more possible values than integer (or fixed point). The results are a gigantic dynamic range and a spectacular signal-to-noise ratio. For 24 bit integer you have a (quantization) noisefloor as low as -144 dB, for 32 bit floating point it´s at a stunning -202 dB.

6. Seeing is believing

The first graph below presents a tiny sine at 1.000 Hz with a level of -90 dB on an original 24/48 wave file (-> lossless, see Fig. III). This is so low-level that you won´t ever hear it. But you can see that the sine still looks like a perfect sine. 24 bit obviously allows for enough possible values for any signal at -90 dB to be properly represented, no wonder considering a signal-to-noise ratio of -144 dB. The next graph (Fig. IV) shows the same signal, though this time encoded with 16 bit. The situation now changes: this low-level signal gets very close to the -96 dB noisefloor limit of a common 16 bit system. In place of a sine you now have the very famous 'digital staircases' signal so often used to (mis-)represent supposed flaws of any digital system.

Fig. III: 1.000 Hz sine, -90 dB, 24/48 WAVE

Fig. IV: 1.000 Hz sine, -90 dB, 16/48 WAVE

Now let´s have a look at several codecs. Are they able to encode with high resolution so that the sine will look like Fig. III? Please be aware that I used only the latest codec versions; I also employed their highest possible bitrate. In case of MP3 (LAME) that´s 320 kBit/s, for AAC (Apple's implementation) it´s 320 kBit/s as well and for WMA Professional it´s 440 kBit/s. 

Fig. V: MP3 320 kBit/s, 1.000 Hz sine, -90 dB, 32/48

Fig. VI: AAC 320 kBit/s, 1.000 Hz sine, -90 dB, 32/48

Fig. VII: WMA Professional 440 kBit/s, 1.000 Hz sine, -90 dB, 24/48

Duh! The three graphs (Fig. V to VII) prove that we have true 24 bit resolution with every lossy codec represented here. Not one of the sines looks like Fig. IV. As I said: high resolution - despite lossy compression by a factor of 7. Wanna hear a word from the inventor of MP3, Herr Brandenburg (the guy up above), regarding this?

There are some kinds of deficiencies of standard audio equipment which cannot be found in properly designed Layer-3 and AAC codecs. They are listed here to mention the fact that it does not make sense to test for them. Most noticable are Dynamic range: MP3 and AAC both contain a global gain adjustment parameter for every block of music data. According to the word length and resolution of this parameter, the dynamic range of both MP3 and AAC is well beyond the equivalent of a 24 bit D/A resolution. In short, MP3 and AAC represent the music in a way that the dynamic range of every known audio source is perfectly retained. (Source)

High resolution again, he´s well aware of it. Of course he is, he developed it. But how will all those lossy codecs react to those plain old measurement signals he describes as unneccesary? I mean, lossy codecs are designed to work very well with music. Measurements are entirely different... those don´t fully adhere to all the psychoacoustic principles so vital to every lossy codec so effectivey they should measure horribly. To find out how they react to these difficult conditions I used RMAA and compared all encodings to the original .wav file they were derived from (a 32/48 kHz file). Therefore, the resulting, decoded test files were at 32/48 too (WMA Professional: 24/48). Look below at Fig. VIII and see for yourself how they managed to deal with this situation.

I have to repeat that lossy codecs behave worse with measurement signals, they are simply not developed for this kind of signal.

Fig. VIII

Duh... again. Think about it: roughly 90% of the test signal has been removed... gone forever... Poof! (WMA Professional: 80%) - yet they still perform so fuckin' well... and of all things with signals they aren´t even designed to encode well! But you probably glimpsed that there are quality differences between all those codecs. OGG for example ends up badly, this is due to OGG not being revised often enough. The same goes for WMA Professional; while it clearly is one of the winners in this contest it has an advantage because of its comparably high bitrate. Ironically, the oldest codec, MP3, fares best in RMAA's quality assessments. The picture changes somewhat when one looks at the graphs... to make a long story short, the best codec is AAC as used and continiously developed by Apple for iTunes. The AAC version from Nero isn´t able to hold up that well, again caused by long pauses in development.

7. Listening to differences

Measurements are one thing, listening to results another... and I´ll simply show you the residue that is produced when MP3 encodes music. I chose something you might already know, a song you´d hear on the radio: Woman's World (-> video) by Cher from her first album in 12 years, Closer to the Truth. The residue is called - see above - quantization noise, in this case mixed with imprecisions produced by the filterbanks MP3 needs to find out what it might erase. These quantization artifacts are usually hidden by the music itself (this is supposed to be that way). In order to reveal how MP3 works I simply inverted the phase of the original 32/48 .wav source file and mixed it with the decoded MP3 file. Et voilà, every little speck of dirt MP3 produced for this file is subsequently revealed. I didn´t alter the gain of this file, the level you´re hearing is the actual level of the errors within the MP3 file. Now, for those who´ll now say "Very loud... and that´s exactly why MP3 sucks big time" I can only say that you´re stupid. Just consider for a moment how our ears & brain perceive sounds and have a look at this: the RMS-level of the original file is -15 dB, the residue has an RMS-level of only -43 dB. A gain reduction of just 6 dB feels to our ear / brain half as loud. I think, you can do the math yourself.

Again: you cannot use this noise to point out how badly any lossy codec performs - it´s how it´s supposed to work, nothing else. As I said: (with a high enough bitrate) usually these artifacts are hidden and inaudible.

Sounds funny, doesn´t it? You can hear that the residue mirrors the original file closely, except that it´s stripped of bass and mids. Our ears & brain aren´t very good when it comes to high frequencies. Hence any lossy codec prefers to focus on the treble area. You can also hear that artifacts rise in level when the music gets louder and that their level decreases when the music gets softer or less complex. This is why lossy codecs are able to encode high resolution music - the additional resolution is kept. After all, high resolution is, when it comes to bit depth, nothing more than lowering the static quantization noise floor. To be fair, the artifacts left by MP3 are anything but static, they are chaotic and at all the places where MP3 removed information. BTW, the MP3 used for this example was encoded with 320 kBit/s. When the bitrate is reduced to, say, 128 kBit/s, the artifacts are considerably louder.

Update 30.03.15: Soundcloud used to host the audiofile containing the compression artifacts. But just this day, Soundcloud decided to delete everything I ever uploaded because their automated content protection system detected several breaches of copyright.

Well, of course it did! For my reviews I need to listen to music and in order to make sound differences available to you, dear Reader, I uploaded several samples, each of them - at max - 30 seconds long. Naturally, this isn´t a breach of copyright, because a) I don´t have a commercial agenda nor background for this blog and b) I don´t advertise filesharers nor do I encourage to download things illegally. I don´t even want to mention, that only 30 seconds (!) of a particular song or piece are far too short to be enjoyed properly by anyone who attempts to be an illegal asshole. Yet Soundcloud fears the lables and their paranoia of copyright breaches which in turn prompts them to be paranoid and incompetent ninnies themselves.

I hate paranoia, I don´t want to have anything to do with stupid people / companies and everything was deleted anyway... so I decided to delete my Soundcloud account. Sorry for that, dear Reader.

8. And now... the problem

All of this would be marvellous... if all those nifty lossy codecs would be decoded properly. Have you ever heard about a digital audio player or a smartphone that decodes lossy codecs with 32 bit floating point precision? See, neither have I. And that´s where the beast rears its ugly head: if they aren´t decoded with full precision they produce strong, additional quantization distortions NOT contained inside the signal itself. Software players for the PC usually don´t suffer from this malady, foobar2000, Winamp or JRiver decode MP3 & Co. with full floating point precision for simple playback. This makes perfect sense because the aforementioned softwares work internally with 32 or 64 bit floating point precision anyway (for DSPs, volume control, etc.). JRiver deemed this issue so important that they opened a thread in their forum, talking about it. But should you attempt a conversion from MP3 to WAVE, the basic problem is resurrected, meandering around again like a zombified corpse. Foobar2000 for example assumes that every lossy file was derived from CD; using the 'Auto' bit depth configuration in its converter dialogue converts everything lossy to 16 bit, whether it´s MP3, OGG, AAC or WMA (see Fig. IX). High resolution? Forget it. The same goes for the usually wonderful software dBpoweramp: floating point decoding has to be activated under advanced options (Fig. X).

Fig. IX: foobar2000 converter dialogue

Fig. X: dBpoweramp configuration, advanced dialogue

By all means, the 16 bit, quasi-standard decoding of MP3 & Co. isn´t a good thing. Imagine a CD you ripped yourself to MP3; these files were derived from a normal 16 bit source. 16 bit decoding should be enough then, right? Should, but is not. While the decoded data boasts the sources' original bit depth again, something new and eerie has been added... and I don´t mean the inlying compression errors produced by the encoder. No, this thing from the crypt is additional quantization noise produced by the decoder. Cause: truncating floating point values to integer values. These additional artifacts are produced only because the decoder works at half speed and with half of its options. Look at the graphs below:

Fig. XI: original, lossless 16 bit wavefile for comparison

Fig. XII: MP3, decoded with 32 bit floating point

Fig. XIII: MP3, decoded with 16 bit integer

Doesn´t look so bad, you´re saying? Well, then look again at the signal causing the noise, it´s a simple 1.000 Hz sine only. Fig. XI shows an original, lossless 16 bit wavefile, the quantization noise is evenly distributed across the spectrum. Fig. XII shows an MP3 file that has been decoded with floating point precision and while there are distortions, those are well below audible levels; most of them are at frequencies we cannot hear well. Fig. XIII shows the same file, this time decoded truncated to 16 bit integer. The artifacts have doubled - and only because they were decoded with 16 bits instead of 32 bit floating point. They might still be inaudible... but I´m not sure because now we have additional aliases at frequencies where us humans can hear extremely well. These distortions will be added by the stupid integer decoding, and it doesn´t matter if you have created those lossy files yourself or bought them at some online store. You´ve probably been listening to quantization artifacts all your life, errors produced by dumb decoding of portable players and stupid software. Even files encoded with WMA Professional, files that are clearly marked as being 24 bit by their data stream, suffer from erroneous decoding as most software decodes it to 16 bit only. I´m looking at you, foobar2000. Just because you´ve been programmed by people who give a shit about proprietary, 'bad' software coming from Microsoft, this still doesn´t mean that you have to behave like a silly goose.

9. The solution

I´m afraid that for the time being there is no solution. Companies producing digital audio players don´t seem to be aware of the problem or they just assume that your average-joe won´t notice it. Well, at least you can do something about it when decoding those files yourself with software. Just have a look at dBpoweramp again (Fig. X above) and configure it to decode MP3, AAC or OGG as 32 bit floating point or 24 bit. Do the same with foobar2000 in the converter dialogue (see Fig. IX) and change the output bit depth to '32' or '24', it´ll then decode lossy formats to their full potential when converting files to HDD. But there´s hope that this problem gains attention... people like Bob Katz who´s a mastering engineer and important enough to maybe excert some influence, mentioned this problem in a recent thread at the JRiver forum. Yeah, he was talking about dithering but at the same time he was fully aware that for MP3 & Co. it´s imperative that they are decoded properly and that decoding those lossy codecs with anything else than floating point will result in truncating values which in turn produces the aforementioned artifacts. He wrote:

"Did you know that all current Lame and Fraunhofer and Apple AAC and MP3 decoders run internally at 32-bit floating point? In fact, if you take a "16-bit" source AAC file and reproduce it through the AAC decoder, it produces a 32-bit float output word! If it was a very good encoding, you will lose audible depth if you reproduce it at 16-bit because more than 16-bits come out of the decoder. The output of an AAC decoder should therefore be dithered down from 32-bit float to 24-bits for best reproduction. Almost NO ONE does that, but they should, and I've heard the audible difference when I play AAC in an engine that permits that."

Thank you, Bob, exactly what I´ve been saying! At least someone acknowledges it. I won´t dither to 24 bit but each to his own (I´m not too fond of dithering to 24 bit, using a bit depth like this renders quantization-related problems moot). From now on, I will mention if digital players (portable CD players for example) are able to decode MP3 properly.

10. How to make the best MP3s (& Co.) ever!

In the meantime I´ll present some advice everyone encoding her/himself might find useful. I´ll also recommend the best sounding codec to you, based on my own personal experience. Please be aware that my suggestions will cause any encoding to take longer, if you don´t have the time to wait for the encoder, then don´t even bother. But then you´re not interested in best sound anyway, or are you?

First of all, all those nifty codecs are command line based, on a Windows PC they look like plain old DOS. Usually you can´t see this as the command line window is hidden by the software employing those decoders/encoders. But it nevertheless enables you to make everything yourself by using your keyboard... though I admit that it isn´t very convenient. So we´ll use nice, flashy software instead. The following encoding tips and setups will depend on the codec of your choice; in some cases their setup might be complicated for noobs but once you´ve done it correctly you won´t need to bother with it anymore.

MP3 (LAME)

One of the best and most versatile softwares around the net is the aforementioned dBpoweramp. It costs money but I can recommend it without reservations. It´s a powerful encoder/decoder for any format you can think of, it contains one of the best available CD rippers and its powerful talents are hidden inside an easy looking package. Even better, the CEO of Illustrate (company maintaining dBpoweramp) is a nice guy, discussing things and giving advice in his own forum and on hydrogenaudio. A free alternative would be xrecode II (shareware with a nag screen) but it´s buggy and inconvenient - use at your own risk. So, in case you want to use MP3 and in order to create the best sounding MP3s ever I recommend these encoding settings (Fig. XIV):

Fig. XIV: best encoding settings for MP3 (LAME)

If you´d like to use a different frontend for the command line based encoder/decoder instead  (xrecode II for example) I´ll now give you the commands so that you may copy and paste them:

-b 320 -q 0 --noreplaygain

The most important part is the '-q' switch, it configures the 'quality' option of the LAME MP3 encoder. The standard setting advertised by hydrogenaudio is '-q 2' but we want a choice and the best quality so we opt to set it ourselves. Why the constant bit rate (CBR) of 320 kBit/s when Hydrogenaudio recommends variable bit rate (VBR) in order to save on storage space? Think about it: 60 minutes of music occupy 137 megabytes when encoded with 320 kBit/s CBR. For VBR with an average of 240 kBit/s these 60 minutes take roughly 108 megabytes. A difference of 29 megabyte. In 2013, two to three photos on a smartphone alone consume this. We have to be realistic here: it might have been an issue 10 or even 5 years ago but nowadays with an abundance of storage space anywhere, surely we can afford bigger files. Furthermore, MP3 profits from more bitrate, no matter what the skeptics (hello, my dear hydrogenaudio-ists) are saying. More on sound issues later.

WMA Professional

Fig. XIV: best encoding settings for WMA Professional

I wouldn´t recommend WMA Standard as WMA Professional is superior in every way, it´s also the one codec officially supporting 24 bit output. Which is a fake of course, Microsoft just embedded an additional integer decoding option, internally it works with floating point just like other codecs. As a quality option, you should always use 2 pass encoding, it yields audibly superior results. Compared to MP3 above, 60 minutes of music occupy more space because of the highest 440 kBit/s setting: 189 megabytes. Should you really be concerned about storage requirements, using 384 kBit/s would work too, it´ll still sound well. None of this however hides the fact that WMA Professional enjoys close to zero hardware support. Most portable devices are able to play WMA Standard only - and I can´t recommend that one.

AAC (Apple)

Instead of using an old or proprietary codec I´ll recommend one of the most recent instead, one that also comes with ample hardware support by almost every manufacturer: AAC. Be advised that it now gets inconvenient. To make it easy, you could of course use the AAC codec from Nero (with dBpoweramp for example) but I´d advise against it; the one from Apple included with iTunes is much better (and it shames me to write this as I don´t like Apple). There´s only one way to unleash the iTunes encoder (or more precisely: the QuickTime encoder) with the best quality options: qtaacenc (get it here). You´d have to setup foobar2000 like this (Fig. XV):

Fig. XV: foobar2000 setup for qtaacenc

Just like with MP3 there´s a quality setting allowing for better-than-standard results (the standard settings are used by Apple for music they sell through the iTunes store): it´s simply called '--highest' (another source says '--high'). Don´t forget to instruct foobar2000 to use 32 bit during encoding, you now know well that AAC can handle it. Anyway, here are the commands in case you want something else besides foobar:

--cbr 320 --highest

11. This really is high resolution? Some disadvantages

MP3 and AAC (Apple) have one big disadvantage: they aren´t able to handle samplerates beyond 48 kHz. For those there´s only one codec left: WMA Professional. AAC (Nero) can handle 96 kHz too... but as I said above it´s not very good. Furthermore, devices usually able to handle AAC will react in strange ways (or not at all) when trying to play those 96 kHz AAC files. So you might want to use WMA Professional. But its main problem hasn´t changed: hardware support is extremely limited, not even software properly recognizes it.

What lossy codec to use depends on what you yourself consider to be high resolution. To me, HiRes starts with 24/48, for others it starts with 24/44.1. Strictly speaking, everything that´s not CD is high resolution. Take HDTracks: they sell even 24/44.1 as high resolution. More than one third of the music they´re offering is at 24/44.1 or 24/48. For those releases lossy codecs would be the perfect choice if you could ensure proper decoding on playback. And if you want to save on storage space, you might consider resampling your 96 kHz albums to 48 kHz and encode the result it with AAC or MP3.

12. REAL "lossyless" high resolution

But the best combination to save space and keep any file at its original resolution is... not FLAC. Have you ever heard about WavPack? Thought so. WavPack normally is, just like FLAC or APE, a completely lossless encoder/decoder that won´t ever touch the material it encodes. But it has a second, not so well known setting: WavPack lossy ('hybrid' is the correct designation). This 'lossy' mode is unlike MP3, AAC or WMA, it won´t remove anything within the music. As I said, the other lossy codecs work psychoacoustically and 'hack' into the frequency band at countless places, removing what cannot be perceived. WavPack lossy ignores psychoacoustics and does only this: reducing overall bitdepth of the file according to its level and distribution of frequencies.
Let´s assume a 24 bit file with lots of loud and soft parts. With WavPack lossy the soft parts will retain close to 24 bit resolution, loud parts will be reduced to 16-20 bit resolution. The resulting quantization noise is then moved by a very tame noiseshaper towards high frequencies where it cannot be perceived anymore. Very much like SACD. Unlike SACD though, bit depth isn´t static. WavPack lossy can be described as a 'dynamic bit depth decreaser'. In that respect it also differs from other lossy codecs; with them bitdepth changes a lot within frequencies, with WavPack lossy only from one level change to the next. It truly preserves the full bandwidth and dynamic range of high resolution material and in my experience, the added quantization noise remains completely inaudible. Remember the MP3 noise sample up above? If I would have extracted the artifacts of a WavPack lossy-encoded file, I´d have been required to raise the level by +50 dB to make them even audible!

Fig. XVI: RMAA chart comparing several lossy codec against their lossless source

Please refer to Fig. XVI above where three lossy codecs have to compete against their own source. While WMA Professional is superior to AAC Nero, it´s WavPack lossy winning the contest. It measures almost exactly like the original. Yet it´s 7 times smaller than the WAVE file it was produced from (1.200 kBit/s against 6.144 kBit/s). See Fig. XVII & Fig. XVIII for details.

Fig. XVII: total harmonic distortions - WAVE and WavPack lossy are the clear winner

Fig. XVIII: intermodulation distortions - WAVE & WavPack lossy measure the same

Because of its noiseshaping feature, WavPack lossy is ideally suited for encoding anything beyond 48 kHz. If you employ the '-x6' switch, you can exploit this further. Using this switch, I´ve found out during many hours of testing that WavPack lossy is fully transparent with 96 kHz material the moment the bitrate exceeds 1.000 kBit/s. Because of that I always use it with 1.200 kBit/s, as an additional security I also employ the '-h' switch (high quality). However, encoding takes forever. To me, this doesn´t matter; I only encode those files once and never touch them again. But with you it might be different - so you decide.

Fig. XIX: WavPack frontend with my recommended settings for best lossy quality

Fig. XIX presents my recommended settings for 96 kHz material using the WavPack frontend (for convenience I would´ve loved to recommend foobar2000 or dBpoweramp... but both won´t allow extra commandline switches - as it appears, foobar2000 now can. Very convenient, because it´ll use all your CPUs cores for additional encoding speed). The 'Extra Option' -x6 is most important to achieve the best possible quality... but using it will prolong encoding time (no kidding; while you wait, you could write a novel). The same goes for the '-h' switch and because it´ll also prolong time needed for decoding, it´s optional and not mentioned anymore. Anyway, for 44.1 / 48 kHz those 1.200 kBit/s are overkill, for 176,4 / 192 kHz they aren´t enough. That´s why the bitrate needs to be tailored to the samplerate:

44.1 / 48 kHz:

bitrate: 500-600 kB/s, switches: -x6 (optional)

88.2 / 96 kHz:

bitrate: 1.000-1.200 kB/s, switches: -x6

176,4 / 192 kHz:

bitrate: 2.000-2.400 kB/s, switches: -x6

13. The sound of noise

One of the most important questions of this article is how lossy codecs sound. And I don´t mean what the mainstream public thinks they are sounding which can be answered easily: cold, digital and lifeless (yet all use them - a mystery?). No, I mean the actual sound quality. I´ve said above that lossy codecs are far better than their reputation. BUT: you have to make sure to extract the best possible quality when using them! Please consider my encoding suggestions above again and remember that they take time but are worth every second spent on them. The best encoding isn´t the one that is the fastest (C'mon... a whole album encoded in 40 seconds... really?). To make it simple: the higher the quality, the longer it´ll take to encode. Just deal with it. Back to topic... if you used the best encoding options along with the highest bitrates and if you decoded all these files correctly (-> floating point) the sound of several lossy codecs is like this:

WavPack lossy: perfect. Using bitrates of roughly 1.200 kBit/s with 24/96 in combination with the -x6 switch it sounds exactly like the original. Always. Very limited hardware support... software support is good though. I have been archiving every bit of music with WavPack lossy since 2008 and I´ve never looked back. Not suited for portable use (for lack of hardware support), but perfectly suited for archiving and transcoding to other codecs like the ones below.

AAC (Apple): might produce instable staging, instruments occasionally seem to change size and position. Yet this happens so rarely that I could have been imagining it. Otherwise it´s completely devoid of artifacts. Sounds O.K. enough with lower bit rates. Hardware support is phenomenal. Simply the best mixture of convenience and good sound, therefore highly recommended for portable use. 

MP3: slightly 'dark', 'warm'. Somtimes sounds too dry, as if reverb has been reduced (especially audible with VBR). Smeared transients are another problem. No matter the bitrate, MP3 will always have difficulties encoding really short and tiny transients (the sample size for short blocks isn´t small enough). However, in 85-90% of all cases all of this isn´t audible at all. The danger of typical problems like metallic sizzling, flanging, etc. completely disappears if you use the highest bitrate of 320 kBit/s. Of course, using LAME avoids most problems anyway compared to other MP3 encoders. Hardware support? A 100%.

WMA Professional: on occasion creates instable staging worse than AAC (Apple). Instruments might change size and place, dimensions shrink or expand sometimes. This depends very much on the material, in many cases it´s completely inaudible. When used with 96 kHz, sound is too mellow. Otherwise it´s one of the most neutral and artifact free codecs available. Hardware support: laughable.

AAC (Nero): like WMA Professional, only (much) worse. Instruments always move around slightly, sizes vary as well, dimensions shrink and expand constantly. Furthermore, the sound feels 'blown up' at lower mids. Despite being much more recent and advanced than MP3 (LAME), it doesn´t sound remotely as good. Hardware support is - naturally - the same as for AAC (Apple).

WMA Standard: sizzles. Even with higher bitrates. On the other hand it enjoys almost the same hardware compatibility as MP3. I still can´t recommend it, it just isn´t good enough. 

OGG: was my standard choice more than 8 years ago. Shouldn´t be used nowadays. Obscures much of the virtual stage, sounds harsh (this is the only codec sounding literally 'digital'). Sounded different back in 2005: very beautiful and pleasant. Ignore it.

14. Transcoding Horror

If there´s one thing I hate, hate, hate people doing it´s transcoding from one lossy codec to another... or from 128 kBit/s to 320 kBit/s. You cannot imagine how many think that it actually improves quality - see here & here (I´ve found countless other examples, but they´re in German and my native language is a bitch for most people). The only thing it does is hurting quality immensely. To the encoder, the file that was compressed before is just a new file and it´s treated as such. Example: the really loud encoding artifacts within a 128 kBit/s MP3 are simply treated as they are, to the encoder they are just new musical information, a seemingly natural part of the music. You see, not one encoder in the world can distinguish between noise, artifacts or music... which means that artifacts are now treated as music. Back in 1999, when Microsoft tried to advertise its WMA Standard codec as superior to MP3 out of licensing greed, they tried to fool everyone to re-encode their MP3s to WMA. For this stupidity they should have been shot... so please, don´t you ever transcode from MP3 to MP3, AAC, WMA or OGG. The only allowed trancoding is one from WavPack lossy to another lossy codec (preferrably not WavPack lossy again).

15. Conclusion

My dear constant reader, I don´t know where you´re coming from, if you´re an audiophile, a skeptic or just feeling lucky to be here. But I know one thing: if MP3 & Co. would be sounding horrible no one would use lossy coding in the first place. If you´re an audiophile, do you really think that half of the earth's population fell for a ruse invented by some German scientist in order to 'dumb down' the sound of music? Then let me tell you that all the hills marbles of these 3 billion people are very much alive and there. This is no conspiracy, MP3 & Co. sound well enough that most people won´t even think about other formats. And as I´ve proven with this article, they are prepared to exceed and surpass mass-market sound with ease. Only if my suggestions are heeded, that is. Should you decide to follow them you´ll be rewarded with true high resolution sound... or high end quality, if you will. All coming from some 'dumbing down' lossy codec. I can´t stress it out often enough: take care in using the best possible quality when encoding with lossy codecs yourself. Take care in trying to make sure to employ the best decoding as well. Only then will you enjoy a sound quality you wouldn´t have expected from MP3 & Co. Happy encoding... and perfect decoding!

Last update: 29.12.2013

Spectacular sounding MiniDisc tweaking - FOR FREE!

Introduction

The following Article concerns itself with a piece of software called an Equalizer. Removing certain frequencies with this EQ enables the ATRAC codec necessary for any MiniDisc recorder to encode any piece of music in a superior way. Substantial proof will be offered by means of measurements as well as musical examples.
A warning: if you are lazy or if PC-based audio is a complete mystery to you this article isn´t for you. You also won´t be happy if you think that an equalizer is one of the devil's minions (an opinion clearly not based upon facts). Furthermore, the significant sound improvements described here require you to spend some time with your music, it means work. If you can´t afford the necessary time, you probably aren´t interested in good audio quality coming from MiniDisc at all. Or are you? ;)

Basics of lossy encoding

Lossy (or perceptual) codecs were developed for only one thing: making audio smaller without you, dear constant reader, noticing it. Using a datarate of, say, 128 kBit/s you´ll receive a file the size of only 5 megabytes (before: 50 megabytes). Every lossy codec available on the market performs this 'shrinking', it doesn´t matter if it´s ATRAC, MP3, AAC, OGG or WMA. How do these codecs achieve this? They 'erase' or 'alter' parts of any musical material, getting rid off things our ear cannot perceive anyway. Erasing certain parts of any object reduces its size which is most convenient for portable players not having terabytes of available space; they can store more music that way. In the article that follows I will compare ATRAC (necessary for MiniDisc) to MP3. Both codecs are equally old (20 years), yet MP3 is still in development. ATRAC however was declared dead in 2004 with the introduction of Hi-MD, since then it hasn´t been improved. ATRAC is effectively one of the worst codecs around, it doesn´t help that its datarate is comparably high (292 kBit/s). It could have been more effective if Sony would have decided to fuck their desire for power constraints (a superior ATRAC IC would have drained battery power faster). Well, we have to work with what we´ve got and how we are able to improve it. Anyway, the following chapters will show you the most important basic instrument any perceptual (lossy) codec employs to fool your ear... or more precisely, your brain (Fig. I):

Fig. I: Equal-loudness contour, logarithmic scale (copyright: Wikipedia)

On the left of Fig. I you can see the Sound Pressure Level, measured in dB. The line below presents several frequency points of the entire frequency range our ear can hear (20 Hz to 20,000 Hz), measured in Hz. Now take the highest frequency at 16,000 Hz; it would need a level of roughly 25 dB to be perceived as loud as the 4,000 Hz tone. The key thing you have to understand is the difference between actual levels and perceived levels. Our ear itself listens with almost perfect precision, but our brain decides what part of the received audio material is kept. Our brain has been 'tuned' over the course of several millenia to recognize the human voice with immaculate precision, the area from 2,000 to 4,000 Hz is the most important frequency band  with the highest sensitivity. So listening sensitivity of the ear/brain combination sucks at low and - especially - at high frequencies. Those frequencies weren´t necessary for survival, so they have been 'dropped' as a result. Which means that we have to make signals at those areas much louder. But it also means that we´re unable to hear grave errors at low and high frequencies - and that´s where perceptual (lossy) codecs come in:

Fig. II: Multitone signal, -6 dB, .WAV, 24/44.1 (linear scale, Hanning)

Above (Fig. II) you can see a multitone signal (.WAV) in 24 bit and 44.1 kHz which consists of more than 100 seperate sines. Frequency response ranges from 0 to 22.050 Hz. There aren´t any errors to witness (the thickened bases of the several sines are caused by imperfections of the Hanning windows function used to display the results). Compare that to the same signal, encoded with MP3:

Fig. III: Multitone signal, -6 dB, MP3 (Lame, 256 kBit/s), 24/44.1 (linear scale, Hanning)

It´s evident that something has been added by the MP3 compression (Fig. III). This something is simple quantization noise, inherent to every digital system. In this case it´s at roughly -90 dB (original .WAV file: -144 dB). This amount of noise also reveals how lossy codecs work. It is always said (even I did so above) that they 'remove' or 'erase' parts of the music with surgical precision. Well, it isn´t true, they don´t really 'remove' anything (ignore the cut-off at 19 kHz for a sec'). What any lossy codec does instead is decreasing bit-depth dynamically, taking into account complexity, gain and composition of the signal. Example: a clarinet, formerly having a resolution of 24 bits now has a resolution of perhaps only 6 bits. The remaining 18 bits resolution are non-existent anymore which - BINGO - creates quantization noise. Effectively, the quality of any lossy compression is partly determined by how well it is able to hide this noise. MP3 is very good at this, a wonder given the fact that this codec is over 20 years old. You can also see that the quantization noise floor is shaped according to the Equal-Loudness contour (Fig. I). It is fairly low at 4,000 Hz (I failed to point this out with my red line) and rises slightly towards higher frequencies.

Disadvantages of ATRAC compared to MP3

Fig. IV: Multitone signal, -6 dB, ATRAC 4.0, 24/44.1 (linear scale, Hanning)

ATRAC (in this case 4.0; DSP Type-R doesn´t differ) performs much worse. Quantization noise is stronger (70 dB, Fig. IV, MP3: -90 dB)... WideBitStream my ass! ATRAC has a datarate of 292 kBit/s to its disposal, the MP3 example has even less (256 kBit/s), yet it yields superior performance. But it should be clear by now that ATRAC also follows the Equal-Loudness contour (Fig. I), just like every other lossy codec. However, ATRAC & MP3 do a bit more: they completely remove high frequencies. MP3 did it in my example at 19 kHz, ATRAC at 17.5 kHz. They get rid of those frequencies because A) we don´t hear them well and B) because their encoding hurts the rest of the frequency spectrum. How can the presence of high frequencies distort frequencies below them? Because both codecs use a constant bit-rate, they cannot adapt this bit-rate (and therefore possible quality) to changing complexity patterns. MP3 is of course able to encode with variable bit-rate (VBR). But I used constant bit-rate only (CBR) to create fair comparison conditions. ATRAC is forced to encode everything with 292 kBit/s, no matter what. Should a certain signal require a higher bit-rate, pity, ATRAC simply can´t do it and has to take away more information at other areas, informations it might want to keep instead.

Advantages of ATRAC compared to MP3

Fig. V: 1 kHz sine, MP3 (Lame, 256 kBit/s), 24/44.1 (logarithmic scale, Hanning)

Fig. VI: 1 kHz sine, ATRAC 4.0, 24/44.1 (logarithmic scale, Hanning)

The situation reverses with less complex signals. Fig. V and Fig. VI exhibit that ATRAC performs superior to MP3 with simple signals. Quantization noise is more evenly distributed, resolution of 20 bits is retained. Which means: the less complex a signal is, the better can it be encoded by a lossy codec; apparently, this is even more valid for ATRAC. But it has another advantage: its time/frequency resolution, expressed in block length. Perceptual codecs need to decide between long mode and short mode. The latter enables any lossy codec to encode transients (very short & loud signals, for example the attack of a piano or handclapping). For short mode, MP3 uses a window size of 192 samples (or 4.3 ms). Now if a signal is shorter than that it is ignored, it ceases to exist because it's effectively 'invisible'. In comparison ATRAC has not one but two short modes. For lower frequencies it´s 130 samples (2.9 ms), for higher frequencies 65 samples (or 1.45 ms). In general, ATRAC is better suited to encode very loud and short transients, helping dynamics and precision.

Sonic differences between ATRAC & MP3

Ask 'normal' people (not audiophiles) how they perceive the sound of MP3 and you´ll most likely receive the answer that it sounds slightly warmer to them, that is, if they can hear any difference at all. Yet your basic audiophile will call its sound 'cold' and 'digital' when in reality nothing could be further from the truth (their reasoning: it must sound that way because it uses lossy compression). Listen, all you audiophiles out there: MP3 encoded music sounds a tiny bit warmer compared to the original. The reason is not that it removes any frequency content above 16 kHz (also stated by audiophiles), the one and only true reason is that it fails at encoding short transients responsible for dynamics, attack and precision. When ATRAC still was used regularly it too was described as sounding 'cold' (for example in German STEREOPLAY magazine 15 years ago). Again, this isn´t true. At least not for ATRAC 4.0 and ATRAC 4.5. Both sound significantly more pleasant and warmer than the original (ATRAC DSP Type-R changed the situation somewhat). Responsible for this mellow signature isn´t a too short window size. These shortcomings are caused by the ATRAC codec attempting to encode signals up to 20 kHz. All those years ago, magazines and audiophiles alike (the german STEREO magazine paramount among them) constantly pressed Sony to improve rendering of high frequencies. They believed that if you could retain frequencies from 16 to 22.05 kHz it would yield true audiophile sound. Bullshit! ATRAC would have profited extremely if they wouldn´t have listened to audiophiles, I will show you how.

Tweaking ATRAC

Fig. VII: ATRAC standard encoding (24/44.1, linear scale)

By trying to encode signals up to 22,050 Hz (see Fig. VII) ATRAC is losing precious available bits better reserved for lower, more readily audible frequency bands. This creates an overall pleasant sound signature by producing soft compression artifacts. Of course, this sound is far away from the truth. The final critical band ATRAC encodes is the frequency area from 15,500 to 22,050 Hz, if it wouldn´t be present anymore, ATRAC would have more bits to spare for lower frequency bands (0 to 15,500 Hz). Remember: the less complex the music is (a.k.a. less frequency bands), the better will it be encoded by ATRAC. For that reason we will erase frequencies beyond 15,500 Hz with an equalizer so that ATRAC 4.0 or higher doesn´t need to concern itself with these frequencies anymore! Will this sound muffled? No, it won´t since ATRAC encodes transient responses with utter precision (compared to MP3 which always sounds slightly blanketed depending on the material). Getting rid of frequencies beyond 15,500 Hz improves quantization noise as a result:

Fig. VIII: Multitone signal, -6 dB, ATRAC 4.0, STANDARD ENCODING, 24/44.1 (linear scale, Hanning)

Fig. IX: Multitone signal, -6 dB, ATRAC 4.0, 15.5 kHz CUTOFF, 24/44.1 (linear scale, Hanning)

Look at Fig. IX and compare it to Fig. VIII by clicking on one of them with the left mouse button and scrolling through both of them. Quantization noise floor has been lowered by roughly 5 dB - and only because frequencies beyond 15,500 Hz have been removed. Stunning result, isn´t it?

Fig. X: RMAA frequency response, ATRAC 4.0 STANDARD ENCODING, four passes

Fig. XI: RMAA frequency response, ATRAC 4.0, 15.5 kHz CUTOFF, four passes

Even RMAA recognizes the effect. Fig. X & XI depict that the hole around 4,000 Hz, typical for any ATRAC version, has almost disappeared along with the odd response at subsonic frequencies (20 Hz). Increasing levels from 10,000 to 15,000 Hz on Fig. XI are caused by my equalizer setting (see below at 'Equalizing ATRAC (costly option)').

Equalizing ATRAC (free option)

I told you that this tweak is free, I therefore searched, found and measured a suitable equalizer. This was difficult, not many free equalizers around are able to process with high quality. I will however also tell you about costly alternatives, namely SoundForge and iZotope Ozone. SoundForge is the VST-host while Ozone is the equalizer I work with in that case. I will talk about them because they are yield slightly superior quality. Never, I repeat, NEVER use built-in equalizers (SoundForge, WaveLab, Adobe Audition, foobar2000, Winamp). I´ve measured them and they create so many errors that it´s shocking. Anyway, to achieve the tweak without paying any money while still retaining high quality you´ll need these things:

- foobar2000 (get it here)
- a VST-wrapper (get it here)
- the equalizer EngineersFilter from RS-MET (get it here)

or

- Audacity (get it here)
- the equalizer EngineersFilter from RS-Met 

Fig. XII: RS-MET EngineersFilter setting for ATRAC Cutoff

Fig. XII reveals my configuration for the cutoff filter. The EngineersFilter offers several other filtering methods but I decided to keep it simple in order for less tech-savy people to use it as well. Regarding installation/setup of foobar2000, its VST-Wrapper and the EngineersFilter I cannot help you however, you need to figure that out for yourself, the same goes for Audacity. Other recommendations are: keep the signal at 32 bit floating-point, regardless if you´re working with foobar2000, SoundForge or any other digital audio editor. As you know, the MiniDisc is capable of working with high resolution material so if you´re recording from a PC just keep it at that high resolution. If you don´t want to use a PC I´d recommend a CD-RW (which can be erased and rewritten). In that case, decrease bit-depth to 16 bit without using noise-shaped dither (the shaped and dithered quantization noise would otherwise confuse ATRAC again) and burn the results to CD-RW.

Equalizing ATRAC (costly option)

Fig. XIII: iZotope Ozone 4.0 settings for ATRAC cutoff

Fig. XIV: iZotope Ozone 4.0 general setup (-> click 'Option')

First of all you´ll need a digital audio editor like SoundForge, WaveLab, Adobe Audition or Audacity. With these you´ll be able to load iZotope Ozone (in my case, version 4.0) which you will configure to the specifications pictured in Figs. XIII & XIV. The 1.5 dB amplification of frequencies at 20,000 kHz is optional and used by me to fool my ear into not recognizing that certain frequencies are alltogether absent. Why would you even use the iZotope Ozone EQ? Because it´s in my experience the best equalizer on the market, it doesn´t create phase distortions nor other distortions or errors and generally performs perfectly. Have a look:

Fig. XV: iZotope Ozone, phase response

Fig. XVI: EngineersFilter, phase response

The phase response sadly is very underrepresented when it comes to sonic differences between DSPs or units playing audio material. In this case it´s evident that iZotope Ozone has superior phase performance (Fig. XV) compared to the EngineersFilter (Fig. XVI), yet it is debatable if this is audible at all. Let´s be fair: the EngineersFilter EQ performs admirably compared to all the other free EQs I´ve tested. Impulses play a role too:

Fig. XVII: iZotope Ozone, impulse response

Fig. XVIII: EngineersFilter, impulse response

Fig. XVII depicts a perfectly symmetrical impulse response for iZotope Ozone. A high steepness of the cutoff filter produces better frequency resolution at the expense of perfect impulse response. The same is true for the EngineersFilter, although here the impulse response (Fig. XVIII) is modeled after the first CD players with analogue anti-aliasing filtering. In the end you have to decide, I´ve written it years ago that the effects of impulses are overrated. BTW, the settings I´ve described yield the following measurable results:

Fig. XIX: frequency response, ATRAC 4.0, 15.5 kHz cutoff (logarithmic scale)

Fig. XX: frequency response detail, ATRAC 4.0, 15.5 kHz cutoff

As you can see on both examples above which were created by RMAA I´ve achieved the desired effect - without frequency deviations created by crappy equalizers and, almost (for the EngineersFilter) without phase distortions. The graph depicting the zoomed-in frequency response (Fig. XX) reveals a not too steep cutoff, yet it´s precise enough to get rid of frequencies beyond 15.5 kHz. The result of my procedure is evidenced by Fig. XXI: the picture shows a spectogram derived from an ATRAC encoded/decoded recording (compare to Fig. VII). The precise 15.5 kHz cutoff is clearly visible.

Fig. XXI: ATRAC encoding with 15.5 kHz cutoff (24/44.1, linear scale)

The sound

When I first heard the results I couldn´t believe my ears, the sound had improved by such a margin that I was wondering how I had been able to listen to it before. Precision, attack, stability and holographic impression of the stage were sounding so much better now... But listen for yourself. The following files were recorded digitally with the Sony MZ-R 55 featuring A) the standard full-frequency and B) the 15.5 kHz cutoff discovered by me. In both cases, the original files were at 24/44.1. After recording I played them back using the Kenwood DM-5090 (also digitally) and recorded its output with the optical input of my Creative Labs Soundblaster X-Fi HD USB. After that I merged three 30-seconds examples and uploaded them to soundcloud. INSTEAD OF LISTENING TO THEM ONLINE, DOWNLOAD THEM! Reason: both are ATRAC-encoded/decoded PCM-files, encoded again with MP3 by soundcloud (at 128 kBit/s). Should you just press 'play' you´d only hear a transcoded file, revealing compression artifacts clouding possible differences. Downloading them however you´ll be able to listen to the pure, ATRAC-encoded/decoded, Kenwood-derived, digital files in pristine 24 bit quality without further influence from soundcloud. You would even be able, should you desire, to perform a DBT listening test; these two examples were edited with sample precision.

Three ATRAC 4.0 encoded samples, standard encoding

Three ATRAC 4.0 encoded samples, encoded using my 15.5 kHz cutoff filter

Epilogue

And? What do you say? I feel that the result speaks for itself. I admit that this tweak requires some effort but I think that it´s worth it. I now can use ATRAC 4.0 again! Oh yes, I almost forgot... why didn´t I use a more recent ATRAC version? While the effects will be superior using ATRAC 4.5 or higher, the ATRAC 4.0 equipped recorders I own (MZ-R 30, MZ-R 50, MZ-R 35, MZ-R 37, MZ-R 55) have high quality drives, producing MiniDiscs running without flaw on any other MD recorder / player. Later units (MZ-R 900, MZ-R 909) featuring superior ATRAC ICs fail to do this. With the exception of the Sony MZ-N 510 they all record with unreliable results. I also admit that it isn´t very convenient to use MiniDisc these days. The reason to use them, for me at least, isn´t their sound. Every other lossy codec employed today around the world is superior. I´m sorry, but it´s a fact. Still, I love those little discs. The players/recorders are of high build quality, sound well enough (in some cases more than well) and you get the joy of bringing some amount of 'slowness' into your musical life by occupying yourself with media you can actually touch. Let´s face it, I´m an idiot. An idiot... just like people still listening to vinyl. Like them I believe in an ancient and deceased format. But have I mentionend yet, that it´s pure joy? Oh, I did? Never mind! Anyway, with my tweak you´re able to prolong the lifetime of MiniDisc before it´s completely replaced by superior codecs and playback devices. And while you´re at it, use it in combination with the FiiO E07K, it´ll sound even better this way. Use this chance well and enjoy the results!

Last update: 06.09.2013