Okay, I apologize ahead of time... I am a GM not an Engineer

[LA_Guy]

I respectfully disagree, Goran...

Lossy coding happens ONCE-the first time you encode the stuff. Once it's done, it's done.The MP3 and other algoritihms are designed to work on certain traits of human hearing. The stuff gets taken out and yes, once its gone it's gone, so it CAN'T be taken out again! As long as you use the same or better quality when you re-encode, nothing ADDITIONAL comes out. There's no reason for anything else to come out! Where this is not true is when you use DIFFERENT types or qualities of coding. THEN each encoder WILL take out things according to IT'S algoritihm and you will have build up. It's when you cascade MP2 and ATRAX, and MP3 and WMA and finally the coding that IBOC uses that you get build up-becasue each of these lossy codec systems works in a different way. BUT..if you keep with ONE system and ONE quality level you will be fine.
There are a couple of minor exceptions-if your A/D and D/A converters (and their associated analog circuitry) suck then the audio coming out of them will be a bit different, so some build up might happen that way. But if you use pro quality codecs and analog support circuitry, this will not happen.

 

[gtomas]

No, no, no... And it's very dangerous for people to even think that what you are implying might be true. There is enough confusion as it is...

Every time you do lossy coding, more stuff gets thrown out! The reason for this is decoding. I'm not aware of any encoder that can take an MP3 stream and operate on MP3 bitstream itself. How would that even work?? It's like saying once you record audio on an analog tape, it doesn't get worse with each subsequent transfer to another tape. But because it has to "leave the tape" (so to speak) to get recorded on another, every generation has poorer audio quality.

In coding something similar happens - you have to decode before encoding again. And when encoding again, the encoder has no idea what has happened before. They just see high bitrate digital audio that must fit into (low) target bitrate. Hence they throw information out to get there. Not only do they throw information away, they also add distortion and noise. In a way, lossy coding is all about manipulating noise (or bit allocation if you want).

If what you were implying was true, than you could encode MP3 20 times with the exactly same settings and the same encoder and get no loss in quality. If you still think that's true, it's easy to try. If what you are implying was true, than all the problems in the world of lossy coding would be easily solved by using a single standardized codec and single bitrate. Unfortunately, it doesn't work that way... Every coding (even with the same codec and settings) further reduces audio quality.


Regards,
Goran Tomas

 

[busyradioguy]

Agreed. The analog tape copy process Goran described, should help people understand this. To paraphrase, you can make numerous copies of the master tape, and the copies, known as slaves, will all sound basically identical (give or take, depending on how well the duplicating system is maintained and as the master tape wears out). But take a slave and treat it as a master whereby the duplicate slaves are of a slave master, and the quality quickly degrades. Audio encoding and decoding and encoding and decoding and encoding and decoding in chains, is akin to making a slave of a master, then a slave from the first slave, and so on.

 

[LA_Guy]

But that has nothing to do with the MP3 process. I've spoken to experts about this (who know more then you and me) and they agree with me. The degradation you speak of is due to the actions of the (imperfect) codecs changing the audio slightly during A/D and D/A conversion. Do 20 times and the CODEC build up becomes the problem-NOT the MP3 coding itself. The MP3 (and any other) algoritihm works by removing stuff-and if the stuff is already removed then it's not there to take away any more! It's as if I took white noise and notched out 1 kHz-once it's gone you CAN NOT notch it out again! It's gone, and will not re-appear by itself.

The tape noise example is completely invaild-tape noise is ADDITIVE, while lossy coding is SUBTRACTIVE! Again I say, once something is already GONE then it can not come back by itself!

When you encode and decode, the decoded audio does not magically gain resolution-and when you encode again, the encoder does not take things out that it does not HAVE to!
An MP3 encoder not think-it works according to its hard wired algoritihm-removing redundant information according to its pre-programmed model of human hearing (rumored to be based upon the song "Tom's Diner" by Suzanne Vega). If that information has already BEEN removed-it does not remove it AGAIN! According to your math 1-1 does not equal zero!

 

[busyradioguy]

But as Goran stated, the encoder doesn't "know" what has and hasn't been removed. Another way to explain it is that repeated steps of encoding and decoding, do not equal clones of each other. MP3 is "data" reductiomn, as analog audio deterioration during the analog tape dubbing process. Forget what the experts have told you, and run your own test.

 

[LA_Guy]

But as Goran stated, the encoder doesn't "know" what has and hasn't been removed. Another way to explain it is that repeated steps of encoding and decoding, do not equal clones of each other. MP3 is "data" reductiomn, as analog audio deterioration during the analog tape dubbing process. Forget what the experts have told you, and run your own test.

I think we're saying the same thing, but it's NOT due to the MP3 process itself...

Let's do some math:


Let's say the uncompressed WAV audio has a 'D' value of 100 (equal to its data).

Pass this through an aggressive MP3 encoder and 90% of the data gets tossed. Now it's D value is 10.

Decode it and the D/A decoder adds a small amount of non-linearity. Now the D value is 10.2.

Pass it through the encoder again and the D value gets reduced back to 10 -something further has been thrown away....and it might not be just the stuff that was added.

Decode it gain and now it's back to 10.2 again....then encode back to 10.

This is a form of build up, but it has NOTHING to do with the MP3 coding process itself! It has to do with the imperfect A/D converters (and their reconstruction filters).

Do this 30 or 40 times and OF COURSE the audio will be degraded...but again with PERFECT codecs it would NOT be!

The good news is as codecs get better the format improves as well....BUT it isn't the format that it inherantly bad-it's our real world IMPLEMENTATION of it!

PS: Barix codecs are now VERY, VERY good (compared to the early ones-the Instreamer and Exstreamer 100 are in their fourth hardware version) -and the Exstreamer 110 and new Exstreamer 1000 are even better then the 100s!

FYI....The newest hardware revsion of the 100's is: HW 1.2.

 

[CW]

You can argue anything theoretical.....but in REALITY, NO codec is perfect or will ever be such!! We don't live in a perfect world and never will...no matter how good the electronics, etc....even a 16bit PCM WAVE is not as good as the original analog audio....its just cleaner when played over and over or copied over and over.... Recoding ANYthing other than a WAV PCM will result in more degradation of the signal.....taking audio off a sat rcvr that has been compressed once and feeding it into a hard drive system that compresses it more will result in more distortion.....YUK!!
Yes it IS because of the process....NO process is lossless.....(other than leaving the audio alone in its original uncompressed format....and even then, why stop at 44.1/16bit??) Even every little stairstep of the PCM coding is not perfect....it is NOT an exact copy of the original....but only a CLOSE approximation of it.....in fact, telcos use 1004 instead of 1000Hz as testtone now because of the DISTORTION of the 1Khz tone by the 8kHz sampling rate.....on an real analog circuit (cable, etc), such a problem does not exist and 1000Hz is as clean as 1004....on a T1, you will have better S/N at 1004 than you do at 1000...and thats a technical fact. Been there, seen it...and this is with NO compression...

 

[gtomas]

Let's say the uncompressed WAV audio has a 'D' value of 100 (equal to its data).

Pass this through an aggressive MP3 encoder and 90% of the data gets tossed. Now it's D value is 10.

This "math" is not scientific enough to prove anything. Nevertheless, let me use it to point you to the error in steps:

Decode it and the D/A decoder adds a small amount of non-linearity. Now the D value is 10.2.

Decode it and it's bitrate is back up to 100.

But it's "acoustical" "value" is still at 10. Now encode it again to a bitrate of 10 and you have to remove 90% from the "acoustical" "value" of 10. That's how it works. Because it's psychoacoustical and exploits human auditory imperfections, it doesn't sound as bad as raw numbers would suggest, but nevertheless the loss of audio quality quickly becomes perceptible.

To imply that the AD or DA conversion is the main source of the audio degradation in codec cascading is ridiculous, apart from being totally incorrect.

The audio coming to the encoder (be it software or hardware) has to come in the form of decoded, uncompressed audio. Each coding cycle then removes additinaly more of acoustical information. Coding is in essence a form of noise manipulation - you allocate less bits (raising the noise) in the filterbanks you determine the noise won't be perceptible becasue you have a strong masking signal.

Please, look up on how the psychoacoustic coding works, because respectfully I don't think you understand it. You state yourself that you are not the expert, so why claim something you are not sure of? Please take into consideration that you misinterpreted what other people told you and/or they might not be experts you believe they are.


Regards,
Goran Tomas

 

[rrsounds]

I'll go along with the idea that going through multiple A/D and D/A conversions will degrade audio quality. But I think this is a bit of a red herring. Converter quality is improving almost daily, but even this is a side issue.

Because, honestly, how often does such a conversion happen?

I would answer that, by saying that typically the conversion from one file type or format to another (via compression or not) while remaining ENTIRELY in the digital domain occurs MUCH more often than going back and forth between analog and digital.

This process can be entirely transparent, (some file conversions merely change header information) or extremely detrimental, such as when data is reduced with perceptual coding.

IMHO, blaming the obvious problems of bad sound these days on A/D and D/A coding errors would be misguided. True, multiple instances of such conversions can add graininess and remove clarity, but they will NOT glaringly and dynamically alter the frequency balance the way perceptual coders can and do!

If quality is important, then repeatedly converting between analog and digital is, of course not recommended. But I would above all say that staying out of data compression, unless absolutely required, is a MUCH more impactful rule to live by! Especially important when you have no clear knowledge or control over any past or future perceptual codings which may be applied to the content of your signal.


Kind Regards,
David

 

[SML]

"Because, honestly, how often does such a conversion happen?"

Here is an example:

Field reporter records an actuality on a flash recorder using MP3 compressed format. Reporter sends file via FTP to NPR. NPR edits the file and includes it in the next newscast. NPR sends the reporters story as an MP2 companded audio file via satellite to affiliates. Affiliate records the file in a linear fashion, then sends it via a AAC codec-compressed Studio-transmitter link. Analog sounds OK after two passes through two different companding processes, a little gritty after the third pass and then it hits the HD transmitter audio compression system. Audio is now very gritty and un-natural sounding. A collision of different codecs and compression schemes.

It happens every day out there.

 

[rrsounds]

"Because, honestly, how often does such a conversion happen?"

Here is an example:

Field reporter records an actuality on a flash recorder using MP3 compressed format. Reporter sends file via FTP to NPR. NPR edits the file and includes it in the next newscast. NPR sends the reporters story as an MP2 companded audio file via satellite to affiliates. Affiliate records the file in a linear fashion, then sends it via a AAC codec-compressed Studio-transmitter link. Analog sounds OK after two passes through two different companding processes, a little gritty after the third pass and then it hits the HD transmitter audio compression system. Audio is now very gritty and un-natural sounding. A collision of different codecs and compression schemes.

It happens every day out there.

But honestly (I use the word again), how many of these conversions HAVE to involve analog/digital conversions? In your description above, as few as none, or maybe one or two?

Contrast that with the number of file/format conversions. Which is exactly my point. THERE lies the weakness. NOT in A/D, D/As!


Kind Regards,
David

 

[LA_Guy]

Please, look up on how the psychoacoustic coding works, because respectfully I don't think you understand it. You state yourself that you are not the expert, so why claim something you are not sure of? Please take into consideration that you misinterpreted what other people told you and/or they might not be experts you believe they are.


Regards,
Goran Tomas

I know quite well how it works. Shall I tell you? There are 35 or so frequency bands the audio is chopped up into. 44 thousand times a second, the level of each of these bands is sampled. If there is one or more bands whose adjacent band has a much louder signal, then those bands are silenced (no data) because human hearing only hears the dominant frequency. In this way, much 'redundant' data is removed.

However, when converted back to a WAV file, the stuff removed DOES NOT magically re-appear! It's STILL gone! Encode it again and there are no side frequencies any more. They were already removed! In essence, the audio has already been processed.

I am going to take a WAV file and run it back and forth through LAME a few times, each time looking at the file size. If it changes dramatically, then we know Goran is correct. If it doesn't then we know I am right.

 

[busyradioguy]

I am going to take a WAV file and run it back and forth through LAME a few times, each time looking at the file size. If it changes dramatically, then we know Goran is correct. If it doesn't then we know I am right.

That's NOT the accurate way to do it. If you really want to see what's happening, you must study the graphic images of the audio waveform, through a frequency analyzer.

Your approach will only show dramatic changes in file size from the standpoint that .WAV files are always much larger than MP3 files. This is not a case of taking stuff out then putting it back in. This is all about how stuff continues to get removed in the staggering chain of encoding and decoding.

 

[rrsounds]

I am going to take a WAV file and run it back and forth through LAME a few times, each time looking at the file size. If it changes dramatically, then we know Goran is correct. If it doesn't then we know I am right.

As you may have noticed, different MP3 encoders (and decoders) can sound very different from each other. That is because they use different proprietary methods to meet the MP3 spec, which is somewhat general. Even with identical settings, they do not give identical results, such as file size and/or audible effect.

So, because files rarely get re-processed through identical encoders, IMHO a test which uses the same coder and decoder over and over would be of only a limited usefulness to prove or disprove anything.


Kind Regards,
David

 

[The Beave]

I should know this, as I am a protools certified engineer, but going from WAV to AAC+ which there are a few streaming signals out there starting to do this, and they sound darn good, but what is the comp/loss ratio doing that?

Any guesses?

 

[gtomas]

I know quite well how it works. Shall I tell you? There are 35 or so frequency bands the audio is chopped up into. 44 thousand times a second, the level of each of these bands is sampled. If there is one or more bands whose adjacent band has a much louder signal, then those bands are silenced (no data) because human hearing only hears the dominant frequency. In this way, much 'redundant' data is removed.

However, when converted back to a WAV file, the stuff removed DOES NOT magically re-appear! It's STILL gone! Encode it again and there are no side frequencies any more. They were already removed! In essence, the audio has already been processed.

I am going to take a WAV file and run it back and forth through LAME a few times, each time looking at the file size. If it changes dramatically, then we know Goran is correct. If it doesn't then we know I am right.

How is the file size going to tell you anything? The file size is determined only by the bitrate and the length of an audio. The file size tells you nothing about the changes in audio itself. You may encode one 4:00 song with an MP3 @ 192 kbps and you will get the same file size as completely different 4:00 song. You may encode one song with 192 kbps constant bitrate with MP3, AAC, OGG or RealAudio and they will all give you the same file size. kbps is kilo bits per seconds. X seconds will equal Y kilo bits (or MB) regardless of the codec, what happens to an audio or what audio actually is.

Studying percetpual coding is not trivial. You can't easily measure it, hence all the blind listening tests... Logic I've removed this, so I can't remove it again can't be applied here. In part because nothing is removed, but replaced by noise. Better said "drowned" in noise.

You are still missing the key point in your thinking that the bitrate gets increased when coded audio is decoded. After decoding the audio is back to it's original bitrate (usually 1.411 Mbps). To get it encoded to 192 kpbs (or something) again, requires separate process of perceptual coding that is agnostic to whatever happened before.

The tape analogy I've used before seems to hold well. To be able to transfer audio from one tape to another, you have to read it from the tape (with all the noise and distortion). When you record to another tape you add more noise and distortion of the second tape. You can't say well I've put it on the tape once, it doesn't matter any more which copy is this because it's all on the tape! You have to read audio from the tape to put it on another tape. In coding, the equivalent of this reading is decoding.

The problem is with tape you can use single tones and the usual stuff to measure the noise, distortion, flutter, etc. Because the perceptual coding is nothing like this, you can't actually measure it easily. You can't measure the quality loss of cascaded codecs with an instrument, file size or what not. You have to rely on subjective measurements like listening tests. That is the only way to objectively quantify it.


Regards,
Goran Tomas

 

[Holder]

Hi Goran

Thank you for your explanation was very clear and interesting .

Regards ,Holder

 

[bossbill]

Don Mussell hit it on the head a few posts back.

BTW, anybody ever hear those Sears commercials on the radio that have gone through a few analog to compressed digital back to analog back to compressed digital, etc?

Ever use a minidisc to record off the network from a satellite receiver and then play back through a compressed digital stl?

Fact is, whether it's cascading algorithms (which is a real phenomenon) or imperfect A/D converters , the outcome is still inferior and a good ear can hear the difference. An average ear maybe can't consciously tell the difference, but the brain still can. And that's the digital form of listener fatigue. In earlier times, we had IM distortion which caused listeners to tune out. We would do our transmitter proof and measure harmonic distortion and think everything was fine but something didn't sound right...

WHY WHY WHY are so many trying so hard to justify inferior audio? Just stay linear. Use your IP setup for BACKUP, not your main. Uncompressed digital STL if you have a clear shot, or conventional analog if budget doesn't permit.

Next thing you'll be justifying cellphone remotes, pretending they don't sound like crap, and pretending the listeners can't tell the difference. By the way, that's where you use compression, so you can use a pots line and sound a lot better than a cellphone.