I'm trying to understand the concept of transparency when it comes to lossy encoding.
I know it is a subjective matter, but let's say I have a number of lossy encoders that are all able to produce results that are transparent to my ears.

According to my logic, if I am able to distinguish the original file from any file produced by transcoding an arbitrary number of times in any sequence, then one or more of the encoders are not truly transparent.

Am I completely out of line here?

No lossy encoder is truely, 100%, transparent.

It's just a case of finding one that is as transparent as it can be.

To Gray

yes as long as you remmeber to teste blindly. .e.G you do not know which file is being played.
otherwise you might risc having the placebo effect screw it up.

If you mean successive encodes of the same file, the loses are always cumulative. Even the best lossy process will soon start producing audible differences.

You can't say that.

"Transparent" doesn't mean the soundwave is identical to the original, it just means it sounds identical to a human (which is a totally different thing).

If you made a lossy encoder that made songs .000000000000000000000001db quieter than the original, there is no way that any human could tell it apart from the original, and therefore the codec would be transparent.

Transparency is perceptual and subjective. To say that something is "truly" transparent is kind of meaningless, since any encoding that you consider to be transparent now can theoretically become non-transparent at any point in the future due to a number of factors (you get better equipment, you train yourself through listening to hear more artifacts, etc.). Similarly, if you subject a recording to repeated transcoding with a certain encoder, it doesn't mean that the encoder is "not truly transparent." It just means that that encoding, at that particular moment, isn't transparent for you, for whatever reason (in this case because you transcoded it too many times).

lets keep things simple.....

if you can't hear the difference between the transcoded file and the original, then it is transparent to you

it is that simple

That's not a very useful encoder, though.

I agree with Synthetic Soul. No useful lossy encoder is truely, 100%, transparent.

And that's not a very useful statement either. The point of the 1e-20dB "lossy" encoder example is that there is transparency.



I don't think you understand the concept of "transparency." Transparency is the inability to determine a perceptible difference between a two signals. Even a person with near-godly ears will have some threshold by which audio information can be removed with no noticable lack of fidelity.

Think about a low-pass filter removing audio above a frequency beyond the hearing range of humans: 96kHz. The resulting signal is "lossy" but it makes no difference to you because you can't hear anything above ~20khz. Lossy encoders work on the same principle, only they remove inaudible signals within the frequencies the human ear can hear.

There are 100% transparent encoders. The real test is: can you tell a difference? If you can't, it's 100% transparent. Most people don't need 100% though, so the real tradeoff is how much space you want to save for audio data which may not contribute to your listening experience.

You could use the word 'transparent' to mean whatever you like, but I thought the standard usage was subjective; a three-place relation between an encoder/setting/environment, an audio source, and a listener. eg "encoder X is transparent with music Y for listener Z". No matter the X (whatever lossy encoder, whatever bitrate, whatever tweakings) there's probably going to be some Y out there (ie some problem sample) for which a given Z (ie someone with very good ears who knows what to listen for) will be able to abx it. If so, synthetic soul is probably right that no lossy encoder is 100% transparent.
But conversely, I'm far less confident that I'm that kind of Z! I'm guessing that my ears are so crappy that you'd be hard pressed to find a Y i'd be able to abx from the original if encoded with ogg vorbis q10, for example. It's probably transparent (for me).

Anyway that's more or less the way I've been using the word 'transparent' in this context. Perhaps I'm just speaking a language of my own...

That was exactly my point. Well put. You cannot state that a lossy encoder will be transparent for every sample for every person. And when I say "lossy encoder" I am talking about a useful encoder like LAME, MPPENC, etc. that produce well-compressed files at the cost of quality.

LAME -V5 is transparent to me, for all intents and purposes. Does that mean that LAME -V5 is 100% transparent?

When X encodes Y and Z cannot discern a difference between the original then Y can be said to be transparent to Z. We can say no more than that.

grav's post sounded to me that he was stating no encoder can achieve transparency because, following an infinite number of listening tests, he would be able to prove a sample to be non-transparent. That is why I stated that no encoder is 100% transparent, and that the solution was to simply find one that was transparent in the most situations to him/her.

Stop making sweeping statements with disclaimers attached. You are using ambiguous terms such as "useful" and "well compressed" which require context to be of any meaning. Limiting your scope to encoders which will always cause a loss of perceived quality loads the question, and generalized statements about lossy encoders should not be based upon that set.

There are 100% transparent encoders, because there are thresholds for which no human now, or ever, can blindly detect an audible difference between an original and lossy version of an audio signal, for all cases. This can be extended to any signal/signal-detector relationship. For audio, examples are an encoder such as Grand Dizzy suggested, or even resampling a gigahertz audio signal down to 192khz as a lossy "encoding" (with size-reduction ratios orders of magnitude better than any mainstream lossy codec). Quite useful, but maybe not applicable for what you need.

A sonically perfect lossy-codec for humans can be made, it just may not be as "lossy" as you'd like.

How about a little diagram?



HIGH
Origianl CD quality/lossless compression

Highest quality lossy compression




Perceived transparency by the world's best ears

Most mp3s

Perceived transparency by the worls'd worst ears





Very low quality lossy compression
LOW

I attempted to give context when I used LAME and MPPENC as examples, but you chose to ignore that. You have chosen to make a moot point about impractical and negligable lossy encoders, regardless of grav's first post.

I am not loading the question to make my answer correct, but purely limiting the scope of my answer to the real world.

I accept that my statement, when taken out of context, may not be 100% correct; but it is correct under the assumption of the term "lossy encoder" that is used on this forum every day.

Why haven't you popped up on any of the "What's the best lossy codec?" threads to enlighten us before? You would be very busy correcting us poor fools hourly.

If it bothers you so much to make the leap of faith that most people took when reading my words then I will retract my statement and replace it with this one:

And that's not a very good counterstatement. Nobody said that transparency doesn't exist.


Pardon me, but you're looking at the person who created the "Transparency (data compression)" stub on Wikipedia.


Now that's a very simple way to put it, isn't it? Just the kind of thing that would make audiophiles work up a rash.


There isn't a single real encoder that I know that doesn't suffer from artifacts at some bitrate. True, you can try increasing the bitrate till the artifact goes away or you can use lossless if that fails.

But really, all encoders have problem samples and can't be 100% transparent all the time.

Have you just claimed that it is possible for anyone in the whole world to detect artifacts in all files created by LAME or mppenc?

The "contraposition" of what you said is, "If all encoders are truly transparent, I would not be able to distinguish the original file from any file produced by transcoding an arbitrary number of times in any sequence."

I don't know what to say to that, but please read on.

Lossy encoders aim for two things; one is to reduce filesize, and the other is to maintain audio quality. You can't have a win-win situation here, and there's always a trade-off.

Let's say you have 101 files, from a[0] to a[100]. a[0] is the original, and a[1] is the first-generation copy which has been encoded with, say, LAME. You might or might not notice the difference between a[0] and a[1]. If you can't, then a[1] is transparent, relative to a[0], the source of the encoding.

Now, a[2] is the second-generation which has been made by encoding a[1] with LAME. You might or might not notice the difference between a[1] and a[2]. If you can't, then a[2] is transparent relative to a[1], the source of the new encoding.

The encoder that makes a[2] has no knowledge of what a[0] sounded like. It just tries to reproduce a[1] faithfully. The encoder that makes a[3] has no knowledge of what a[0] sounded like, and it just tries to reproduce a[2] faithfully.

If this goes on, a[100] might or might not be distinguishable from a[99]. However, I'm very certain a[100] would be very different from a[0].

That is a good point! I haven't looked at it that way.

I was thinking more in math functions, but since transparency is not "mathematically defined", that way of looking at it might not make sense.

Thanks to all for their view on the matter. I've read all comments with interest! :-)

Better to say, 'lossy encoded files can be subjectively indistinguishable from source'. Which doesn't mean all of them always are, in all conditions, for all people. It means that it is possible that a given person may be unable to distinguish a given encoded file from source, in a blind test. This has been documented to happen. It doesn't mean it always happens. It simply means it's *possible*.

There are some people, on the other hand, who claim that *they* can ALWAYS tell a lossy encoded file from source, by sound alone. THis is at least as far-fetched a claim as the claim that a codec is ALWAYS transparent.