'FeralA' decoder -- free-to-use

[astrotoy]

John, thanks so much for all your work and your detailed explanation of how to use your decoder.  The first step I will do is convert the files to 192/24. I have a Pacific Microsonics Model Two and Pyramix software, so I normally get really fine results in my digital rips.  I'll then try the decode on the files.  

 

Thanks again,  Larry

[John Dyson]

About the FA mode V3.0.10L -- IT IS READY!!!

I promise -- V3.0.10L is now ready, and will not be reascended for at least a week or so.  If anything new comes out, I'll add it instead of replacing tje V3.0.10L version.

 

Location of snippets:

https://www.dropbox.com/sh/i6jccfopoi93s05/AAAZYvdR5co3-d1OM7v0BxWja?dl=0

Subdir with decoder:

https://www.dropbox.com/sh/5xtemxz5a4j6r38/AADlJJezI9EzZPNgvTNtcR8ra?dl=0

 

This has been a tricky release (like all are.)

The snippets are in the directory above, and the decoder is in a sub directory inside of this one.

There is a new 'Start Using' document for V3.0.10L also.

 

My hearing SUCKS -- tell me if you think that the highs need to be increased or decreased -- I have a few little LEGAL/'following-the-rules' modifications that can be done, but I did try the best choice from a straightforward engineering standpoint, but MIGHT BE IN ERROR.

 

I did not want to delay the release, but the fix is a true correction instead of a hack.   Everything now is 'following the rules', and as far as my hearing allows, doesn't seem to have any extreme problems.   Let me know if you hear any really bad stuff, and I'll add a correction and make the corrected version available also.

 

Here are the bugfix notes for the recent fixes and attempted corrections:

 

BUGFIX from V3.0.10E to V3.0.10L

A couple of problems overlooked/corrected and a true, final mitigation for the overly metallic sound. The ‘metallic’ sound finally was found to be an embarrassing pre/de-emphasis EQ error. There was way, way too much super-HF pre-emphasis, which made any compensation impossible. Also, perhaps just as important and just realized – the anti-distortion mechanism can not be used in the in-between layers EQ. The anti-distortion scheme works very well, but slight errors in my version of it can creep up when using the scheme for more than a few EQ in series. Since the in-between HF EQ is done 7 times, the small percent-level error in my implementation ends up being noticeable in the 0.5dB to 1dB level of total errors. Such errors are simply not tolerable. This series of V3.0.10 releases has been stuttered too many times, and hopefully this is the last fix for a while.

 

ADJUSTMENT from V3.0.10D to V3.0.10E (valiant attempt to fix the problem -- but slightly misguided.)

Still not happy with the V3.0.10D change where 2^nd order EQ was used. The result was too soft/’rounded’ kind of rolloff. Didn’t result in a ‘clean’ sound. Therefore changed the 2^nd order EQ (totaled as -3dB at 21kHz) to be a single first order rolloff of -3dB at 21kHz. The resulting sound is much more clean, but the rolloff actually starts lower in frequency… Kind of odd that more of a lower frequency rolloff results in more clean, precise sound, eh? This EQ stuff is strange and sometimes counter-intuitive.

 

[John Dyson]

* This message is intended to assert that V3.0.10L is a real release, no matter if a new one is produced.   The V3.0.10L info will be maintained for

approx 2wks, even if a V3.1.99ZZZZ is created.  I intend no releases for today as the next day or so will be used for 'data collection' and 'decoding recordings', but not ' program coding.'

 

V3.0.10L seems to be 'solid', but with the caveat that my hearing does not allow me to be critical in certain areas.

 

Specifically, I am a little worried about the 'dynamics' or 'dynamism' of the higher frequencies.   There are some things that can be done to increase the strength of the dynamics, but the settings become more and more fragile.   The scheme/settings that I used in the V3.0.10L release is very 'vanilla', without any flourishes.  There is the possibility of adjusting the dynamics of the upper-middle highs and highest highs.   For example, using asymmetrical pre and de-emphasis can retain almost the same general sound, but the dynamics can be ratcheted upwards by a step or downwards by one step.   I believe that it is very wrong to make such changes without a compelling reason - yet, I am unable to choose.

 

The 'dynamics' just might be perfect, so I am not intending to imply that here are any known defects, but it is just that I do not know.

Frankly, I am VERY happy with the decoding results -- more so than ever before, but recognize that my judgement is poor.

 

Thanks!!!

John

[John Dyson]

Part of the purpose of these releases is to help bring the project to completion/perfection.  Unfortunately, there are still some technical choices that need to be made.    There is a parallel release (probably V3.0.10M) coming out in the next day or so which will be very similar to V3.0.10L.   The 'M' version is not an improvement over the 'L' version, but instead it is a 'choice' whereby users can help me judge which is better.   Eventually, we'll chose if we want to use the 'M' technique or the 'L' technique.  BOTH VERSIONS ARE VERY SIMILAR.   Think like the comparisons in the optometrists office, but I sure hope this won't need to be done very often (maybe this is the last hurdle?)

 

 

Hopefully, someday soon, the specs for the FA scheme will become public, but until that happens, we might still need to make some of these 'choices'.   This second guessing can either be a chore or a challenge.  Myself, I have accepted this as a challenge...

 

The difference:  the 'M' (or later) release will be similar except for having stronger dynamics than 'L'.  Stronger dynamics does not necessarily mean 'better' dynamics, just simply is a difference.   Since I cannot hear differences above 9kHz for much of the time, the eventual users of the decoder might have to help.  With some luck, I might have some 'clear' hearing sometime during this phase, but we cannot depend on it.

 

The 'M' version will include the 'demos' on the public site and all of the normal stuff on the other sites.

The public site will have BOTH the 'L', 'M' and up to one other version to choose from.   I am not planning on a third version, but it might happen.

 

Again, this is NOT a replacement for the V3.0.10L version, but just might be the better choice for the future, either V3.0.10L or V3.0.10M might be the 'best' version.

 

* This comparison version MIGHT NOT BE V3.0.10M, but will be something after V3.0.10L.   I increment the letter on every major test edit/build to keep from losing control, so I cannot really predict the exact next subversion letter ID.

 

 

[PeterSt]

3.0.10L-0 is the better one over 3.0.10.M-0.

 

Yesterday I found 3.0.10L-0 already quite good in absolute sense and today I right away found 10M the lesser one.

Later, out of the blue my wife told me that the music sounded "sharp". Indeed I think this is true, especially after presenting her the 10L from yesterday. She told me without any hints:

-Warmer

-Less dynamic (ha !)

-Less pumped up.

 

The pumped up would be about "stressed" (overdoing its best to sound right and with that fails on everything).

The less dynamic is one you are looking for BUT:

 

Yesterday it was my observation that 10L could show less of that. Although I don't know the examples any more, but "ragget" would be my description of the highs at times. Say the opposite of silk.

 

Generally 10L lacks punch and 10M lacks even more of that. I am confident there is no difference in the low-mid in both, but the more highs in 10M depicts it.

The lacking of punch seems to be prevalent in all of the 1 million versions so far. To me it seems that this could be your means to perceive a more clear sound. Clear = less distorted, or the better perception of that when so. Thus:

 

Fuzz it up !

[John Dyson]

The 'lacking punch' was the reason for the attempted improvement when moving to 'M' from 'L', so I missed the mark.   I will be working on it...   I *know* that there is still something not 100% correct, and I will further process (think about) your comments over the next few days.

 

I have been getting good feedback about 'L', but not as good about 'M', so your results somewhat correlate with others.  I'll revert my mainline to 'L'.  (Because I don't use version control like 'git' or even 'cvs', I keep every version online at all times, and simply assign a certain version as my mainline.)   The convenient factor of keeping everything online all of the time is that I can refer back to old versions and run them directly.  (Alas, the big problem is remembering the command syntax because of my poor docs!!!)  I am a 'hoarder' of sorts, so much keep everything visible at all times -- thus the reason for avoiding version control and the obscured versions of the code.

 

After my rants and commentary, I do think you understand that for the results to be correct, I MUST follow some kind of 'rules', otherwise the results will be muddied up by a 'tweakfest'.   In the short term, 'following the rules' makes getting results more of a challenge, but in the longer term the results will be more correct.

 

Unfortunately, this project has been VERY VERY long term, and would have stopped a long time ago if I realized the length of the journey :-).

 

Will 'stew' about the current state and the recent feedback, and will find valid ways to improve the results, mostly by attempted mind-reading of the designer of the FA scheme. :-).

 

About the 'fuzz' -- I don't think we really want more fuzz -- but I understand the humor, but I am starting to think that the transients are somehow suppressed.   On the DolbyA decoding mode, the decoder is very noticeably more intense, so the problem isn't directly in the dynamics processing.   I really do think that this is still a pre-emph/de-emph incompatibility.  MAYBE EVEN A MISSING OR MISPLACED LAYER?!?!?!?

 

I think that the missing 'fuzz' that you describe is what I call 'slippery' sound.   It is like the sound is too 'lubricated' so the sound slips and is slimey, but clear in a 'soft', 'clean' way.   'Being dirtied up' or 'Being fuzzed up'  is an interesting, but apt way to describe it :-).

 

Will be spending at least several days to a week on this -- to avoid wasting people's (your) time with micro-changes.

 

 

[John Dyson]

Follow on to my previous comment responding to @PeterSt-- gawd, people have better hearing than myself.

 

I just checked on the parameters for decoding 'Supertramp' recordings, and recognized that the highest level decoding layer was -29dB down.   That is way too far down, and leaves the higher levels of transients compressed with FA compression.   Clean FA compression does sound slippery and slimey when it isn't mangled together with the 7 other layers.  The 'dynamics' of FA comes from the layers beating each other up.   Basically, we were left with 10dB of slippery, slimey FA compression that hadn't been dealt with properly.  The results sounded good to me because the modulation distortions were mitigated as far as they were, but there was insufficient processing to fully undo the FA damage.

 

Even though I did realize (and test/verify) the fact that the top layer of processing had been missing, that doesn't mean that I can produce a corrected version right away.  Such a profound change as re-organizing the layers requires re-testing on numerous recordings to understand if their layering is still similar to each other.

 

Thank goodness.   Basically, both the 'M' and 'L' versions were missing the top (most important) 10dB of expansion.  Also, my attempt to improve the dynamics by creating the 'M' version was without technical merit and in vain.   The asymmetrical pre/post emphasis would not be needed if/when the top 10dB of expansion is inserted into its rightful place.

 

THANKS to @PeterStand others who have privately provided good commentary.  (The only reason why I don't divulge their names is a respect of privacy, not keeping secrets at all.)

 

 

[John Dyson]

Follow-up on the previous comments/evaluations...   Some of the comments might not be operative on all recordings.   Supertramp, for example, is processed profoundly differently than many other recordings.

 

However, the benefits/corrections attained by making Supertramp correct might help other recordings.   Supertramp recordings are especially egregious, with a very different layering pattern.   Just now, I have written a new layering scheme for experimentation purposes.   Instead of the default behavior when using '--fa' (and the secret '--fb and --fc' switches), there is a literal scheme just added, where the layering is explicit.   For example, most normal recordings when using '--fa' would be equivalent to the new '--fs' command with the following full spec:  '--fs="0,1,2,3,4,0,1"'   On the other hand, during these current tests, Supertramp recordings don't work well with that spec.   My initial tests using the new 'experimental' version of the decoder and using the experimental '--fs' switch, Supertramp recordings appear to be happiest with: '--fs="0,1,2,2,3,4"'.   My first estimate of this setting was done by running two FA decoders in sequence using '--fa=3' and then '--fa=3' again, but with different calibration numbers.   When that seemed to be going in the correct direction, then I added the '--fs' command to further the experiments.

 

The actual, complete solution isn't going to be ready for a few days -- but AGAIN -- not all recordings might have the same troubles as talked about in other comments.  The four Supertramp recordings that I test with are very different 'animals' from the normal 'Simon & Garfunkel' or 'ABBA' stuff.

 

No matter what, I WILL be double checking a lot of my test recordings using variant '--fs' sequences.   My 'general' '--fa' command just might not be getting ideal results on other

recordings also.

 

I promise you -- I am very very critical of the decoder, and all I need are cogent, constructive and kind comments to benefit the project...   This helps to direct the work towards true completion, which I thought was done over 1yr ago!!! :-).

 

John

 

[John Dyson]

Have one more status report before 'radio silence' for at least a day or so...   I found some additional possibilites for better pre/de emphasis.

 

Since the complaint was basically a lack of dynamics and there is really no additional 'expansion room' to add more dynamics, it all has to be attained

by pre/de-emphasis working the expansion a little differently -- basically to move the highs further into the expansion region.

 

I sure wish we had specs, but since we don't, a lot of difficult things need to be done by listening.   The correction might be the '--fs' switch concept or might be more severe pre/de-emphasis to push the highs further into the expansion region.

 

The frustrating thing about processing the highs is that even a primitive DolbyA HW design doesn't produce noticeable 'grainy' distortion, but instead increasing the expansion will increase the fog, no matter if correct or not.   This FA decoder won't even normally produce much fog, so there is  doubly no indicator (tell) for correctness.   The only way to determin correctness is if the dynamics are correct -- really tricky to do.

 

The adjustments need to be made iteratively, and since I know the kind of thing that is being criticized, I can focus on that mattter for a few days!!!

Will report back when I have more solid information about the improvement.

 

 

[John Dyson]

New demo of upcoming improvement!!!

Great news -- found the problem with too little expansion.  In fact, the 'fix' was again re-inserting a module that I had previously removed.   There was a midrange specific pre-emph/de-emph even before the HF pre-emph/de-emph.   I had removed it because it seemed like it broke a lot of things.   Since I seldom remove code, all I had to do was a minor update, and insert the new pre-emph/de-emph back into the chain.   I didn't have to add another layer of expansion, and also could revert to something very close to my original HF pre-emph/de-emph.   The MF pre-emph/de-emph goes hand in hand with the per-layer pre-emph/de-emph also -- the midrange (1k to 3k) is a special beast because of the 10dB difference in the compression/expansion curve threshold on DolbyA/FA.

 

The resulting decodes sound more natural, less edgy sound but stronger, much stronger dynamics.   I think that this gives everyone closer to what they might expect, in fact, I believe that this is yet another major improvement.   The dynamics are MUCH stronger -- I did a comparison between '10L' and the current experimental '11B', and this makes '10L' sound very anaemic.

 

Most of the distorted sound on Elton John snippets is gone, and the dynamics of most vocals are much more naturally strong.

Below is the public review site.   Private request demos of the 'demos snippets' will be updated within an hour.   Other, on request demos will be coming after tonight.

I am not ready to release a V3.0.11B type decoder until I do a LOT more testing for dynamics problems.  (Of course, frequency balance is impossible for me, but I believe that the frequency balance is fairly close to the 10L type decoder.

 

The site below is NOT the normal site, but are special V3.0.11B review  snippet demos;

https://www.dropbox.com/sh/ho83ztnv9fmuyoi/AACYFehMVI2DgDJ-7oSCbCWha?dl=0

 

 

[Skip Pack]

I agreed with @PeterSTs comments on the dynamics with the sense that, while they were still an issue, they have steadily improved with the frequency balance was good the the clarity was distinctly better than the input. With this last group of snipps, I think the improved dynamics are a jump in the clarity department as well as sounding richer and more vital. Not helpful vocabulary, but I think this is the first time 'Orange Colored Sky' has not had gross problems either in the normal sections or the ridiculous blasts. It all sounds like music you could hear someone perform in real life this time. 

 

I hope you can give any percieved minor flaws a lot of time before you chase them. We'll need that time to give  you more considered feedback. We might want to take months with this one.

[John Dyson]

5 hours ago, Skip Pack said:

I agreed with @PeterSTs comments on the dynamics with the sense that, while they were still an issue, they have steadily improved with the frequency balance was good the the clarity was distinctly better than the input. With this last group of snipps, I think the improved dynamics are a jump in the clarity department as well as sounding richer and more vital. Not helpful vocabulary, but I think this is the first time 'Orange Colored Sky' has not had gross problems either in the normal sections or the ridiculous blasts. It all sounds like music you could hear someone perform in real life this time. 

 

I hope you can give any percieved minor flaws a lot of time before you chase them. We'll need that time to give  you more considered feedback. We might want to take months with this one.

Thanks.   I have been careful to avoid publically quoting any private comments made by anyone, but I do thank you for the continued criticism.

Right now, as of one version a step beyond what I just demoed (from 11B to 11C), I am running out of flaws that I can hear.

 

I think that all three of us (and a few others) have been worried about the dynamics, but all too often I run out of ways to fix the flaws.   In this case (like a few other previous), some of the 'fixes' were already in the program, but I had disabled them because I thought that the modules were damaging the results.

 

As the bugs in each section had been fixed, some of the sections that I had removed have been able to be re-inserted.

 

In the last few months,  there have been 3-5 major problems (depending on how you look at it) in the code:

 

1)  LF output EQ & LF inter-layer EQ

2)  LF pre-emphasis/de-emphasis

3)  HF output EQ & HF inter-layer EQ

4)  HF pre-emphasis/de-emphasis

5) MF pre-emphasis/de-emphasis

 

At the beginning of about 2-3months ago, 1->4 had plausible, but incorrect versions of the associated modules.   item 5 had been disabled for about 6 months, but was originally essentially correct.

 

Each of the 1 through 4 had been adjusted to be more correct.    The item 1 changed from all EQ as output EQ to LF EQ being the correct mix between inter-layer and output.  Item 2 was added to correct the sound of pianos at low level and to generally correct the bass at low and high levels (keeping the bass sound the same vs different levels.)   The same thing on item 3 as item 1, except for HF.

Once items 1, 2,3 got fixed, the results started being more plausible.

Item 4 being fixed is where we were as of a few days ago.

Item 5 restored the dynamics to be where we are now.

 

All of this progress was impeded because I had to make many A vs B choices, but could not do so reliably on immediate demand.  So, I started being very careful as to when my my hearing is accurate.    My HF hearing is accurate at different times than my LF hearing (odd, eh?)

 

The last step (5) came as I had realized that the lack of dynamics might be because the midrange (1k to 3kHz) levels might not be correct relative to the 3k->inf and 20->1kHz bands.  On a lark, the MF pre-emphasis/de-emphasis was re-inserted, then the dynamics strenghtened.

 

A lot of these final versions had already existed 6 months ago in almost exactly the same form, as I had predicted that these blocks might be needed, and had estimated what the blocks should do.   There have been many, many possible choices based on 2nd guessing the choices of the original designer.

 

The number of possible choices and possible directions have been mind-boggling.   Some of the choices had 'lesser quality' aliases, where plausible, but incorrect choices 'work', but produce suboptimal results.

 

I'll produce more complete examples/demos and slightly more up-to-date some time tomorrow.

 

Thanks again for the feedback/help that you have been kind enough to contribute.

 

[John Dyson]

After a little thought and review, there have been some incremental improvements/accomplishments in the last iteration, but after doing several A/B comparisons, I still havent been 100% happy.  Still haven't gotten the extreme transparency that I remember from the past.   I am really good at sensing transparency, and the 'decoding' shows an improvement, but doesn't quite meet my new goals.  (My older goals were less strict, with the feedback that I have gottten over the last year or so, now have become more picky.)

 

On a lark, I enabled 8 layers instead of 7 (easily done by using --fa=+8 instead of just --fa.)   After a few initial tests, it has become very obvious that the previous improvements plus the proposed addition of the 8th layer seems to be making further major improvements.

 

Running 8 layers instead of 7 can slow down the decoding a little bit (probably 15% slower), but the benefit is likely worth it.   This might further dissuade using one of the highest quality modes, but there are some speed improvements coming in the anti-MD arena also.

 

Demos with the recent improvements & 8 layers will be available probably in about +12Hrs.   Even though the demos can easily be produced more quickly than 12Hrs, need to consider the possible need for a redo.

 

The intital 10 second result from one recording gives an initial sense of wonderful depth, but that might be an expectation bias of some kind.

 

The proof will be in your own personal review.

 

 

[John Dyson]

V3.0.11N -- nicer, more real sounding.

NOT 'flat-image' sounding at all.

 

Another perhaps interesting step.

1)  Another layer of expansion (done correctly, of course).

2)  Modified pre-emphasis/de-emphasis.

3)  No increase in distortion, perhaps perceptual decrease in distortion.

 

The result is noticeably stronger dynamics, less 'slippery' sound.

I am a little worried about the HF, wondering if it might be a little weak.

Bass has been tricky -- the line between 'muddy' and 'insufficient' is very very narrow.   I did everything possible to widen the bass

response to below 20Hz while not making the sound muddy.

 

The stereo image is MUCH MUCH stronger, in some cases can be 'a wake-up'.

 

https://www.dropbox.com/sh/4vv8hg9xuavhtr8/AAD5sGCgDu1KCxbX0d0-0ZHKa?dl=0

 

[John Dyson]

32 minutes ago, John Dyson said:

V3.0.11N -- nicer, more real sounding.

NOT 'flat-image' sounding at all.

 

Another perhaps interesting step.

1)  Another layer of expansion (done correctly, of course).

2)  Modified pre-emphasis/de-emphasis.

3)  No increase in distortion, perhaps perceptual decrease in distortion.

 

The result is noticeably stronger dynamics, less 'slippery' sound.

I am a little worried about the HF, wondering if it might be a little weak.

Bass has been tricky -- the line between 'muddy' and 'insufficient' is very very narrow.   I did everything possible to widen the bass

response to below 20Hz while not making the sound muddy.

 

The stereo image is MUCH MUCH stronger, in some cases can be 'a wake-up'.

 

https://www.dropbox.com/sh/4vv8hg9xuavhtr8/AAD5sGCgDu1KCxbX0d0-0ZHKa?dl=0

 

Important comment -- there might be an excess layer of HF cut, so the highs might be a little suppressed.

A new version will be created later on today.

Sorry for this 'stutter', but just realized the problem after some feedback.

 

 

[PeterSt]

18 hours ago, John Dyson said:

A new version will be created later on today

 

Hi John,

 

I did not keep track of the latest version relative to your last post here. I suppose 3.0.11N-0 is not a newer one ?

Not at all to rush you, but currently I am waiting, and you could be waiting too. 😊

 

-----------------------

 

One other thing; I tried to tell you that the dynamics were too high (the M version already a tad and the L version more so), but do I read correctly that you want to add more of that ?

IOW, maybe you misread my Dutch.

 

Those too high dynamics should be in the 1000-2000Hz range - might that help.

And otherwise no worries. ... You could be trying to solve it with more dynamics UNDER 1000Hz (I feel that that could help just the same so the higher freq dynamics become less profound).

N.b.: Not many people may know, but transient speed in the mid range is way more important and interesting then the higher frequency (e.g.) needle ticks. That usually leads to "fake" resolution. The mid range speed is quite another beast.

 

Peter

[John Dyson]

1st -- the new version will be infinitely more dynamic than previous.   10M and 10L are both bugged similarly, just with different building-block choices.   Let me explain what I just realized this morning - after describing the next step:

 

The new version will be in the V3.0.12X series, and is TOTALLY different (much less crazy-intense.)

I am working on some correspondence on something that might really help the project, so am a little delayed, but am working on the new '12x' version.

 

Here is the revelation/admission -- IMMEDIATELY after waking fully up after a good sleep, I listened to the current test version (without the new fix), and found that the results sounded like HELL to me.   I am assuming that my hearing first in the morning is best on the HF.   This morning, when listening to the FA originals, I can easily hear the highs, where the FA originals later in the day sound very suppressed/muted.   However, when listening to the recent decoded versions, they sound 'better' in the dynamics sense, but also sound like HELL because of much HF EQ.

 

When listening to the bad EQ, I noticed that it sounds similar to the effects of the HF pre-emphasis/de-emphasis, so I turned it all off.   The LF pre-emphasis/de-emphasis and the MF pre-emphasis/de-emphasis are still totally necessary -- I have proof/tells that they are totally critical.   However, the HF pre-emphasis/de-emphasis just might be an artifact of my bad hearing, which degrades quickly after an hour or so in the morning.   Normally, I cannot very easily detect this change because I don't normally sleep much in a 24Hr period, but for some reason I started sleeping above 6Hrs per night.   Now, my hearing is VERY sensitive in the morning.

 

*The HF pre-emphasis/de-emphasis got added because of artifacts that I heard -- but are probably misleading.   This is probably one reason why the HF EQ had kept changing so much.  My hearing was varying over different parts of the day.  THIS IS A HUGE CHANGE through the day -- probably 6-10dB or more at 9kHz.

 

The result of this is that the next demos will have ALL of the HF pre-emph/de-emph EQ disabled, except for the necessary inter-layer EQ (that keeps the gain stable between each layer, or chaos ensues.)

 

In a way, this might only be a test, but also might be the best version yet.   The last of the 11x versions had the expansion added, but the most recent 12x series will have the new pre-emph/de-emph EQ that is basically totally disabled.  (It would be nice to remove any part of the code, if possible.)

 

* The decoder has LOTS more code in it than is currently being used.   As building blocks were used for testing, and they didn't appear to be applicable, they weren't removed, but instead were disabled.   This is why the MF pre-emph/de-emph, which eventually was determined to be totally necessary, was so easy to add -- it was already in the code.   Even the *exact* gain and frequencies were already in the code.  My engineering predictions were correct, but other bugs caused the correct code to act incorrectly.   Some day, before source disclosure, the decoder needs a massive house cleaning.

 

 

John

 

 

[John Dyson]

Here are some demos as a status report.

 

Lots of minor fixes/improvements.

Don't expect the highs to be exactly the same as the FA originals.  As far as my hearing can detect, vocals might seem a little

stronger and sound more intense, but much of that comes from the expansion and recovery of dynamics.  Remember, FA

material is fairly heavily compressed in layers, where general levels are maintained, but near those levels there is compression.

It is wierd, but does have the effect of squishing the dynamics in a more subtle way than a normal expander.   This difference in

intensity appears to be in the 1-2dB range at most, and is plausible considering the expansion.  The dB or so stronger highs have

been slightly mitigated by some ad-hoc EQ (yech!!!)  This was done because of a demo to a potential well known contributor.

 

THIS IS ONLY A STATUS REPORT SET OF DEMOS...   NOT FULLY COMPLETE and ONLY SNIPPETS:

https://www.dropbox.com/sh/sw78c7wzewwskmp/AADr90Ax_9NOfFS5JsRNV4jca?dl=0

 

 

1)  Minor improvement in dynamics.  I figured out exactly how to handle the HF pre/de-emphasis, and the major importance is to soften

the dynamics in the 3k to 6k range.   If the dynamics are left too strong, the sound can be unpleasantly intense.

2)  Better bass.   The exact tuning of the bass has been tricky.  Since I don't know the specs, it has been a balancing act.   There needs to be the correct boom vs thud, while not obscuring the vocals.   (The boom is approx 70-100Hz, obscuring vocals 200-500, but includes down to 30-40Hz, thud is in the 15-30or40Hz range.)   Making these correct with 1st order EQ (and a special 1st order EQ at 75Hz) is tricky with poor hearing.

3) Revisited all of the inter-layer EQ.

4) Minor modification of output EQ where the higher frequencies in the 9kHz to 15kHz range are de-emphasized by -0.75dB.   This helps to give a more similar balance

as the RAW FA versions.  The vocals will still generally be stronger/more forward in the decoded versions.   I believe that the need for the de-emphasis is the accumulated

error of 8 DA decoders in series.   A true DolbyA won't do this, but has other, more profound problems.  (The sound being mucked up severely with fog.)

 

The major issue right now is the LF pre/de-emphasis.  I believe that the latest HF pre/de-emphasis changes slightly threw the LF pre/de-emphasis out of whack.  The problem doesn't manifest very often, slight errors are not usually noticeable.   The LF pre/de-emphasis isn't too far off, but you can notice it at the beginning of 'Even in the Quietest Moments' as it fades up.   With the LF pre/de-emphasis being totally incorrect, a calibration of --coff=-7 was needed, but now the calibration is a more normal --coff=-4.0.   The only problem is now there is some modulation/expander effects during the fade-up.  This doesn't manifest on any of my other test recordings including 'Year of the Cat' or any other piano pieces.  The

LF pre/de-emphasis problem is likely a wrong choice for one of the EQ freqs.   Will be fixed on actual release.

 

 

 

[John Dyson]

VERY IMPORTANT QUESTION FOR THE PROJECT.

IT IS CLOSE TO BEING FULLY FUNCTIONAL, should I work on high quality decoding speed soon?

 

As of the upcoming release, I believe that *by far* most of the problems in the decoder are resolved.  Of course, the proof will be in both private and public reviews, but I am at the point where EVERY impairment that I can describe is fixed in the test version.  (WITH GOOD HEARING.)

 

Most of the time, when suggestions are made about problems in the recordings,  I understand the suggestion PLUS my hearing is functioning,  then I have agreed.  At this time, my hearing is at 100% of what it can be, and I have been searching for defects and tells for imperfections....   I have come up without no further *common* impairments in the upcoming version.

 

I CAN be convinced of other impairments, but when showing defects,  COMMON defects for all (or most) of the recordings is critical.  MUCH (not all) of the time, defects on one class or another of recordings is an EQ matter (e.g. Supertramp, Carpenters,  Anne Murray.)   Sometimes pre-encoding EQ is done to make the FA version sound better.   We need to distinguish in those cases.   Another possible impariment is the need for something like -3dB at 9kHz on all recordings, it is VERY possible that my hearing is missing that need (Supertramp recordings need that kind of EQ.)

 

Recent corrections (tell me if I am wrong):

1) Bass should be in correct proportion.

2) Upper midrange intensity is under control (it was too-strong gain control intensity, not signal level.)

3) HF timing/phase distortions appear to be fully under control (you can hear incredibly clean line up of detail on some recordings.)

4) Bass problem that caused beginning of 'Even in the Quietest Moments' to have expander effects -- FIXED.   (Bass pre-emph/de-emph.)

5) Additional layer of decoding (up to 8) gets rid of the 'slippery' sound.  8 is a nice, round number -- 7 always bothered me, but seemed to work.

 

Various anecdotal improvements is that there is no hiss leakage on older recordings (modulation of hiss is gone), eg Carpenters Bacharach Medley.   All of the frequency balance problems have been resolved as either decoder fixes, or pre-emphasis before FA encoding.   (Supertramp is much better.)

 

The sound of 'S' has no slurring or 'hard stop' once the correct choice of two calibration levels are used.

 

The only matter that might call this into question is about the 3k to 6k range sound is still a little strong.   This brings the vocals perhaps a little too far forward.   I *DO* believe that the strong 3k to 6k region  to be part of the mixing/mastering to make a recording sound more bright.   Most of the time, when a vocal is alittle too bright, a small amount of EQ (mastering) can fix it.  If one recording is brought back down, then others suffer, so MUST be either EQ or mastering/mixing manipulation.

 

(About my hearing -- I did some research, and also from my memory -- some kinds of age related hearing loss can be mitigated by Lecithin suppliments -- real research has been done in that direction.   My hearing has been consistently good for two days now. )

 

Thanks for your consideration and thinking about this.

 

 

[John Dyson]

Here are the current V3.0.12M snippets -- with the fixes that I previously described....

There was an attempted release, but there was a logistical problem that will be resolved by Thursday.

 

However, this is the sound that I expect in the next release.   I'll also being decoding the requested demos tonight.

 

https://www.dropbox.com/sh/sw78c7wzewwskmp/AADr90Ax_9NOfFS5JsRNV4jca?dl=0

 

Unless someone gives a big veto, this will be the sound of the next release.

 

 

 

[John Dyson]

About the decoder project....   I just realized that most people likely do not understand the reason...   Do you think that I want to sell or even distribute decoders?

 

Think about MQA, yet another layer of garbaging  up the recordings.

The FA encoding scheme is the same kind of nonsense.

 

My crazy goal is to embarass the industry into producing clean recordings.   That WILL make me and the project a target of ridicule by the insiders, but I just don't care.

 

Sometimes the truth hurts, and even if the decoder never becomes perfect (it really is close to being as good as possible), that isn't the actual goal.

I want EVERYONE to have clean recordings -- not just use the decoder.

MQA is not even as bad as FA!!!

 

 

[John Dyson]

New version of the decoder will be almost infinitely better than any before.

However, must be delayed because of personal logistics (I won't be available to work on the decoder.)

Trying for Wed release, but I consider Thu as a professional-type goal.  I want to get the release out well before

European/UK early Fri evening before the weekend.   This means 'Thu' night my time.  Public demos should

be ready at release time, but one-on-one decodes (e.g requests and others) will be planned to be completed

on Saturday (hopefully before.)

 

To avoid reading the 'blather' below, here is a shortcut description:

1) HF is more correct, less intense.   Double checked many times.  No more hearing damage from excess HF.🙂

2) HF dynamics are more correct.(dynamics strong when they should be, smoother when should be smoother.)

3) No need for HF variants.

4) LF needs variants, but perhaps 5 at most, where 1 or 2  other than 'default' (prob 3 total) will cover almost all cases.

5) Old Carpenters and ABBA recordings are no longer variants from more common, newer recordings!

6) Might add a full classical mode that covers all settings, including EQ, calibration, stereo image, etc.

7) Correction for Windows builds where there was mistaken use of single precision math insead of double precision

in a very very very sensitive piece of code.  (Yes, such precision is seldom actually needed in most applications, but without

the correct precision, the sound quality is very noticeably poorer.)   This WILL have noticeable impact on non AVX2 or AVX512 machines.

8)  Other improvements in DolbyA compatible sections.   Better attack dynamics,  even less modulation distortion, etc.

9)  Useful cheat-sheet (2 pages max) that cover 'normal' switches and easy refer-to and understand definitions of variants.

10) Full advantage of the 8 layers over 7 layers.  Dynamics notably more strong (less 'slippery' sounding.)

 

I want to forewarn that even with the more balanced frequency response (might even be perfect?!?!?) vocals will be slightly more forward than FA (only slightly), and instruments like cymbals sound a little more real (they wont' squish like on FA.)   Pianos will sound better than older decoder versions (less muddy.)

 

But also very important, the highs are NO LONGER ear burningly bright sounding.   Once in a while, I can hear how intense the highs actually were, and there was considerable consideration about the most likely choices of EQ that the original designer might have used.

 

Only the most intrepid need to read further.

-------------------------------------------------------------------------

Below is the 'blather' with more explanations and details:

 

Changes:

The dynamics will be slightly improved, tested/reviewed over and over and over to get it right.

The internal DolbyA compatible decoders have some small general quality improvements and Windows corrections (used the wrong math precision in a critical place on Windows.)

 

The HF frequency response balance is now correct, or at least much better.   A lot of tests were done, being careful about my hearing problems.   Basically, there is some correction in the high end.   The end result is that the highs should be less overwhelming for those with good hearing.   One caveat -- the FA raw material has suppressed HF dynamics, so expect stuff like cymbal crashes to sound like cymbal crashes, and not 'swish'.   They aren't overwhelming, and less intense than before, but much more real than the FA raw.   All of the advantages of 8 layers vs. 7 layers are being utilized.

 

The LF frequency response is better.  Previously has also been 'trouble'.   Even though a lot of things were correct, there were also some incorrect assumptions in the post decoding EQ.   The EQ looks much more 'regular' and as if an engineer would have designed it.  In previous decodes, some recordings might have been okay (ignoring excessive HF), but then others would have a muddy sound.   That problem is corrected with a caveat below.

 

Since everything really does sound better, variations in LF EQ between recordings become more apparent.  HF pre-emph and de-emph are no longer an issue.  With the corrected HF output EQ, the pre-emph/de-emph oddity dissappered.  

 

Unhappily, there are variations in LF on perhaps 30% of the recordings, at 1kHz, 500Hz, 75Hz, 37.5Hz (AFAIR).  Much of the time, the variant LF EQ are audibly optional.   When setting the decoder for a given recording, one might think of the LF setting as a user-tweak in perhaps 1/2 of the variants.

 

Once I noticed the LF variants, I was worried about usability.   However, the LF EQ exceptions are much less of a problem than I had worried about.   There are several prominent LF EQ methods used on recordings, but I am going to CLEARLY document each one, and explain when to use each one.   Hopefully, the decoder  documentation will make it easier to get used to the method for EQ.   I plan a single page cheat sheet similar to what @lucretiusdid when trying to document the unfortunately and unknowingly primitive version of the decoder.   I still regret my too much optimism, but without the optimism, the project would never have gotten this far.

 

Support for LF variants will be an 'A', 'B', 'C' type thing.  I am going to try to make sure that normally only one letter will be used, and only a few bounded variants.   I expect that along with the 'default', only two or three of the 'letter' scheme will used in by far, most of the cases.   There will also be a more general purpose EQ scheme that will handle practically every possible case that I can think of, and will be tedious to normally use.  I wont' encourage using the very general version of EQ.   I believe that out of default plus  'A', 'B', 'C', 'D', 'E' instead (one each, no multiple characters like 'AA' type stuff), will cover all recordings so that they sound reasonably better, and will probably make 99% of the recordings absolutely correct.  I WILL explain each one, and make 'A' used more often than 'B',  'B' more likely than 'C', etc etc.   Frustratingly, there just is no way to get a 'perfect' decode without some LF EQ. 

 

Just to reassure -- even ABBA and the old Carpenters use the standard, default LF EQ.   Classica/pure instrumental recordings appear to require a slight modification, but a modification nonetheless (less lower midrange.)   This might be a general vocal/instrumental optimization, and will try to make this a 'first choice' for EQ.  Interestingly, I have found that practically all classical recordings need the same general settings, and JUST MIGHT create a 'classical' mode that sets EVERYTHING.

 

I will be trying for Wednesday of next week.   I consider that on/before Thu night my time would be like a 'commercial' workplace level goal.

 

  

[John Dyson]

Been staying quiet on singular technical matters, but I finally found a good 'tell' for the pre-emphasis/de-emphasis, so now I am 99.99% sure that it is now correct!!!   This is a real

relief.   Likewise, associated with THAT tell, I can get the output EQ approx correct without fully depending on my hearing.    The output EQ accuracy is a litlte less demanding for lack of 'jumble' than the pre-emph/de-emph.   Normally, 2nd order and other 'sharp' EQ schemes dont' create the 'jumble' when incorrect.   However, 1st order EQ must line-up for such wide bandwidths that the timing/group-delay/phase errors become important.   I cannot 'hear' phase errors, but anyone can hear when phase errors create unexpected cancellation  in vulnerable parts of software and HW designs.   This cancellation and timing wobbles is the best description of the 'tells' that I am talking about here.

 

What is the 'tell'?  The 2nd Telarc 'Ein' disk, track 10, approx 55 seconds in.   Since these wide dynamic range recordings are pretty good, they result in sound similar to my master tape copies.   When the pre-emph/de-emp is incorrect, the instruments/strings jumble badly.   It almost sounds line IMD/MD that the anti-MD couldn't get rid of.   In fact, initially, the jumbles made me think  that there was a bug in the anti-MD code.   The high quality recordings are so clean that these little errors become more audible.

 

After exhausting everything else in the decoder, I finally thought that it might have something to do with an error in the pre-emph/de-emph.   After several iterations of plausible pre-emph/de-emph, the results stabilized and cleared up.

 

One might reasonably suggest:  maybe just found a local minimum?   maybe it isn't the best 'correct' solution?    I believe that to be unlikely, because I have found that literally *everything* must be correct in the timing for the results to avoid a jumble.   At this time, the 'jumble' avoidance in complex recordings is better than I have ever heard.   My attitude about the quality of recordings available to the consumer HAS improved.   However, all such recordings ARE FA encoded.   I have found only about two out of many thousand recordings are FA.   Even though the recordings are good quality, the consumer is NOT able to realize the full quality.

 

A side-benefit is that there was the opportunity to maximize the anti-MD behavior -- simply to maximally clarify recordings.

Interestingly, when the anti-MD is fully functional, bugs/distortions/timing errors (jumbles, group delay) become more and more obvious.   In a way, the anti-MD can make recordings sound WORSE.

Just doing a decode without anti-MD leaves so much fog that a lot of these details are 'fuzzed' out.

 

This is the best technical news that I have to offer in a few weeks...

 

[John Dyson]

Positive results -- moving further forward!!!

 

Got a 2nd pair of headphones (cannot tell much through speakers at the detail that I need...)   (DT 990 special with straight cable.)   The original main reason for the DT990 is pressure against my head possibly causing HF hearing loss.   IMPORTANT DIFFERENCE FOR THE USER -- the results will be MUCH more mellow.

 

Learned something very interesting -- with the new DT990 headphones, with just a few dB difference, the effect of FA decoding is more subtle.   Doing the decoding is still worthwhile, but changing the bar is so very interesting indeed.

 

Also, the difference of just the few dB are demonstrating the big difference in sound than what I expected.   The DT990 is much more transparent and has a more profound stereo image than the DT770/80 Pro.    Out of both headphones, the 770 is more 'fun', but might have misguided me in a few settings.

 

I can just guesstimate that the response curve was a +6dB at 9kHz (more at higher freqs) and +6dB at 75Hz (even more at lower freqss) and 0dB at 2kHz.   That can be nasty sounding indeed.   The original FA version  (raw recordings) sound so very intense that when I matched them (per the 770), the result was way, way too intense.

 

Re-organizing the building blocks based on the DT990 produces a much more mellow sound.   When listening to the settings for the DT770 with a DT990 about burnt my ears with intense sound..

 

Also, the rolloff above 9kHz (starting about 6kHz) and at 75Hz (starting at about 120Hz) is much greater when using the DT990.

Along with my hearing, the DT770/80, which sounds REALLY good, creates the impression that 'more is better'.   When listening to the decoder adjusted for a RAW/decoded match with the DT990, the DT770 also sounds more mellow and produces a sound similar to the DT990.

 

*  I do not yet understand the effect that I describe, but suffice to say, I can understand why the end  results have sounded 'excessive'.

 

My hearing is going to need to readjust, hopefully it happens before Thursday.   I'd suspect that a much more *normal* sounding result will be coming out this week.  Still trying for Thursday as EVERYTHING else is going really well, but might be delayed by a day.

 

Most interesting -- I can understand why some people might not 'hear much difference' (assuming the response is correct), but there really *IS* a major difference, just perhaps more subtle than what I was hearing.   The advantage of the DT770/80 is that I could hear the little distortions so much more easily.   So, the DT770/Pro80 was useful, but I probably should have had DT990 available (or something similar) also.

 

Moving forward!!!

John

 

[John Dyson]

There is a release with much better controlled highs.   V3.1.0A.

Caveats (minor, compared to the older versions) after the links below.

 

Location for snippets:

https://www.dropbox.com/sh/i6jccfopoi93s05/AAAZYvdR5co3-d1OM7v0BxWja?dl=0

Location for decoder (with new, simpler usage leaflets):

https://www.dropbox.com/sh/5xtemxz5a4j6r38/AADlJJezI9EzZPNgvTNtcR8ra?dl=0
 

The new leaflets are 'work in progress', but the short version has everything needed for 'normal' decodes.

 

This release comes from using a new set of headphones as the basis for EQ.   Unfortunately, my old headphones apparently were defective, and had a very profound rolloff at 6kHz or so.  That rolloff encouraged very intense highs in the result, because the decoder was tuned to EQ against those headphones.   Now, the highs should be much more normal, but we have a little trouble right now: Unfortunately, my hearing just 'switched', so at the last minute cannot tell how good it is.  I am now hearing bass a little heavy.   The results might be wonderfully perfect, or LF just might be about 1.5dB to 3dB bass heavy.

 

Fewer building blocks are now needed for the highs, and the EQ makes more sense WRT DolbyA behavior.  Apparently, there was some additional EQ that I could now do better to reverse it's effects.

 

The HF EQ is now much more 'normal'.  I really like the EQ that was needed, and really does make sense.   The LF EQ, as always, is a little tricky, but all in all -- your ears wont' hurt when listening to these results.   The decoder and short snippets are available, and if you can, please let me know if there is too much bass or not.  (Some also have access more complete set of demos -- just trying to keep things from being out of control!!!)  I am 99.99% sure about the highs.

 

 When designing the BASS EQ,  bass requires some EQ where there is a 'boost' in the 100Hz to 1kHz range, but I am having problems with judging which freq: 1k, 750, 500, or 250.   And, how much?  Most likely 3 dB at one of these.   There is also the possibility of a boost at 150Hz and 75Hz, probably 3dB on each.   Then, there is the needed rolloff at 50, 37.5, 25 or 18.75Hz.  How much for the rolloff?  probably -3dB, except for the last freq, which might best be -6dB.

 

This version sounds TOTALLY different, and I think that you are going to hear something much closer to what I had been hearing through my bad headphones.   Again, the bass might be heavy.   THAT WILL BE FIXED SOON (days, not weeks.)   To fix it, if I am hearing what I think is happening, maybe -3dB at 75Hz or 150Hz?   However, that problem just might not exist.   I have to  wait until tomorrow morning to find out if my hearing has been 'telling lies' again.   Up until the last minute, I have been searching for 'tells', and then my hearing died about 2-3Hrs ago.

 

Tell me what you think..  Maybe this is a little less embarassing -- I didn't realize that 'this' emperor had no clothes. :-).   People kept saying so, but I was hearing something that didn't represent what was on the decoded recording.

 

TRULY SORRY AND REGRET MISJUDGING MY HARWARE AND ANY WASTED TIME BY THOSE WILLING TO HELP...