'FeralA' decoder -- free-to-use

[John Dyson]

Corrected Decoder and demos available.  V3.1.0C.

This includes better, more detailed BASS than V3.1.0B

 

lf you have heard a real tape direct from a mix, you will be impressed how close the new decodes are coming.  I had been misled by my headphones, and frustrated explanations to me about the specific defect.   As long as the *new* headphones are close to flat, these decodes ARE VERY CLOSE TO FLAT.   The V3.1.0A had some bass problems (and a trip-up in the high end), but these have been fixed by resurrecting a previous test version that I had intended to release instead.

 

Demos below, and decoder is in the subdirectory.

 

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

[John Dyson]

Just got some minor feedback, and will do a release before late tonight US time -- approx +9Hrs.   I must be careful about my hearing, but the results (with my new headphones) are good.   In the past, I did previously get feedback from some audio people (not here), and previously did get favorable results with the excessive highs.   This could be because of EQ that they were willing to do.

 

 

I plan, along with the release, to have the  normal demos and blind < 55sec snippets for Supertramp, and maybe even some ABBA stuff.

 

Sorry about this modification, but it is a last tweak type thing.   There was a block of EQ that I had missed, and really wakes up the sound.   Interestingly, it is a 1.5dB step, which I am going to try to figure out futher why it is only 1.5dB difference.

 

With my new, correct headphones, the decoder is starting to sound like what I used to hear with the old ones :-).   There was one heck of amount of EQ for the old ones, but I had no way to tell other than feedback from users that I couldn't interpret.   It was VERY far off -- NOW it is lovely sounding with my new headphones.

 

John

[John Dyson]

I am having REAL troubles holding back about the release coming tonight.  I am trying for +3Hrs, really, really hard.

 

1) Buried vocals problem fixed.

2) Better control of the bass.

3) The highs are a little more transparent.

 

What was the major problem?   Another one of those really wierd things that are so common in the decoder.  This thing is the 'reverse engineering' challenge of a life time.

 

I am doing test decodes (same as what I upload as the demos) right now.   If those pass (no botches), then the release WILL happen.   Next will be the Supertramp stuff (snippets in the public area, and made available to some other supporters, critics.)   ABBA is coming after that, but will DEFINITELY not be available till the morning (my time.)  VERY SELDOM do a full set of decodes -- only when major improvements.

 

The new release will be V3.1.1A (hopefully), and the change from 3.1.0C to 3.1.1A shows the fact that the improvement is pretty important.  Maybe not perfect, but that is what you guys and gals are for -- TELL ME WHAT IS WRONG.   I think it might be tricky to find major bugs now.

 

THANKS, and a new toy is coming (hopefully tonight.)

 

John

 

[John Dyson]

The decoder is VERYVERYVERY close to ready, but frankly, I wasn't totally comfortable with the bass on the run of the demos.   So, I cannot make another time estimate until at least +12Hrs.

 

This really MUST be the best that can be done with the current techniques, but the bass ended up being JUST A LITTLE too thin.

 

Working on this very intensively.

Did get some encouraging feedback on the last version, and don't want to regress!!!

 

John

 

[John Dyson]

Release V3.1.1K is available.

Sorry about the delay, but I have had a *really* tough time on getting the balances correct.  My hearing has been changing, and as soon as everything is 'good' , then two hours later, I have to re-organize the building blocks again.   FRUSTRATING!!!  PLEASE TELL ME THE ERRORS...  Once I can quantify the errors, I will do what is needed ASAP.

 

* I just realized that it might not be *fair* for me to ask you to do a blind estimate on the BASS.  I'll be producing a V3.1.1KA version (added 'A') that has more full 100-150Hz range.   I'll ask for a comparison -- the files will be in a test directory that I'll supply in about 1Hr.   At that point, I'll ask for a comparison between the default V3.1.1K and V3.1.1KA versions...  Give me 2Hrs...

 

Decoder:

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

 

Samples:

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

 

1)  The HF isn't naturally as bright as FA on a normal recording.   The super highs on FA are 'enhanced', but decoded versions can come close.   Since all HF EQ is 1st order, any change to the super highs will affect the lower highs, so it can be tricky to get the right balance.   Eventually, I decided that the recordings JUST MIGHT be bright sounding to begin with, so I added an '--hfrl' (HF rolloff) switch.   Without specifying the switch, there is about a -1.5dB at 18kHz with the smooth rolloff from about 7kHz -- 9kHz EQ actually starts a little way below that.   If you use '--hfrl=0', you turn off all of the rolloff.   If you use '--hfrl=2', then you get more HF rolloff.  These are carefully chosen rolloffs, so I suggest that as a beginning user, try the '--hfrl' switch instead of EQ yourself.

 

2) I just got feedback on the V3.1.1F version that the 70-150Hz bass might be a little thin.   This is a *REALLY* complicated thing because of the wierd, paradoxical effects when dealing with bass rolloff.   I can make the 100Hz range sound more thin if I add signal in the 18-25Hz range.  So, this is so complicated that I really do need feedback

 

Believe it or not, esp when considering feedback that I have gotten, this EQ is likely no more than 1.5dB in error.  Because the EQ is 1st order, the 1.5dB error isn't just starting at a narrow frequency range, but can affect the gain as a curve over a range perhaps twice as wide as the low 2nd order Q value of 0.50, so this is TRICKY to say the least.

 

Two questions:  (private or public response is welcome);

 

1) is the upper treble too strong?  give me an estimate about how much.   if it is like 3dB at 18kHz too much, I can easily fix that and stay within the rules.   if it is like 1.5dB too much or too little, then that can be tricky to do right.   Eventually, it can become tricky as the EQ gets closer and closer to correct.

 

2) tell me about the bass...  Remember that FA will produce a more resonant sound, but the low bass can actually be stronger on the decoded version.  If the lower bass is too strong, it actually hides the upper bass.   Tricky balancing act.

 

HELP ME, and you'll get the best possible software toy.

John

 

 

[John Dyson]

After all of this turmoil about the LF being too thin, I made a very simple change from 75Hz to 150Hz for one of the +3dB EQ, lo and behold, much more full sound.

 

Bottom line, there will be another release tomorrow morning at about +12Hz (maybe later, but really trying for the goal.)   Also, will revisit the HF again.

 

Sorry about my release stuttering...   You might still want to listen to the demos just for an idea of what I am talking about.

 

John

 

[John Dyson]

There is now a V3.1.1L release available.

I listened to the last minute criticism -- at first, as usual when trying to figure it out, and was a little skeptical.  After thinking about it, and trying some changes, eventually I agreed that the upper bass was thin in V3.1.1K.   I thank the several reviewers, but especially the reviewer who did an excellent prescription.

 

The snippet site has V3.1.1H, V3.1.1K and V3.1.1L.   Make sure to choose the correct version that you want to try.

(Thanks again to reviewers!!!)    Suggestions REALLY helped!!!

 

Snippets (V3.1.1H, V3.1.1K, V3.1.1L):   * ADD-ON, added some examples -- from Steely Dan, Dionne Warwick, etc.

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

 

Decoders (V3.1.1K, V3.1.1L) -- docs are valid for BOTH -- minor change for V3.1.1L, but simple enough to ignore for now.

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

 

Unless more suggestions otherwise, we will be moving forward based on V3.1.1L, and is probably the 'gold standard' right now.   Making adjustments from this point (whichever of 'K' or 'L' that we choose) will keep changes under control.   Since the decoder is algorithmically TOTALLY complete, now we are at the point of  final adjustments.

 

DESCRIPTION OF CHANGES FROM 'K' to 'L':

 

The change for the bass was very simple...   There needed to be +3dB LF shelf, single pole at 75Hz.   Instead of 75Hz, I moved it to 150Hz.  One correspondent did an excellent job of prescribing the missing upper bass, even to the extent of a good estimate of the frequency range.   The LF response in general wasn't significantly changed, just the location of the +3dB EQ was moved.   (I cringe when making a dB increase or decrease, which usually requires reshuffling and a lot of work.)   Just moving the EQ is often very very simple.

 

For the treble, I finally gave-in  and did a -1.5dB HF shelf at 18kHz.   This has an effect of perhaps -0.75dB at 9kHz (or less).   Since the HF was probably between 0.75 and 1.5dB too hot anyway, the only real problem with the EQ is that there might be a little too much change below 9kHz -- any difference from the straight 1st order shelf requires some consideration.  In the future, the -1.5dB at 18kHz might be changed to -0.75dB at 18kHz twice.   This would have much less effect at 9kHz, but effect 18kHz the same as before.

 

 

[John Dyson]

Gawd -- my hearing is bad for sure.  I have also lost a lot of my ability to discern freq response balance in he general case, whether hearing or not!!!

The results coming in V3.1.1N will have the SAME general response balance as the FA version -- finally got it through my thick skull.    The decoder affords access to the same screwed up EQ mechanism in the command line as the '--pvh' and '--pvl' commands.
 

I was just doing some comparisons again between the FA recordings and my results.   Like all FA recordings, the highs have a rough kind of highs, which I heard as distortion.

 

WRONG -- I should have matched the levels of the distorted highs of FA instead of the part without the distortion.   That is, my previous  results were 6dB too weak at the high end, and about 3dB too weak at the low end.

 

The release delay is happening because of the upcoming correction....   It still might happen tonight, but I have to finish the demo decodes, and do a long listening session as my hearing gets more and more dull.

 

First -- in V3.1.1N (it is coming) the LF is flat because there is very very little EQ except the end rolloff at 25Hz and below.  That is -- now, in V3.1.1N, the LF cant be anything but flat.   Previous versions were botched because I was trying to EQ the bass down to match the high end.

 

Next -- in V3.1.1N, the highs will match the 'bright part' of the FA recording instead of the average levels.   (this is hard to explain), but will result in about 6dB higher signal at 9kHz, with the curve starting at +3dB at 6kHz.   The interesting thing about this is that the pre-emph is still correct, and to get these results only a few of the EQ steps are REMOVED from the de-emphasis.   These results will still be much less screechy intense when I was using bad headphes -- on V3.1.1N, the sound is clean and balanced -- but there is more of a high end than the V3.1.1L/V3.1.1M.

 

Like usual, I was listening for the wrong thing to match.  It is so sad that the beautifully distortion free signal isn't easy to enjoy because of my broken LAST PHASE EQ.   Everything else creates severe damage to the signal...  The last phase EQ just damages the EQ, which I cannot handle or fix very well.

 

I am watching the decodes go by -- just started about 20minutes ago, so it will be at least 40 more minutes before I can start doing the review (looking for distortion and any response balance errors that I can detect.)

 

[John Dyson]

 

Release V3.1.1Q is ready, and very profoundly improved.

This release is NO JOKE, and if you are going to try any release, this one is very worthwhile.

All versions prior to V3.1.1Q are a waste of time, other than for historical review.

 

Demos:

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

 

Decoder: (V3.1.1Q ONLY. Docs updated)

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

 

Comments about the V3.1.1Q release:

The quality is very profoundly improved. Been dealing with a tedious approx 5minutes of accurate hearing, much shorter than my previously assumed 15minutes. This short time frame of accurate hearing has caused some weird results. This has been an incredibly frustrating moving the EQ building blocks around based on A/B choices that can last only 5minutes at a time. If the EQ attempt lasts too long, the goal is no longer accurate. I know that this has been an ongoing problem, and I do regret the frustration.

(There still might be a bit too much bass – but to get ‘Crime of the Century’ to sound right, the bass is a bit strong for some other recordings.) I am thinking that I might have to sacrifice the sound of ‘Crime’ a little for everything else to be optimal. The error < 1-2dB at 20-30Hz.

 

I wanted to get this release out today – to show progress, and also show the trickiness of balancing the sound of recording whose EQ is impossible to determine which one is correct!!!

 

1) The highs are very clean, and I dare say very close to accurate now. The original EQ effectively had a strong rolloff starting at about 6kHz, and proceeded to above 12kHz, with then a boost again at 18kHz. The HF EQ is now corrected along with a latent approx +1.5dB tilt from about 9kHz up to 20kHz. The tilt compensation, with anti-distortion optimizations, is enabled by default, but compensation can be removed if desired. (refer to –hfrl command).  The tilt compensation is different than what can be effected by a simple 1st order EQ, otherwise the tilt compensation would have been a part of the normal EQ.   I suspect that this might be a result of IIR filters being slightly different from real analog filters.  No matter, the tilt compensation works well.

* I suggest listening to the FA highs and compare with the decoded. You should find that the sense of the extreme and compressed FA highs are changed into truly dynamic highs. DO NOT compare with earlier versions of the decoder as the comparison will be misleading.

 

2) The lows are truly flat, perhaps with a small excess around 20Hz. The EQ has NO 1^st order gain below 125Hz. The 2^nd order gain has always been +9dB at 75Hz, however there is also a combo of 1^st order and 2^nd order rolloff down to approx 12.5Hz. (Remember, a 1^st order EQ at 12.5Hz has some effect at 20Hz.) Frustratingly, since there is only +-3dB and +-6dB EQ at discrete frequencies per the ‘rules’, there appears to be a very slight peak in the 20Hz to 30Hz range, thereby making some recordings (e.g. some Anne Murray stuff) having a very slight peak in the lowest bass. When I try to add a slight rolloff (per the rules), the results are inferior on some recordings like 'Crime of the Century'. On the LF, we are down into the approx 1dB error range. I’ll continue to make some decisions on this LF issue.

* The decoded lows will sound slightly different because of the compression, and the fact that the midrange is compressed differently from the low-lows. This will give a more ‘V’ shaped response character for the FA version. On the decoded version, the results are more typical of flat recordings.

 

3) The pre-emphasis has been correct for a long time. However, the HF correction results from a modification of the de-emphasis. The pre-emph and de-emph are not 100% mirror images, but I suspect that this design choice optimization was made because of the effect of compressors/expanders.

 

4) Unfortunately, a few recordings have obviously excessive bass. The ‘–lfeq’ switch is available to correct it. Since there are so few necessary adjustments, with few choices needed per adjustment, the addition of a relatively simple-to-use one shouldn’t be overwhelming.

 

 

[John Dyson]

 

I just made the following post, and it was TOTALLY WRONG...   This is an example of how unreliable my hearing is.   Just did a very careful comparison between the FA and decoded version, and they are too simialr to be as bad as what the result sounded to me..

FALSE ALARM -- but just showing you how bad it is for me!!!  Everything,  so far, is cool...

 

Still investigating, but bad hearing might have caused troubles again.   Trying really hard, but I ran tests over periods of multiple days.   If there is an error, it is just a matter of a few 3dB errors.   Everything is in the increments that keeps things somewhat under control -- really frustrating for sure.

 

[John Dyson]

The responses have been encouraging, but after a lot of testing, I did find a minor bug of approx 1 to 2 lines.   The 6kHz->inf range is 3dB too hot.  That kind of mistake is easy to make with my 65yr old hearing...  It will be fixed ASAP, and will be V3.2.0A, and at that point will be the baseline version for speed improvements, and minor quality fixes.   This change is *literally* one or two lines of change!!! It is also possible that 12kHz->inf also has an additional 3dB error (that is where two lines might be needed.)

 

* It does appear that the pre-emphasis and de-emphasis WILL be identical.   My attempt to match the FA original was fraught with trouble because my HF hearing had been so weak that I couldn't hear much difference.  I wish I did this project 10yrs earlier!!!

 

Expect the fix to be available within 1 day, but probably +13Hrs, assuming I don't botch anything and have to do multiple passes of debugging/testing.

 

John

 

[John Dyson]

7 hours ago, John Dyson said:

 

I just made the following post, and it was TOTALLY WRONG...   This is an example of how unreliable my hearing is.   Just did a very careful comparison between the FA and decoded version, and they are too simialr to be as bad as what the result sounded to me..

FALSE ALARM -- but just showing you how bad it is for me!!!  Everything,  so far, is cool...

 

Still investigating, but bad hearing might have caused troubles again.   Trying really hard, but I ran tests over periods of multiple days.   If there is an error, it is just a matter of a few 3dB errors.   Everything is in the increments that keeps things somewhat under control -- really frustrating for sure.

 

PS: the first attempted correction is:

Remove -3dB at 6kHz HF shelf

Add -6dB at 6kHz HF shelf

 

Remove -3dB 12kHz HF shelf

Add -6dB at 6kHz HF shelf

 

Add -3dB at 50Hz LF shelf

Add -6dB at 37.5Hz LF shelf

 

(The LF shelves don't have the profound change that the HF shelves do, the change is actually relatively minor.)

Both of these can be the result of 'bad hearing', but notice that EVERYTHING is a 'standard' EQ value or 'standard' frequency...

[John Dyson]

V3.2.0A -- sounds very different, very major (but simple) change.

(This is a status update release, not fully ready -- but getting closer -- still too much HF.   Will no longer try to do a full release without 2nd opinions.)

More 'hearing fixes'.

Serious inter-layer bug fixed.   It broke trying to decode some of Supertramp -- previously made recording shrill, not just HF boost.  The first cut of 'Quiet' was insanely bad on the previous release.   This version is much more sane, but just probably too strong highs.

Material like 'Nat King Cole' with the piano sounding muddy, now that is fixed.

 

Demos: (decoder in subdir)

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

 

What bug fixed?

Between each layer, there needs to be an HF cut, or all h*ll breaks loose.   That EQ is basically two 1st pole EQ.   During experiments, I found that 2 each of -6dB at 18kHz worked well.  Unfortunately, that wasn't quite correct.

 

The actual solution is a similar EQ to the 2each of -6dB at 18kHz,  instead is -3dB at 9kHz, and -6dB at 18kHz.   The difference makes the sound much better.   Also, the dynamics end up being more like what I remember.  Normally, both of the EQ sound similar, but obviously aren't 100% the same.   I just made the wrong choice, and didn't realize it until doing a review of normally untouched sections of code yesterday.

 

Sometimes I have been accepting of the sound, but not totally satisfied.  Do the recordings actually sound like that?  Often, I believe the sound to be correct, only because of 'accomodation'.   This inter-layer change makes the sound much closer, but I still think that the highs are too intense.   My hearing is very clear during this session, so will try to do the HF EQ for a release later today.

 

AGAIN -- since I have made misguided 'releases' in the past, I am only claiming this to be a 'status report'.   I have gotten feedback that a very recent release is 'good' -- I do think that there are still some problems, but are getting better and better.

 

This is a very difficult thing to do, not simple like a single layer.  This multi-layer thing is probably 10X more difficult at least.

 

I'd expect the next release to have highs that have more realistic levels.

 

[John Dyson]

New V3.3.0A status report test version (with demos):

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

 

(I need this feedback every several major changes because my judgement gets twisted.)

 

1)  Better/More < 40Hz low bass.  (Been getting complaints since day one.)

2) Better joining up the lower midrange/upper bass  (previous version caused a 'bump' that I didn' realize)

3) MUCH more clean high end.

 

My major worry is if the bass vs higher frequencies is balanced well?   I am pretty sure that there is enough low bass now.  Spent a lot of time trying to get the right balance -- not that the mechanics of choosing the correct building blocks is difficult, it is judging what the sound  should be like based upon the sound from the FA recordings.

 

Details of what was done over the last few days:

 

I gave up on the 'trying to stick with 1st order rules', and moved to 2 layers of 1st order below 75Hz instead of one.  I believe that this fixes the low bass matter.   I really wanted to avoid another 2nd order EQ, but it seems like the 1st order rule is relaxed for low bass.   Every time using 1st order as a fix, the low bass became over tubby.

 

Also, a new general LF EQ -- joined directly the 500->1500Hz compensation to the <250Hz bass without any EQ.   The new 500-1500Hz compensation enabled this new EQ, but I didn't realize until this release.

 

I figured out a problem with the HF pre/de-emphasis.   This is related to my self imposed limit of HF EQ not being done above 21khz to avoid problems with 44.1kHz sample rates.   After a lot of futile attempts at getting 'that' sound, even though it was close to what I wanted, I tried looking at 24kHz EQ.   There is a place where the pre-emphasis EQ is alternated at various 3kHz increments starting at 3kHz with +6dB and -6dB.   For the best sound, the last iteration apparently had to be at 24kHz instead of my continual attempt at using 21kHz.   This keeps the decoder from internally running at below 48+kHz, but normally it doesn't run at below 66.15kHz (for 44.1kHz inputs) anyway.   I REALLY tried to make 21kHz work, but it wouldn't give that ultimate, precise high end that I wanted.  Once the 21kHz restriction was relaxed, then all works very well.

 

Most of the time has been working on choosing the correct Q and associated 1st order EQ at each LF EQ frequency.   I had pretty much decided that 75Hz/37.5Hz were needed, but when doing the 2nd order EQ, both the gain and Q needs to be chosen.   Also, should there be a mix of 1st order and 2nd order at each frequency?   After a lot of A/B tests, I found that both 6dB LF shelf and 3 each of 1.5dB/Q=2.0 are needed at 37.5Hz and 75Hz.   Previously, only 75Hz was 2nd order, and was all done in 2nd order instead of the mix.   Basically, at 37.5Hz and 75Hz, there needed to be approx 10dB boost, instead of the original approx 10dB boost at 75Hz only.  (This is the reason why <40Hz was weak.)  (6dB + 3*1.5dB = 10.5dB)

 

Once this was determined (it is all based on +-3/+-6dB building blocks and choosing the correct frequency from the limited list that I have mentioned before.)

 

I think that those complaining about bass will be happy now.  HOWEVER, as mentioned above: is the general bass a little too strong, or is it a good mix between the bass and the rest of the upwards spectrum?

 

 

 

[John Dyson]

Gosh -- one LF fix coming, and then I'll suspect that everyone should like the decoder much better.  There have been lots of challenges in the design, but everything is done carefully with understanding the concept and CONSERVATIVE rules .  The problem is guessing at what the original designer had chosen.   The other problem is that my hearing is failing.  It makes NO difference when I work on this, my hearing is simply unreliable.  Because of numerous considerations, it is better to make this thing work sooner rather than later.

 

* The improvements come from little bits of improved understanding about what the original designer was attempting to do.

 

Had been having troubles with the HF/LF balance (each HF and LF are much improved independently), so the 'join' between MF and LF has been difficult to do right.   By my hearing, I had though that the 'join' didn't need a +3dB EQ downwards at 500Hz or 250Hz?   (My broken hearing said that a join wasn't needed -- later on, that was found to be untrue.)

 

However, after some more review, a 'join' is needed, and the amount should be +3dB (architecturally), but should it be at 500Hz per the MF EQ, or should it be 250Hz per the inter layer EQ?   This is where listening comes in, and how my hearing screws things up.  (The bad decision about not needing a  join happened because my HF hearing had shut down,.)

 

After some comparisons, and noting that people have mentioned before that the midrange was lacking, I had decided that adding a 'join' between the 500-1500Hz EQ and the LF EQ should be at 500Hz and 3dB.

 

Basically, in V3.3.0A, the bass was light by 3dB.

The V3.3.0B release will have the bass added in by implementing the join.

Architecturally, it would have been much more clean to not need the join, but that was not done in a 'beautiful' way.

 

A new release V3.3.0B with demos should be ready in an hour or so.

 

 

 

 

 

[John Dyson]

The V3.3.0B demos,  decoder with binaries is available.

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

 

(DEMOS DELAYED for 2Hrs -- I screwed up a build -- Plan to look +2HRs from posting time.)

Before going further -- remember,my hearing is unreliable...  Which is best?  V3.3.0B or V3.3.0A?   Just let me know.

 

About the release relative to V3.3.0A and further comments about V3.3.0A/B.

 

This includes the fix for the join between MF and LF (+3dB.)

One might note -- how come he says that the LF is good, but the level is 3dB in error?   There are two reasons -- one is that I listen to the LF individually, but don't have reliable relative hearing between LF and MF.   Also, there are some 'tells' on some recordings where if the freq response has some bounces (variations), the sound on the recording becomes very odd.   Sometimes I trick myself into thinking that the recording has troubles, but when a change/improvement is determined, then the improvement becomes obvious.

Also, on LF, I was in a 1st order EQ-only brainlock.  When allowing the use of 2nd order EQ, the problems were relatively easy to solve.  There were two places where the LF got additional 2nd order EQ:  37.5Hz and 25Hz.   The 25Hz EQ was real trouble for 1st order EQ, but got much easier with 2nd order EQ.

 

The HF problem was less pronounced -- but there was a lingering minor problem.   It was hard to chase down because of my < 44.1kHz/2 EQ limit.   When removing that limit, then the 24kHz EQ fixed things.   One might ask:  WTF 24kHz EQ, when we can only hear up to 20kHZ?    The answer is that 24kHz 1st order EQ has a substantial effect down to about 12kHz or less.   The HF hasn't changed since V3.3.0A, and I won't want any changes unless minor adjustments (but we aren't doing 'minor adjustements' right now.)

 

 

 

 

[John Dyson]

Here is V3.3.0B again...   There will be some private demos and requests for feedback, but this will be the last version for at least several days.  This should be sufficient for almost anything.

 

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

 

Sorry about the mistaken build – I took a wrong earlier decoder version and assigned the version number to it. Once I listened to the demos, found that the first V3.3.0B was wrong. The version originally uploaded was never intended to be V3.3.0B, but this one is correct.

 

Here it is again:

 

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

 

My hearing appears to be getting worse, and that is one reason why I have been hurrying this along. I’d hate to see the really beautiful music be unavailable because of minor physical limitations.

 

Before uploading, I listened to at least 10-20 seconds of each cut to make sure all was okay. In fact, I caught a bug in the Linda Ronstadt recording where I used an earlier style EQ setting, so did a re-do. Also, I found that an Eric Clapton snippet had not been created because there was a specification error in the build file.

 

The correct V3.3.0B is now uploaded and available.

 

TECHNICAL COMMENT:

 

Previous versions (before the V3.3.0B series) had a bit of HF rolloff, no matter what, to soften the slight error because of an EQ tradeoff that I thought to be needed. This version, whether or not the highs are at exactly the correct setting (need feedback for that), they are

beautifully clean now.

 

So, now – everything is correct.  Waiting for arguments to the otherwise :-).

I will be working on the speedup now… I’ll be waiting for about 1wk or longer for feedback.

 

 

[John Dyson]

Wasn't going to do another release for a while, but almost 2X faster  on the advanced modes and a bugfix due to a mistake with conditional compiles.  Even without the HF bugfix, the quality should be slightly better because more code is double precision.

 

There are NEW EXAMPLES.   After this, as I had mentioned before, the private demos will start.  It will be nice to do the demos with the faster decodes!!!!

 

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

 

I was in the middle of the speedup changes, and noticed a bug.  This bug was related again to the highs being too intense, but as you know I cannot hear well.   I noticed the bug because I do understand what the program is supposed to do, finding it was not dependent on my hearing.  The choice of using the correction WAS dependent on my hearing.

 

The highs should really be good now (I am hearing clarity that I have never heard before), but the highs are much less strong.   Everything and all choices are based on a coherent set of rules, but sometimes intelligent exceptions are made.   So far, there are only 3 (three) exceptions to the 'rules' and all three are 2nd order EQ using good, standard values instead of depending entirely on 1st order EQ.   The 2nd order EQ are at a standard frequency and the Q values are 2.0 and 0.8409 (Chebyshev).

 

The speedup is for --fx, --fz, --xp, --xpp, and --xppp.

--xp, --xpp and --xppp should be about 2X faster, but --xp is usually the best choice for really good results.

 

--fz is best for 'very good' results on an AVX2 4 core machine.   The speedup is a little less.

 

Using the default, or more ideally --fx is best for slower machines.  'Default' is NOT sped up.

[jabbr]

Careful, I am hearing that music with real high frequency content is still getting substantially cut off. It is very likely that things like early Genesis and Yes etc are not Feral A because I'm still getting the highest frequency content amputated, whereas pop like Whitney Houston, Simon & Garfunkle and Michael Jackson are much better.

[John Dyson]

3 hours ago, jabbr said:

Careful, I am hearing that music with real high frequency content is still getting substantially cut off. It is very likely that things like early Genesis and Yes etc are not Feral A because I'm still getting the highest frequency content amputated, whereas pop like Whitney Houston, Simon & Garfunkle and Michael Jackson are much better.

 

Before reading this -- I found a strong 'tell' that Lamb is one of the few that are NOT FA.   I could tell within 1-2minutes.  So, you are right about 'Lamb', almost 100% likely.  However, the note below is still accurate.   Refer to my comment about the 'Break through', and that is the immediate tell!!!

 

 

They just might not be FA, but I do have some 'YES' that can be checked.  It can usually be proven definitively by misguided 'decoding' creating almost non-linear sounding distortion.  (The fast attack/release times, whacking at the signal 7 times over can create one heck of a weird distorted sound, but the distortion isn't strong -- just ugly.)   Sometimes, instead (or in addition), there is a break-through distortion.   Sometimes, the distortion is mild, but still is more or less still there.

 

 

 

There are still some frustrating software bugs because, honestly, I have been very clumsy lately (even physically), so I am RIGHT NOW trying to be careful to STEP BY STEP make sure that everything is in the code that should be.   Today, I built (privately) three versions, each time I thought that I had corrected a bug, but did not. The 'bug' was a simple missing text substitution, not a 'programming' bug!!!   This has been very frustrating indeed.    There are all kinds of things waiting in the wings, and too often I'll experiment on something, and by mistake leave part of a test inside of the working codebase.

 

Addressing it: I do have some 'Yes' also, and will check it.   I didn't hear any problems with the YES that I have, but again, it might need the EQ that I talk about.   If a recording needs EQ, and there is, perhaps needs a 6dB boost at a certain frequency can loose as much as 12dB additional!!!  Usually, in that situation, distortion will not be created, just a dead, dull sound...

 

Also, non-FA materials will often loose frequency response at very low levels, but that CAN be confused with the EQ problem.  So, it is POSSIBLE that no-FA recordings will sound like there is an HF loss, but the higher signal levels will then tend to 'break through' and jump at you with another kind of distortion.

 

 

I will visit 'Lamb (Lamb determined NOT to be FA)  and the Yes stuff later today.   It IS posssible that the material is not FA, but again -- if there is no 'graininess' or 'itchy' sound, just dull  and/or no 'breakthrough' -- then that really sounds like EQ or something else that I haven't figured out yet.  (It has been difficult enough to read old, Ray Dolby's mind on this project -- he had Alzheimers near the end, and it makes the mind meld a little more difficult :-)).

 

I am still trying to chase down my version skews -- the decoder works SO well, but I seem to botch the release.   Back in the good old days, I had someone to do the release stuff for me!!!

 

John

 

 

[John Dyson]

4 hours ago, jabbr said:

Careful, I am hearing that music with real high frequency content is still getting substantially cut off. It is very likely that things like early Genesis and Yes etc are not Feral A because I'm still getting the highest frequency content amputated, whereas pop like Whitney Houston, Simon & Garfunkle and Michael Jackson are much better.

Well -- I was very curious about Yes, and I do think that the copy that I have is VERY FA.

Attached is the snippet version...   Sounds good to me (but I don't know the group very well.)

 

This was done with the prospective V3.4.0F release (hopefully all together now!!!)

 

 

 

01. Yes - Owner of a Lonely Heart-SNIP.flac

[jabbr]

11 hours ago, John Dyson said:

Well -- I was very curious about Yes, and I do think that the copy that I have is VERY FA.

Attached is the snippet version...   Sounds good to me (but I don't know the group very well.)

 

This was done with the prospective V3.4.0F release (hopefully all together now!!!)

 

 

 

01. Yes - Owner of a Lonely Heart-SNIP.flac 3.45 MB · 2 downloads

I'm using "Close to the Edge" https://en.wikipedia.org/wiki/Close_to_the_Edge_(song) specifically listening to the "running water" and wind chimes in the intro.

 

Similarly Genesis: The Lamb Lies Down on Broadway" has chimes or bells in the long intro.  These are destroyed by FA decoding so I'm assuming that these recordings aren't FA. I'm also assuming that the decoding interferes with the harmonic structure.

[John Dyson]

5 minutes ago, jabbr said:

I'm using "Close to the Edge" https://en.wikipedia.org/wiki/Close_to_the_Edge_(song) specifically listening to the "running water" and wind chimes in the intro.

 

Similarly Genesis: The Lamb Lies Down on Broadway" has chimes or bells in the long intro.  These are destroyed by FA decoding so I'm assuming that these recordings aren't FA. I'm also assuming that the decoding interferes with the harmonic structure.

I do have a copy of 'Close to the Edge'.   In the middle of making the family meal -- the best chicken soup that is possible to do.  However, once I am less busy and don't have to watch the stove, I'll check into it.

 

One caveat about judging against something being FA -- the decoder WILL destroy the sound of FA encoded materials if there is errant EQ involved.   The signal must NOT be substantially EQed, or all h*ll breaks loose.   This is one reason why I believed 'Lamb' might be FA, and still *might* be, but I doubt it at this point.

 

Well, FA encoding also destroys the harmonic structure, FA decoding of FA encoded material restores it.   FA decoding of non-encoded material will certainly cause damage.  Normally, FA decoding of non-encoded material sounds like explicit distortion, but in the case of 'Lamb', it does not.  It is mostly manifested as frequency response   problems.  FA recordings just don't sound like a real recording to me -- and that is why I have been tricked into believing that there si more of an improvement than the is as manifested by most people.   Below -20dB of the average signal level, FA encoding  totally destroys the signal dynamics.  (again, this is -20dB of the current signal level, not just -20dB FS.)  At times, there is actively about +-20dB of gain control going on every 100-500msec, but the DR value doesn't change much.   Most of the FA damage is done at low signal levels.   It is REALLY TRICKY to 'fix' the signal at -20dB and below while trying to maintain the dynamics at above -20dB!!!

 

In general, if the midrange is preferrentlly attenuated, then the sound gets all screwed up, but this is true any time an attenuation of a midrange (or lows with harmonics) can happen.  This is one of the careful constraints that I have when getting the decoder 'just right', but most people probably don't realize all of the constraints that decoding really has. You obviously have recognized (when mentioning harmonics structure)  that one of the 'tells' for correctness is that there is enough midrange to support the manifestation of harmonics in the higher frequencies.   It isn't sufficient to say 'sounds bad', but the reasons for the problems must be well understood to remedy them.

 

 

 

 

[John Dyson]

Checked 'Close to the Edge' and the likilhood is probably 90% that it is FA, but is NOT FA=+7 like most recordings are.

It is apparently FA=7 instead.  It is very important to use --fa=7 instead of the default --fa.   Initially, I thought that it might just be an EQ problem, but is not.  It took me a while to remember the --fa=7 mode, simply because it is not used very often.  --fa=7 seems to happen more often than non-FA in commercial materials, but not all that often.

 

Why do I claim that it is FA?   The low levels have high end reasonably consistent with higher levels (that is, there isn't a pronounce shift in spectrum balance beyond what one might normally expect.)   There is no distortion or 'dead-zone' effect either.   Non FA material gets profoundly less HF at lower levels, but also can be caused with --coff being too high (e.g. --coff=0 will suppress the highs at low levels on most FA recordings.   Likewise, using --coff=-2 instead of --coff=-4 will also suppress the highs at low levels.

 

I can hear the low levels at the beginning, but they aren't very loud.  This is to be expected when the input (raw FA recording) is also low level.  This is similar, but in reverse to the 'Ein' Telarc recordings where the cannons end up being much more realistic levels -- and suppress the music down to -20dB..  (If you clip or limit the cannons, you can bring the levels up to being more listenable.)   I can barely hear the beginning on your examples, but I gain-up by 10dB or even 20dB to make sure that it is all there.  It does apparently appear to be complete and consistent.

Basically, it appears that the decoded version has at least 10dB wider dynamics, and the dynamics appear to be consistent.

 

Is it worth decoding the material even if it is almost assuredly FA?  Maybe, maybe not...

 

 

 

 

[StephenJK]

John,

 

For purposes of comparison, I've dropped Yes - Close to the Edge and Genesis - The Lamb Dies Down on Broadway to this DropBox link:

 

https://www.dropbox.com/sh/77crt0gvyofdqaj/AAAb4xUBq6fo50_sNviUnHhGa?dl=0

 

Both are from the 1973 and 1974 LPs that I recorded.  That may give you an analog reference.  Note that Close to the Edge is from Yessongs, and may not have the same progression as the studio track.  

 

I seem to recall we did this once before with a Supertramp track.  All my LPs were recorded on a much better than average turntable and MC cartridge, but of course can only be as good as the source material.