'FeralA' decoder -- free-to-use

[John Dyson]

10 hours ago, bogi said:

The decoded version sounds IMO much much better than the original - I had feeling of more depth of the recording space, instruments were much better separated and their timbre was clearer. I have no objections to EQ balance of the decoded version. Keep the good work. I hope you will release a stable version with complete manual and enough simple usage instructions soon.

Thanks!!!

About the manual -- I feel *internally generated*  pressure to write a useful, true manual.  Also, respecting the frustration/needs/comfort of those trying to use the decoder, I owe each user better docs.   The 'StartUsing' document is only meant to be a quick set of hints, and is absolutely insufficient as a full manual.  I think that everyone understands that I am 110% occupied doing the technical stuff.   When not eating, fixing family dinners and doing other personal necessary things, I am working on the decoder. HOWEVER, there WILL be a manual soon (planning for weeks, not months) because the decoder is finally 'zeroing-in' to being what we all likely want and enjoy.  After the next release, I'll allocate a few days time-off (4Hrs of testing/listening instead of 10+Hrs/day).  THEN, I'll have much more freedom away from hard-core programming and listening to tests too much of the day.  The highest priority of interest will be the actual manual (or a prelim version) that has ALL known information.  A pretty version will probably start emerging a little later.  I want to make sure that the nuances and tricks are available to everyone.

* There is ONE more set of *simple* EQ issues that need to be implemented before the next release -- the closer-to-FA sound decoded version, it *does* sound more like FA EQ (of course, doesn't have the FA grit).  The fuller sound of the normal decoder mode will also be available.  ALSO, there are recordings done with different EQ, probably intended to artfully improve the sound after FA encoding.  Those additional EQ issues for about 20% of the recordings will be accommodated with a few simple commands.  (The decoder can already handle the EQ differences, but the commands are currently ugly.)   I intend to do these wrap-up issues late today/early tomorrow, because 99% of the technical matter is already known (or already internally implemented.)   Hoping the next release (pray it is essentially final) in under +2 days.  New mods/improvements will then be possible, but instead of 1-2 days for a simple mod, implementing new things might sometimes be 1-2 wks.  Focus will be on *quality* instead of simple function.  Also have user interface ideas for the future.

 

Side notes:

 

Important also, for my own satisfaction, is an 'internals' manual.  There are some truly *interesting* things about how it works.  If the decoder was comprised of textbook mechanisms, it probably would have been easier to design/implement.  Also, there were, and still are complications associated with 'accommodation' or becoming 'accustomed' to certain kinds of sound qualities.  These matters manifest both in the 'developer' (me) and many of those who try out the decoder. About understanding the internal design,  there is a simple-crude ASCII-art diagram in a little message to another correspondent.  On request, I will share it publicly when I clean it up a little bit, as a crude version of the current block diagram.  (Private requests for an immediate copy might need to result in the already existent version.)  The way that many other sound processing schemes work has been expanded at an insane level. 

 

When working at the Labs, I was often paired with a 'sane' and 'organized' person so that the results of my thinking could be better captured.  I truly miss those days, and perhaps one reason why I announced prematurely -- almost begging for another person (or persons) to join the team and capture the design concepts, better organizing things.   That did *not* happen and caused some confusion about the decoder and the quality of the design.   The quality of the *work* isn't really great, but the quality of the *concepts* are beautiful.  Frustrating, and I'd really like to document the *concepts.*   Some of the *concepts* might really benefit other forward looking audio software designers/developers.  (Some people DID attempt to help with documentation, but unknowing to me, the decoder still had 2yrs of development to go.)

 

Of course, right now, the true 'Users manual' is the most important and highest priority documentation.  At least very soon (weeks, not months), expect to complete a preliminary, but very USABLE users document.  The biggest weakness in an initial users manual will be a lack of usable drawings, even though I will attempt to do some.  (I have a deficit in visualization -- there is NO mapping of my concepts to visual documentation -- it is blank.)   Even though the 'words' will be good, the 'drawings' will probably be ASCII-art.   Later on, there WILL be drawings.

 

 

.

 

 

[John Dyson]

Found some cool things, interesting improvements, another little 'secret' nit in decode -- clarifies material like ABBA, other recordings with complex vocals.

Delay release/new docs so that stuff will start being available about Friday.  (Absolute goal is for weekend.)

 

 

[John Dyson]

Sometimes, some reviewers have made comments about a profound difference in freq response between decoded and FA.   The idea of 'frequency response' is nonsense when looking at a multi-band audio processor, but there IS a way of doing a qualitative comparison.   The *general* spectrum should be similar between input and output on FA vs the output.  There ARE differences, and sometimes when the differences are totally eliminated, then the sound becomes *terrible*.

 

This was done on the upcoming V5.0.1B version, recent previous versions are similar (perhaps slightly wider variance), and the future actual release (coming in days) will be similar to below. 

 

Here is a primitive table.  That describes the relative signal density on a specific recording (first 10 seconds of Anne Murray 'Shadows'.)

You might notice that the density is very, very, very similar, where there is a small bump in the 20-50Hz range and the HF >3kHz is about 1dB too hot, with a slight, PLANNED rolloff in the 18kHz on up region.  These differences can be attributed to either response anomalies or state/behavior of the natural decoding process, esp considing the HF response *IS* dependent on signal level.

 

*The 18kHz rolloff is a side-defect an anti-garble technique, cleans up stuff like ABBA chorus, and appears to be in all recordings.  I might not have found the ideal method yet.  It *can* be disabled, but no real reason to do so because the rolloff is only at the very highest frequencies.  It doesn't really affect below about 15kHz.

 

(The first numbers are most interesting, the second two numbers are the individual L and R channels.)

 

LEVELS 20Hz to 20kHz

RAW FA

RMS lev dB -24.22 -23.78 -24.71

DECODED

RMS lev dB -24.29 -23.80 -24.84

 

LEVELS 20Hz to 50Hz

RAW FA

RMS lev dB -37.83 -37.28 -38.46

DECODED

RMS lev dB -35.85 -35.19 -36.64

 

LEVELS 20Hz to 200Hz

RAW FA

RMS lev dB -26.31 -25.82 -26.87

DECODED

RMS lev dB -26.95 -26.37 -27.62

 

LEVELS 50Hz to 1000Hz

RAW FA

RMS lev dB -24.38 -23.94 -24.87

DECODED

RMS lev dB -24.42 -23.94 -24.96

 

LEVELS 3000Hz to 6000Hz

RAW FA

RMS lev dB -39.35 -38.82 -39.97

DECODED

RMS lev dB -38.17 -37.53 -38.91

 

LEVELS 3000Hz to 12000Hz

RAW FA

RMS lev dB -38.12 -37.59 -38.71

DECODED

RMS lev dB -37.11 -36.49 -37.84

 

LEVELS 3000Hz to 20000Hz

RAW FA

RMS lev dB -38.19 -37.67 -38.78

DECODED

RMS lev dB -37.26 -36.64 -37.99

 

LEVELS 10000Hz to 20000Hz

RAW FA

RMS lev dB -46.69 -46.22 -47.22

DECODED

RMS lev dB -47.30 -46.72 -47.96

 

 

 

[John Dyson]

Phew!!!   This total rework of the frequency response balance is now complete.  It really doesn't sound very different, but the HF response is more clean and extended, and the LF is pretty much the same, but a little better methodology.

 

Also, there are two EQ modes -- one, which I *think* is what some (hopefully feq) people might prefer when complaining about the response, but sounds NOTHING like the FA recording.  Then, there is a slightly better version of the most current  emulation that sounds similar to the FA original (like before), but with a few bumps removed.   Unlike previous versions, the measurements that I can do shows pretty darned flat/good out to 20kHz.   Previously, there was an HF dip at the highest frequencies, but perhaps a 9k->12kHz bump.   Those are gone.   Of course, a direct measurement is *impossible* because there are no specs, so only a response comparison with FA, also with consideration for the midrange response distortion done by the MF encoding process.  (The 100Hz to 3000Hz range doesn't change gain much, but the 3kHz->inf and the <20Hz to 120Hz range has lots of gain variation.)   So, the idea of decoding being 'flat' is total nonsense.  (Please don't do measurements and make claims that the decoder isn't flat -- measuring non-flat, direct input to output is a GOOD THING.)

 

The 'compensated' mode has a 'bigger' 100Hz to 300Hz signal level, but not more correct.   As is, by default, the decoder *really is* correct and by hearing that is not FA accomodated, has the same response balance (within a small delta.)   This is the second, and most complete re computing of the internal EQ.   Absolutely NO personal biases were allowed.

 

At first, I was getting worried -- because I did start totally from scratch, and avoided any prejudice caused by previous EQ.   These new results came from a very tedious, non-trivial kind of measurement instead of listening.    I am hoping for tomorrow night, perhaps Friday night.  Also, I will start the new users guide, and try to make some early, raw information available.

 

Since always trying to be responsible, it seemed like it would be a good idea to redo the EQ.   The results *ARE* audibly similar to previous, but the new results are definitely technically better.

 

John

 

 

[John Dyson]

First, in the most recent version, there had been EQ drift, so I had to go through an entire EQ test/revamp.   The default result coming from the decoder is similar to when the decoder was aligned properly, HOWEVER...   It is very easy for the alignment to drift -- it is a complicated thing to do.  I am going to beg an indugence...   There are two versions of EQ, changeable by a command switch, and I cannot decide which one is correct:

 

1)  Same as previous 'correctly operating' versions, except very carefully aligned for the soon to be demos/release.

2)  Another plausible kind of EQ, actually matching the FA better in certain ways, but sounds so very wrong. (my listening judgment is poor.)

 

Both styles have plausible measured comparison results, one sounds 'correct' in comparison, but measures in a different way than the new alternative.   The new alternative sounds very, very, very bland -- but just might be correct.

 

The release/demos/etc are in process.   I'll also try to update a better users manual, but will still be interim.   Getting this done will fully occupy me for the next couple of days, because there are LOTS Of details.   There will be private messages and a public announcement for those who want to join this clean-up effort.

 

John

 

[John Dyson]

I regret the confusion and previous comment about the alternative EQ.   There is no need for a test.   After some reflection on the new, re-corrected EQ, the default output of the decoder is indeed correct, and has been double checked by some new measurement techniques.

 

I am *very very* sure that a sense that the EQ is very much in error will come from a bias of being adapted to the FA recordings.   The most recent one or two releases HAD drifted, I should have re centered the EQ as the successive minor changes became major.

 

The resutls are measuring spectral density from input to output to be essentially flat in the 110 to 1000Hz, 3000 to 9000 and 9000 to 20kHz ranges, and the transitions between the regions are essentially mathematically as expected.   (Again, there is NO 'flat' in the FA scheme.)    The 1000Hz to 3000Hz range has the expected dip of approx 3dB, and the region below 100Hz has the expected 5dB dip.   The reason for the dip in the midrange is related to the curves in the 500 to 1250Hz transition region and the dynamics in the 3kHz on up region, so that several dB dip compensates for the downward gain control.

 

As far as can be discerned from the spectral information while testing more than several recordings with differing character, the decoder IS flat.

 

The MF (1kHz to 3kHz) was the EQ matter discussed before.   There is a way to mostly compensate for that rolloff, but after studying the problem, it is very clear that the 'dip' is very necessary.   Also, with the 'dip' the sound is more similar.

 

 

[John Dyson]

Sorry, there *will* be a delay.  I am making a *major* push for being as accurate as possible.  'Accuracy' is hard to determine, but have been developing some more methods for doing the determination, and NOW have a scheme to eliminate the frequency balance drift.   The precision of the decoder, even when maladjusted, is very tight.

 

The good news, since starting the EQ almost totally from scratch for the first time in a long time, and taking everything that has been learned, the results are 'interesting'.

Much of the 'anti' distortion scheme (other than the simplest methods) comes from a very simple thing that was done to the phase/frequency response in the 15kHz on up region.

 

The LF, MF and HF DO have an interleave scheme, but not as much of an add-on, but more 'organic' in the EQ itself.

 

The HF EQ is now greatly simplified -- 6 steps on input (increment at 3k or 6k intervals), and 3 steps on output.  Also, finally found a *repeatable* way to mitigate the HF burn-out problem where hearing is overwhelmed by about +3dB at 12kHz, +6dB at 18kHz -- excess gone.   If there are future changes, I found a good way to re-evaluate the 'flattness' of the HF (or LF or MF) resposne.   One major thing changed:  the lowest freq is NOT 3kHz, but instead 2.75kHz.  That has been a major problem for EQ.   However, all kinds of things 'fell out' by using 2.75kHz for the first step.   I'll be doing an experiment later on today trying to use the 2.75kHz as a base offset througout the EQ system...   That is, of course, 2.75kHz is 2.75kHz, but should 6kHz actually be 5.75kHz?  This small difference is actually quite profound when at 2.75kHz, maybe more improvement when using 5.75kHz instead of 6kHz?

 

The LF EQ is also greatly simplified, where the EQ is back down to sanity -- 6 required 1st order LF at 1000(-6), 500(-6), 160(+3), 80(+3), 40(+3), and 1st order +6 at 160, 80, 40.   That is almost all there is.  I have also added some very low frequency rolloff, gratuitiously, but just as a matter of responsibility.

 

I'd rather spend more time for improving the decoder based upon some newly learned ways to prove it's correctness instead of doing a release today.

 

 

[John Dyson]

Very interesting matter about the sound of 'FA' style of EQ.   It is very very similar to the 'flat' style of EQ that the decider has normally provided.

 

I still do believe that there is an emphasis in the midrange and 'tubby' part of the bass on FA materials, but that isn't important -- the new version of the decoder can decode while producing results similar to either sound style.

 

The interesting thing about both decoder modes,  is that the stereo image widens significantly, like there is a full left-to-right stereo image in front of the listener.  

 

I will be doing more comparisons between my original album copies vs the two, selectable modes of the decoder.   I will DEFINITELY determine which mode of the decoder matchesthe originals  most accurately.   When I get the results, I'll make the information public.

 

A VERY frustrating, but reassuring thing is that the EQ schema for both the raw FA EQ and the EQ that I prefer -- the EQ is very very similar.   I cannot chose between them based on complexity.   If one scheme is overly complex, then it can be taken as likely incorrect.

 

It took an entire day while using my OBJECTIVE method of measuring EQ, in addition to a subjective component to determine the kind of EQ that matches for both modes.  Even though the modes sound *very* different, the resulting EQ only requres a few minor mods, and oddly -- the level of complexity is also similar, so choosing from an intuitive method is impossible.    Several recordings were used as a driving function for the comparison tests, and the results came out very similar.  (Well, on some materials like the cymbals on Brubeck, there is very little midrange information, so the midrange result might be distorted on parts of Brubeck. )  There are SUBJECTIVE aspects of my much more OBJECTIVE set of measures, and with the OBJECTIVE measures, very repeatable results will now be possible...

 

The biggest difference in EQ is all about an unintiutive reversal for the singular operation needed before decoding .  Each step requires that there be a continued dip in the 1250 to 2750 midrange, or the midrange levels go nuts.   The new, perhaps nowadays prefered  'FA' style of sound is partially driven by one extra inverse step of the midrange EQ AFTER DECODING.   This means that for 8 layers of decoding, an inverse copy of the midrange EQ needs to be done.   This is intuitively INCORRECT, because if the levels are kept similar from step by step if there is a dip before encoding, then The new, perhaps nowadays prefered  'FA' style of sound is partially driven by one extra inverse step of the midrange EQ AFTER DECODING.  an additional, compensation seems to be 'abnormal' because the decoding mechanism cannot stomach the version without the EQ.

 

No matter what, many people prefer what I consider to be 'tubby sound', but I am pretty sure that others consider what I prefer as being the opposite (perhaps enhanced lows and highs???)   The decoder can now do both styles of EQ, and IN BOTH  CASES, BOTH VERSIONS THE FINAL LF EQ ARE ALGORITHMICALLY VERY SIMILAR AND PLAUSIBLE.

 

John

 

 

 

[John Dyson]

It will be about a week before I am back.

 

The decoder response has been drifting too much.  Per some of my previous comments, I have been very frustrated about my own inability to do QC on the equalization.   I have developed a very reliable method for QC, but all by itself, is also a project to implement.

The scheme is based upon using the recordings as a driving function instead of a sine wave.   Using a sine wave for finding the spectral response is TOTAL FOLLY, and trying to do so would show a misunderstanding about what is going on.

OTOH, a real recording CAN be used to *estimate* (not measure) the spectral response, but using multiple recordings of different statistics can produce a fairly reliable estimation of the 'frequency response' (more correctly, spectral response.)

 

This 'QC' project isn't simple, but it will also be done along with a repeatable implementation of the EQ on the decoder.   The complexity of the 'estimation' mechanism is significant all by itself, but I expect that there can be a new decoder with FULL QC along with the QC mechanism in about 1wk.   I already have prototypes running, but still some work yet to do.  The eventual goal (probably only partially attainable) is the ability to do QC by a mostly objective measure.

 

John

 

[John Dyson]

With the new measurement method (which isn't an obvious way of doing things), I found that it is borderline impossible to get a REALLY flat response with 1st order EQ.   It is emotionally  really hard for me to convert to 2nd order, but after giving up on the 1st order quest, getting proper EQ with 2nd order literally took a few hours at most.

 

The resulting sound now -- PRECISE AND CLEAN...

 

Basically, the decoder is now getting approx +-0.25dB response in the HF and LF bands, except the lowest freqs below 100Hz, where there is a rising response of about 4-5dB, likewise at the highest frequencies, there is a slight declining response of about -1dB between 9kHz and 20kHz.

 

The measurements that are being done are NOT as precise as using a sine wave, because the driving functions are recordings themselves.   So, if the spectral content doesn't have much information >10kHz (typical recordings), then the error in the measurement increases rapidly, even though the general trend is true.  The HF decline of approx 1dB is incredibly variable as some recordings show no decline (strong cymbals.)   Likewise, the LF naturally had a rising response relative to the FA recording, which seems true since FA recordings are not flat on the low end.

 

I tried correcting for the -1dB at 20kHz, but the results became quite noticeably shrill.   Likewise, correcting the low end +5dB or so made the bass go away.   It is pretty clear that the decoder is relatively flat.   The midrange (150Hz to 1kHz) and (3kHz to 10kHz) are really good.   I haven't carefully measured 1kHz to 3kHz, but the average level in that bad does match the 20 to 20kHz deviation from FA.  (That is, the delta gain is the same, meaning wideband flat relative to FA.)

 

The big difference vs. previous sound is that it is VERY VERY clean sounding, and there cannot be a basis of claiming a non-flat response.  Previous versions DID have bumps, and earlier versions had rising HF response.   They weren't *that* bad (*except excess HF) when compared with the instantaneous variable response of FA recordings.  (The frequency response of FA recordings is LITERALLY +-20dB, depending on the recording!!!)   My guess is that the decoder brings that *instantaneous* variation down to the +-3dB range -- MUCH MUCH better.

 

I think that *anyone* listening to the upcoming version will NOT be disappointed with the extreme clarity and truly very good frequency response in comparision with FA.  Think that the improvement over previous versions that just happened to be properly EQed is perhaps 2 to 5 times better,

 

 

 

[John Dyson]

Still have some repeatability tests to complete, but below are the results that are achieved so far. The results are truly VERY GOOD.

The release will come in days, not weeks.  Even though the decoder is truly flat considering the input material and characteristics of the FA encoding,  the goal of this clean-up is to create a method that can reliably be used for quality control (QC.)

 

Note that the results depend a little on the material used to drive the calculations.   I used Anne Murray's 'Shadows in the Moonlight' as the driving function.   In a way, the recording represents random noise, but the spectral content & dynamics of actual random noise would produce questionable results.   These results below and the summary immediately following are raw and have not been massaged.  There isn't much to contest, except that the measurements are likely WORSE than the decoder actually produces.

 

SUMMARY:

between 250Hz and 1.5kHz, approx 0.3dB variance.   below 250, slow increase, below 100Hz slightly faster increase, with total just over +3dB.

between 3kHz and 9kHz, approx 0.6dB variance.   above 9kHz to 20kHz, approx 2dB loss.

the difference between 1000Hz-3000Hz vs 3000Hz to 20kHz is about 0.15dB

the measurement of the difference between 20Hz-1.5kHz and the rest is specious because of the general increase below 250Hz,

if you look at the general levels between 300Hz and 1.5kHz, it is very close to the 1000Hz-3000Hz and 3000Hz to 20kHz.

 

Basically, the variance is very, very small considering the driving function can cause a -1dB error vs frequency (generally won't over-estimate.)

(ALL OF THESE ARE DEPENDENT ON SOURCE MATERAL, likely the variance of the decoder itself is *less* than the measurement.)

 

The major caveat is that if the spectral content is too weak, then the resulting measurement might come out too low.   This is exactly the problem >12kHz or so on 'Shadows', where the energy content drops significantly, so the HF measurement becomes inaccurate.   If using material with lots of HF content, the resulting response has shown a slight increase above 12kHz instead of the slight decrease with the current test material.

 

General overview of the results:   a few dB increase starting below about 250Hz, accelerating below 100Hz for approx 3-4dB increase.   I did run an experiment to remove this increase, but the result was bass-deficient.   Likewise, on the high end there is a slight decrease starting in the 7.5kHz to 9kHz range, and dropping about 1.5dB at 20kHz.   This HF decrease doesn't happen on well recorded material with strong HF (e.g. strong cymbals.)  Removing that slight HF decrease produces a brittle, artificially bright sound.

 

raw data is below:

LEVELS 20Hz to 20000Hz
dB raw: -23.47 dB dec: -26.50 dB diff: -3.03
LEVELS 1000Hz to 3000Hz
dB raw: -32.96 dB dec: -36.39 dB diff: -3.43
LEVELS 3000Hz to 20000Hz
dB raw: -37.42 dB dec: -41.00 dB diff: -3.58
LEVELS 1400Hz to 1500Hz
dB raw: -37.60 dB dec: -41.06 dB diff: -3.46
LEVELS 1300Hz to 1400Hz
dB raw: -37.30 dB dec: -40.78 dB diff: -3.48
LEVELS 1200Hz to 1300Hz
dB raw: -36.94 dB dec: -40.45 dB diff: -3.51
LEVELS 1100Hz to 1200Hz
dB raw: -36.51 dB dec: -40.04 dB diff: -3.53
LEVELS 1000Hz to 1100Hz
dB raw: -35.97 dB dec: -39.52 dB diff: -3.55
LEVELS 900Hz to 1000Hz
dB raw: -35.33 dB dec: -38.89 dB diff: -3.56
LEVELS 800Hz to 900Hz
dB raw: -34.62 dB dec: -38.18 dB diff: -3.56
LEVELS 700Hz to 800Hz
dB raw: -33.88 dB dec: -37.42 dB diff: -3.54
LEVELS 600Hz to 700Hz
dB raw: -33.12 dB dec: -36.65 dB diff: -3.53
LEVELS 500Hz to 600Hz
dB raw: -32.23 dB dec: -35.76 dB diff: -3.53
LEVELS 400Hz to 500Hz
dB raw: -31.04 dB dec: -34.64 dB diff: -3.6
LEVELS 300Hz to 400Hz
dB raw: -29.68 dB dec: -33.34 dB diff: -3.66
LEVELS 200Hz to 300Hz
dB raw: -28.22 dB dec: -31.60 dB diff: -3.38
LEVELS 150Hz to 200Hz
dB raw: -28.53 dB dec: -31.57 dB diff: -3.04
LEVELS 100Hz to 150Hz
dB raw: -28.44 dB dec: -31.45 dB diff: -3.01
LEVELS 50Hz to 100Hz
dB raw: -29.81 dB dec: -31.88 dB diff: -2.07
LEVELS 40Hz to 80Hz
dB raw: -31.63 dB dec: -33.14 dB diff: -1.51
LEVELS 20Hz to 50Hz
dB raw: -37.57 dB dec: -38.18 dB diff: -0.61
LEVELS 20Hz to 30Hz
dB raw: -46.07 dB dec: -46.37 dB diff: -0.3
LEVELS 3000Hz to 20000Hz
dB raw: -37.42 dB dec: -41.00 dB diff: -3.58
LEVELS 3000Hz to 6000Hz
dB raw: -38.53 dB dec: -41.88 dB diff: -3.35
LEVELS 3000Hz to 4500Hz
dB raw: -39.68 dB dec: -42.97 dB diff: -3.29
LEVELS 4500Hz to 6000Hz
dB raw: -41.20 dB dec: -44.64 dB diff: -3.44
LEVELS 7500Hz to 9000Hz
dB raw: -43.40 dB dec: -47.31 dB diff: -3.91
LEVELS 6000Hz to 9000Hz
dB raw: -41.32 dB dec: -45.08 dB diff: -3.76
LEVELS 6000Hz to 7500Hz
dB raw: -42.09 dB dec: -45.74 dB diff: -3.65
LEVELS 7500Hz to 9000Hz
dB raw: -43.40 dB dec: -47.31 dB diff: -3.91
LEVELS 3000Hz to 9000Hz
dB raw: -37.54 dB dec: -41.01 dB diff: -3.47
LEVELS 9000Hz to 12000Hz
dB raw: -44.83 dB dec: -49.14 dB diff: -4.31
LEVELS 9000Hz to 20000Hz
dB raw: -44.65 dB dec: -49.16 dB diff: -4.51
LEVELS 12000Hz to 20000Hz
dB raw: -48.73 dB dec: -53.78 dB diff: -5.05
LEVELS 12000Hz to 15000Hz
dB raw: -48.82 dB dec: -53.73 dB diff: -4.91
LEVELS 15000Hz to 18000Hz
dB raw: -52.65 dB dec: -58.07 dB diff: -5.42
LEVELS 15000Hz to 20000Hz
dB raw: -52.61 dB dec: -58.09 dB diff: -5.48

 

 

[John Dyson]

I found a *perfect* non-FA copy of ABBA Arrival.   The decoder is nearly perfect, except a minor flaw needs correction.  Will be working on it for a few days.

I spent a lot of time making the decoder *really flat*.   However, to match the non-FA copy of 'Arrival', some of the EQ 'shims' need to be removed.  This is actually encouraging, because the basic EQ without 'shims' appears to be correct!!!

 

WOW!!! 

 

 

 

[John Dyson]

Finally - V5.0.5A is up for veto.  Tests are on the reviewers site, and hopefully will be able to do release tomorrow night.  The demos are running right now (a LOT >200 recordings/songs.)

[John Dyson]

I am going to make myself a liar, but must delay for a few days.   I am sometimes hearing a 'hitch' in some vocals, and that worries me.   This is about the layering and how the curves butt-up against each other.   There are a couple of places where there *might* be a problem -- one is between each layer, and the other is the transition back from -30dB to -70dB again. (The sequence is -60, -50, -40, -30, then back to -70, -60, -50, -40, etc.   The other problem is to double check that there should be a 10dB interval (very likely true.)

 

AFAIR, I had designed the sequence to be -60,-50,-40,-30,-60,-50,-40,-30 or -60,-50,-40,-30,-50,-40,-30.   I don't remember restarting 10dB lower. This doesn't mean anything really evil except SOMETIMES the gains will be inconsistent or sometimes inadequate expansion.   At the lowest level, the problem wont' be noticed, but at the in-between levels, there can be an inconsistency.   Because of the design, most distortion would be hidden, but SOMETHIGN seems to be wrong.  If this bug is what I think that it is, the only audible change will be to remove the hitch and other very minor POSITIVE effects.   The only real-world change will likely be a 1 day delay.   I'll be able to verify the fix (if any) rather quickly, and redo the demos within 6Hrs of the fix.

 

I am demanding perfection, and have gotten beautiful help from a few *patient* people who understand the challenges.   This commitment requires that I know of NO leftover bugs.  I just realized the possible problem an hour or so ago, and stewed on it for a while.   The onlly correct solution is to do the review and delay.   I have a Drs appt till this afternoon, but will plan and work ASAP on it.

 

ADD-ON:  I just checked with a 'secret' command that re-orders the sequence above (I forgot about it), and the results show that there is a super-high liklihood that the sequence is wrong as I stated above.   Cymbals are much more clean -- stupid me.

 

Apologies.

 

[John Dyson]

Public message for the local reviewers:

The resulting EQ with the new sequence will essentially be identical.   The real difference is that the dynamics will be slightly stronger & and more intrinsically clean sounding.

 

Still testing, but using the 'acid test' is the vocal on 'Crime of the Century', which I didn't check until yesterday.  I found an incorrect 'hitch' in the singers voice on the first selection.  I found that the second sequence started with a threshold of 10dB less than the first sequence.   That was a bug, and really doesn't make much sense.   Instead, the 2nd threshold should be the same as the first sequence or 10dB greater.   The second decision is 'which choice'.    The new choice is 10dB greater.   It seems to work, and on test recordings so far, the sound of cymbals is improved.   Previously, was a little dissatisfied with cymbals, but I had attributed it to compression.

 

So, currently testing with +10dB, and so far the vocal on Crime of the Century appears to have a natural contour, and the dynamics are much more interesting.   The amount of expansion is effectively increased by +20dB, but in the case that the +10dB threshold is incorrect, then I KNOW that the threshold would be +0dB.   I will also be checking if the change in threshold might be +5dB instead of 0dB or +10dB, and it is plausible that it is, but seems to me to be suboptimal & incorrect.

 

Again, the EQ will be almost exactly the same as the previous successful tests, but expect a much more interesting (and correct) listening experience.

 

The reason why the decoder sounded reasonable, even with the incorrect offsets between sequences is partially about R Dolby's robust design decisions and the basic design of the decoder being so precise and of high caliber.  OTOH, my problem for the last year has been all about the proper EQ.   But with literally over 100 variables, it can be really difficult to make proper choices, especially because one can get all the way up to the last choice of the 100, then find out the previous choices were wrong.   There are lots of aliases of 'almost correct'.  When looking at the challenges, it is amazing (to me) that I am still sane.   The sanity is re-enforced by the finally successful results, not just at my level of eureka about the general design, but also from the exacting demands of most audiophiles.

 

 

 

[John Dyson]

Preliminary test decodes for V5.0.7D are available on the public and private demo sites.

 

Upcoming (already created) V5.0.7E version  fixes  known (VERY MINOR) V5.0.7D bugs such as correcting the expansion ratio, minor stereo image issue (moving the background and vocals to the correct places), and minor EQ (<1dB).   The new version very successfully passed the first layer tests, but the demos must be decoded and reviewed before the new V5.0.7E is ready for release.   If the new version decodes pass the quality requirements, then the release will very likely  be available +24Hrs from this posting.

 

Relative to V5.0.7D, the expansion ratio and the stereo image are most important.   The EQ is minor, but important.

Decoding results with even V5.0.7D mostly blows away the never-FA ABBA examples.   The hopefully soon-to-be released is better, but there is a slight matter about certain recordings having manifesting a very minor over-correction when moving from V5.0.7D to V5.0.7E, but only on the most fragile recordings.   All in all, V5.0.7E is very noticeably better, and per the NEVER FA examples, produces results superior to vinyl originals.*

 

* I just found a pure, never FA copy of SuperTrouper.   I had it all along and didn't notice it.  Sadly, the title song also has quality problems on the non-FA version -- so my goal of producing a perfect result on the song 'SuperTrouper' was simply impossible!!!

 

 

[John Dyson]

Snippet release of V5.0.7D and V5.0.7E (same directory, you can select which one) are on the public snippet site.  The private site also has some updates.

The V5.0.7E release was going to be available tonight, but because of a program editing error, there is a missing 40Hz, -6dB LF equalizer.   The effect is that the bass slightly overwhelms the rest of the signal, and appears to be missing some highs.   The highs are 'all there', but the sound effect is that there is too much 80Hz to 200Hz bass.  (Below 40Hz isn't all that irritating, except perhaps too much.)  That upper-bass/lowerlower midrange can be very distracting.

 

Unless I get a veto from some more reviewers, there will be a V5.0.7F coming some-time tonight. (only adding a 40Hz,-6dB equalizer.)   The same effect can be produced by using V5.0.7E decoder with --pvdl=40,-6.   I haven't uploaded that decoder because at this point appears to have known faults.

 

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

[John Dyson]

We are still working on the release.   There is currently a V5.0.7F semi-release version.  It has been deemed very good by the project testers, but also some suggestions for minor change have been made.  In the past, V5.0.7F would have seemed to be acceptable to almost all testers.   Given the feedback being received now, the goalpost has been moved to a more challenging/higher quality level.  This is a GOOD thing.   I do believe that relative to a clinically perfect version of the decoder, we are partially into the realm of personal preference.   Both suggested changes below will manifest as <1dB of response change in their respective frequency ranges.    As noted on the first, already complete change, the result was approximately 0.75dB increase in energy over the approx 20-80Hz range.

 

If interested, the V5.0.7F demos & decoder are available in the usual spots.   Given the feedback and comparisons with actual, NEVER-FA recordings, the results are very close to the ultimate goal, perhaps already meeting it? -- probably not yet.  In most cases, if the listener doesn't perceive the decoding result as an improvement, the answer is because of either/both of two reasons:   I got the EQ wrong for the demo or, the FA version is simply deemed preferable.  There is a near match on even the most difficult to process materials...  In most cases, the decoder is most likely producing correct decodes, even including the EQ.

 

Note:

One of the windmills that I had been 'tilting' was decoding the SuperTrouper title song.  It never sounded 'just right'.  Even with the V5.0.7F, the SuperTrouper decoding results  don't sound correct to me.   However, the 'imperfect' decoded version DOES match the never-FA version, therefore making a 'perfect' sounding decode of SuperTrouper is impossible!!!  No sense in trying harder on SuperTrouper.  BTW:  the decoder DOES come close to the never-FA version, and actually produces *superior* results when comparing with some never-FA examples.   The main reason for superiority  is that sometimesthe reference never-FA is from vinyl.  In the specific case mentioned here -- VERY VERY VERY good vinyl for Waterloo, and perfect digital for the never-FA SuperTrouper.  The result is also impressive on the Arrival album.  (Another reason for superiority is that the decoder is pretty good at mitigating the modulation distortion from previous noise reduction encode/decode cycles.)

 

Expected improvements from V5.0.7F to next version:

 

1) Do a slight bass increase.   This is actually implemented as decreased bass EQ.  There NEEDS to be a bass rolloff, but the shape & amt of the rolloff is an original design decision.  The suggestion was that there should be more bass, and that goal has been achieved for the next release.   There will be about 0.75dB more real bass below 50-80Hz (aggregate power increase based on the test material.)   The LF change is small, and makes a subtle improvement.   This improvement & several tests took an hour or so -- minor, but probably important.with a good side effect.   More clean/less disturbed vocals are a side-effect of this change.   Because of the slight timing improvement, it allowed more spatial distinction between the vocal and the bass sounds.

 

2) The upper midrange vs. lower midrange balance seems to be biased too much to the lower midrange.   That is, the lower midrange is making the signal sound very very very slightly more muddy than it probably should be.   This change will be tricky, so might take a full day.  Just as a matter of the shape of the power spectrum, vocals should also be a bit more clear sounding after this change.   The improvement here will be more based on signal power instead of timing like in the lower bass improvement.

 

Each of the above changes will likely result in a *simplification* of the Depending on my working efficiency, it might take a day for me to produce a test version instead of a few hours.  (Normally, I can make good changes in a few hours -- MOST of the time delay on producing releases is for making the test decodes.) EQ sequence.  As noted before, these changes are sometimes like playing Tetris.   Luckily, we are so close to the canonically correct result that the 'game' is becoming less challenging.

 

I intend that the upper bass will remain similar (perhaps a little stronger), and there is a  slight but important distinction between lower midrange and upper bass.  It is tricky to maintain a distinction between the upper bass and lower midrange, but it can and has been done.   When looking at the energy curves, there is currently a 'hitch' in the current energy curves when decoding several of my test examples.   The 'hitch' or non-monotonic shape in lower midrange/bass transition gain is about 0.5dB of unexpected gain increase instead of expected decrease.   This 'hitch' is the likely reason why there was a request for a slight midrange rebalance.

 

Because of non-audio processing matters, an expected test SW version with both of these changes (improvements) might take a dayor so, instead of a few hours.   No matter what, the demos also need to be run, so even if I had the test release ready right now, it would take approx 1/2 day to produce correct/accurate test versions.   There just might not appear to be much of a slowdown in development, even though there will be a slowdown. Depending on my working efficiency, it might take a day for me to produce a test version instead of a few hours.  (Normally, I can make good changes in a few hours -- recently, MOST of the time delay on producing releases is normally for making&verifying the test decodes.)

 

We are 'getting there'!!!!
 

 

 

 

 

 

 

[John Dyson]

WOW -- one of the very kind reviewers made a very simple comment:

 

paraphrased:  hows about a little more upper midrange and a little less lower midrange?

answer:  remove two equalizers

result: more open sound.

analysis: more simple, obviously more 'natural' EQ design.

 

Then, when I compared with one of the never-FA recordings, I couldn't tell which was more correct.  (this was a vinyl copy -- not quite as good as digital, sorry!!!)   However, I have a pure digital copy of another recording - the change results in almost totally perfect representation.   Which recording?   title song of Supertrouper!!!  One of the baddies.

 

This is a very interesting result, and will be doing a LOT of tests!!!

 

John

 

[John Dyson]

About the last comment about correcting the lower/higher midrange.  By doing so, a lot of recordings really clear up and have fewer defects.  It doesn't fix all defects as some recordings do have defects, but there are some real quality improvements in some cases...

 

Dionne Warwicks' vocals are less dull/muddy.

The cymbals on Carpenters old recordings are less odd.

Anne Murrays vocals are a bit more clean...

 

Lots of other improvements also...  I do SO appreciate those with better hearing and better listening judgement.  Even I, after listening to the correction can hear the improvement.  THANKS SO MUCH TO THOSE WHO HAVE BEEN HELPFUL!!!

 

 

[John Dyson]

Status report:

Currently, just completed the fourth iteration for this upcoming release.      There are now objective measurements that allow understanding the gross level of spectral balance errors, but it is very tricky to compare to get sub-dB accuracy. The comparision of the larger numbers of the NEVER FA materials vs decoder output now show the mastering differences, and not a serious qualitative difference.   It appears that the decoder output tends to be superior over some NEVER FA materials because of the anti-distortion reaching back to remove previous DolbyA encoding/decoding distortions.

 

*  Recently, the whole EQ system was reworked,  Decided to ignore previous suggestions to use 1st order EQ, and have successfully allowed using carefully considered 2nd order EQ with fantastic results.  The recent releases have shown significant improvement and much more positive feedback from testers.

 

The current challenges are:

1) HF vs MF balance.   Even doing it by measurement is tricky because the density of the FA source material changes its gain against the original, nonFA version.  Right now, I believe that the decoder is providing a <0.50dB stronger output at 3kHz on up than it should, but some   other  tests say that isn't true.   Some comparisons show that it is dead-on accurate.   Not sure which evaluation is more accurate, but might have to release this as is.   It does NOT sound like the highs are overly strong, but a subjective comparison isn't adequate, and there are now some subjective measures that really help.

2) LF limit.   It seems like there might be a sub-dB excess bass in the 30Hz range.   It manifests as a slightly strong bass, but not as extreme as previous releases.

3) HF boost?   The FA sound is compressed in the 3kHz-20kHz range, and is actually audibly stronger than natural in the 3k-6kHz range or so.   This had tempted me to add approx 0.50dB boost in the 6kHz range, and currently the test version does this by default (disabled automatically in some cases.)   I believe that it is better to disable this boost by default, and take the criticism about the highs not being as intense.   (The highs are actually stronger on the decoder output, but the density is less because the material is less compressed.)

 

The good news is that the decoder is working perfectly on material that was previously a challenge.   For decoding purposes, the FA decoder is no longer making errors, but the EQ will always be a matter of opinion.

There will probably have to be more more iteration, which will delay by another 12Hrs at least.

 

 

[John Dyson]

WOW -- this has been a very rough couple of days.   As mentioned before, the excess HF has been a problem for quite a while.   There was always a balancing act between 'burning hearing' with too much HF vs.  HF and secondary distortion elsewhere, and I realize the excess HF as been a problem for a very long time but wishful thinking had controlled my judgement.   There was a reason for the 'too much HF' madness, but the alternative was also bad.   The ONLY reason why I could tolerate 'too much HF' is my very limited hearing >12kHz MOST OF THE TIME (sometimes it actually works much higher.) 

 

I am really hoping that this is the last layer of the 'onion' so that the HF problems are fully mitigated.   It *really* took a full understanding of what is going on to intuit the need for an additional  1st order before/after EQ.    This additional EQ 'nudged' the HF downwards in the cases where there was excess but still maintains the frequency response for the less problematic situations.   Hopefully more people will be happy with these upcoming results!!!

 

A release will NOT be done until I am fully satisfied, but will keep the interested parties up-to-date with a 'heartbeat' of status reports.  The HF problem is of paramount importance, and solving it is definitely in the release criteria.

It is VERY POSSIBLE that a release can be done tomorrow morning (+15Hrs), but any problem with final, last minute QC tests will force another day delay.

 

John

 

[John Dyson]

Status:  delayed another day.

*  Important note -- I forgot to mention previously -- Historical 'hole in the bass' has been fixed for a while in the development decoder.   Using 1st order EQ simply cause frustration and poor results.   Using 2nd order EQ now, much easier, sounds better.

*  The HF dynamics are still not correct today.   The dynamics are very difficult because of accomodation, and almost impossible to measure.

 

There is an HF spectral measurement now, and that statistic helps to maintain a relatively sane HF response.   Likewise, the LF response is also maintained as relatively consistent.

Unfortunately, as I had mentioned in the past, the decoder can NOT be 'flat' in the conventional sense, however. comparisons can be used to track behavior.  There still isn't a good method to track the dynamics, so it is still a tredious iterative procedure.

 

I fully expect the release will be likely before the weekend.   The release will NOT happen until I am totally happy with it.

 

 

 

[John Dyson]

Unless there is a crash during running the demos or a major botch, the release should be tonight by +13Hrs from now.

A lot of fixes have been done, the most frustrating problem is that my hearing changes by like -10dB at 10kHz through the day, so I get confused from the normal, simple EQ that the decoder is using.

The new EQ is truly simple (but still does what it is supposed to do.)   Almost all final EQ is intended to undo the pre-emphasis, as usual, no real tweaking.   There is an optional  'finish up' that does a combo rolloff, small peak whose center is 21kHz, but that peak is <0.5dB and only has the effect of slightly improving the definition without getting edgy AT ALL.  *(Yes, I can hear near 17kHz for an hour or so in the morning.)

 

The new version uses a hybrid 1st order/2nd order approach.  With the new EQ, some of the most difficult recordings now sound almost pristine.

 

The most interesting revelation has been the DolbyA compatible steps had two important 'bugs', but also had some calibration issues.   The calibration issues could not be verified by any objective test materials that I have.   Since the FA decoder uses 8 layers of active DolbyA compatible decoding, any DA errors are magnified very strongly.  Most of the errors cannot be easily detected in pure DA mode except for the highs are a little too harsh.  

The bugs included 1)LF threshold about 3dB too low (too much <120Hz LF) 2) the HF0/HF1 split feedback was too fast, giving an edgy sound, 3) The dynamic attack/release scheme had a bug in the gain feedback, therefore made the attack/release too fast, 4) Very carefully tuned attack/release slowdown, including consideration of the faster HF1 control, keeping the spectrum < 10kHz, avoiding too much spectral wrap around.    Lots of other minor DA improvements also.   This release is NOT ready to use for pure DA yet as I have a lot more stringent testing and verification before there can be an official DA mode.   I had thought about disabling DA (easy too do, capability already built in), but that would not be helpful anyway...   Just don't depend on it.

 

All of the FA EQ is still very similar as before, but not distorted by bad hearing 'reworks' through the day.  Also, found that the lower MF EQ must NOT be flat, or you get the 'hole in the lower MF' matter that had been previously complained about.   There is a pronounced curve around 100Hz to 300Hz so that the FA bass is matched (and sounds more FA-real.)   My attempts at flat bass put a hole in the 100Hz to 300Hz range, but instead gain should have been allowed to increase.   To make the bass flat, there needed to be explicit and tricky EQ to make it so.   The EQ for the more FA sounding bass is more consistent and no special mods around 100Hz to 300Hz.   The EQ is now no longer 'tricky'.

 

The LF pre/de-emphasis was made more moderate.   There is still very good stability of mixing the LF/MF/HF sounds at low level, but also the bass is more tight.

 

There are now some objective measures for decoding results, and finding no strange readings.   The decoder can NOT be flat, or hearing will burn and the bass will have the 'holy' sound :-).   Those problems shouldn't be manifest anymore.

 

Again, if everything goes well, release tonight.   This HAS been a very detailed evaluation and re-jiggering of the EQ WHICH IS NOW MUCH MORE SIMPLE.  If there are any pronounced HF problems, they will be trivial to fix, but I need to know about them.  The HF should really be VERY VERY clean now, just don't trust my hearing anymore  -- maybe the EQ should be really contorted -- but that would be unreasonable.

 

 

John

 

 

 

[John Dyson]

Release delay for a day or so -- personal matters.  The decoder is doing well.   Actually, made a few more improvements yesterday evening.

Figure +3 days for now.   It might be ready +1 day if I can free myself up from a personal activity.