Hertz Overload - new demo for AGA Amigas

[Gorf]

Originally Posted by pandy71

It should be averaged over time and if volume adjustment occurs at least twice per sample then everything should be fine.This is job for reconstruction filter.I'm curious how this volume regulation is implemented in Paula - i can imagine various solutions (like feeding to DAC reference voltage from PWM, interrupting analog signal from DAC by switch controlled by PWM or feding data into DAC from some register but data are controlled by PWM - so DAC is switched between register and for example 0 level - all of them seem to be valid).Good question!

no - it just sets the digital output signal to zero(1) after the volume counter reaches the number stored in the volume register. So never for 64 and after one click for 1.

To get to "zero" it alternates pulses:

So what Paula does is to change its operation mode for volume 63 and below. It establishes the aforementioned raster of T = 64/3579545s in which it will perform an operation cycle similar to a PWM. Within each cycle, an impulse is generated which will be canceled out by a negative impulse of the same amplitude after a number of time steps corresponding to the volume setting, i.e. Tv = v/3579545s with v ? [1...63].

http://bax.comlab.uni-rostock.de/dl/...mTheoretic.pdf

[no9]

Originally Posted by Karlos

I was just curious about the general problem. Since quantisation noise is strongly correlated with the input signal, it's more conspicuous than white noise. Does a basic noise injection pre-quantisation help change the perception of the quantisation noise into something less intrusive?

Yes, but it is a matter of taste which one you prefer or consider as 'less intrusive'. (video hosted temporarily) https://streamable.com/9v5xz4

[pandy71]

Originally Posted by Gorf

no - it just sets the digital output signal to zero(1) after the volume counter reaches the number stored in the volume register. So never for 64 and after one click for 1.To get to "zero" it alternates pulses:

So what Paula does is to change its operation mode for volume 63 and below. It establishes the aforementioned raster of T = 64/3579545s in which it will perform an operation cycle similar to a PWM. Within each cycle, an impulse is generated which will be canceled out by a negative impulse of the same amplitude after a number of time steps corresponding to the volume setting, i.e. Tv = v/3579545s with v ? [1...63].

http://bax.comlab.uni-rostock.de/dl/...mTheoretic.pdf

This is not clear to me - Paula schematic is required - you can imagine sample value is feed to DAC and current from DAC is sent to switch controlled by AUDxVOL where there is 65 states - 0, 63 levels and PWM disabled (i.e. full scale) as PWM with 6 bit counter will be from 0/64 (i.e. zero) to 63/64. Until Paula schematic is done then there is big question mark on this.

[Gorf]

Somewhere here on EAB Toni uploaded a oscilloscope screenshot of that volume reduction behavior, but sadly I can't find it anymore ...

only this quote:

Originally Posted by Toni Wilen

Scope tests above 100% prove it can't use 14-bit DAC:
- Pulses (Frequency = CCK/64) even when sample value does not change, except when volume is 64.
- Pulse width = AUDxVOL value.

https://eab.abime.net/showpost.php?p...2&postcount=18

[Karlos]

Originally Posted by no9

Yes, but it is a matter of taste which one you prefer or consider as 'less intrusive'. (video hosted temporarily) https://streamable.com/9v5xz4

Both noise methods definitely a bit better here than the straight quantisation. I am thinking of this in the broader context of a volume-normalised replay similar to those discussed already.

[copse]

Originally Posted by Gorf

Somewhere here on EAB Toni uploaded a oscilloscope screenshot of that volume reduction behavior, but sadly I can't find it anymore ...

only this quote:


https://eab.abime.net/showpost.php?p...2&postcount=18

Attachment at bottom of @absence's post?

https://eab.abime.net/showpost.php?p=824979&postcount=6

[Gorf]

Originally Posted by copse

Attachment at bottom of @absence's post?

https://eab.abime.net/showpost.php?p=824979&postcount=6

no, that's not it ... it was a constant signal or a rectangular signal and different volume settings showing the pulses, Toni was talking about, cutting stripes out of this signal
(maybe it was even a short video?)

[saimo]

@Pandy

I ran the scripts using the last 5 seconds of the Sony example music.
The resulting sound, in both cases, differes strongly from the original one and is affected by perceivable noise. For comparison, I also downsampled the same snippet with Audacity to 16-bit and then converted it to 14-bit. Despite all the issues that affect the 14-bit playback, it still is better by far.

Here are the files:
original, 96000, 24-bit
Audacity, 64489, 14-bit, AQA1
Audacity, 64489, 16-bit, WAV
Audacity, 192000, 32-bit, WAV, sampled from A1200
script 1, 64489, 8-bit, AQA0
script 1, 64489, 8-bit, WAV
script 1, 192000, 32-bit, WAV, sampled from A1200
script 2, 64489, 8-bit, AQA0
script 2, 64489, 8-bit, WAV
script 2, 192000, 32-bit, WAV, sampled from A1200
waveforms and spectograms pictures

Side note 1: the second script also generated a 48000 Hz 24-bit file, which I did not test.
Side note 2: the WAV->AQA converters processed the WAVs just fine, so, even if the scripts specify signed-integer, the data must be unsigned anyway (I searched around it does look like WAVs are always unsigned when the resolution is 8-bit).

[pandy71]

Originally Posted by saimo

@Pandy

I ran the scripts using the last 5 seconds of the Sony example music.
The resulting sound, in both cases, differes strongly from the original one and if affected by perceivable noise. For comparison, I also downsampled the same snippet with Audacity to 16-bit and then converted it to 14-bit. Despite all the issues that affect the 14-bit playback, it still is better by far.

From my perspective 8 bit NS audio deliver same quality as 14 bit (more than 84dB dynamics).

You can try to generate 16 bit with 14 bit or lower precission and with noise shaping - modifying parameter 'p' (dither -f %noiseshp% -p 8) from 8 to for example 12 or 13 or 14 overall noise level perceived by noiseshaping will be significantly suppressed - with '-p 12' you probably will not hear any noise yet overall dynamics will be improved.
Of course instead 8 bit file you need to use 16 bit file - also you may wish to apply level reduction to 14 bit (by applying gain 'gain -n -12.0411998265592' before so in theory 16 bit will have 14 bit data and noise shaping to additionally improve dynamics and can be mapped directly to Amiga).

Originally Posted by saimo

Side note 1: the second script also generated a 48000 Hz 24-bit file, which I did not test.
Side note 2: the WAV->AQA converters processed the WAVs just fine, so, even if the scripts specify signed-integer, the data must be unsigned anyway (I searched around it does look like WAVs are always unsigned when the resolution is 8-bit).

This 48kHz 24 bit file is ideal representation what is possible with noise shaping - assuming perfect Paula DAC (i.e. ENOB 8 bit - ideal linearity) and no other artifacts you should get as the real HW Amiga audio output such quality - i assume A500 low pass filter characteristic if your Amiga source is A1200 tehn depends on revision you may (you should) set different low pass frequency in script (i used data from already referenced by Gorf https://eab.abime.net/showpost.php?p...3&postcount=81 paper about Paula theory).

Checked carefully logs - obviously SoX silently ignore switch to produce signed integer 8 bit PCM wav and use unsigned integer 8 bit instead so your assumption on wav file is probably right- they accept in 8 bit PCM only unsigned integer and that's why your converter works correctly.

[saimo]

Originally Posted by pandy71

From my perspective 8 bit NS audio deliver same quality as 14 bit (more than 84dB dynamics). You can try to generate 16 bit with 14 bit or lower precission and with noise shaping - modifying parameter 'p' (dither -f %noiseshp% -p 8) from 8 to for example 12 or 13 or 14 overall noise level perceived by noiseshaping will be significantly suppressed - with '-p 12' you probably will not hear any noise yet overall dynamics will be improved.Of course instead 8 bit file you need to use 16 bit file - also you may wish to apply level reduction to 14 bit (by applying gain 'gain -n -12.0411998265592' before so in theory 16 bit will have 14 bit data and noise shaping to additionally improve dynamics and can be mapped directly to Amiga).

I don't mean to belittle your efforts, but the problem is that, due to the filtering, the resulting sound is noticeably/completely different from the original (to the ears, I mean). The quality doesn't matter anymore if the result differs (so much). Even bare 8-bit quantization, with all its noise, sounds more faithful to the original.

https://eab.abime.net/showpost.php?p...3&postcount=81 paper about Paula theory).

The bit about the impulse cancellation method for volume regulation is very interesting and, I think, confirms that non-64 volumes can't be precise when the period is not a multiple of 64.

[Gorf]

Originally Posted by saimo

I don't mean to belittle your efforts, but the problem is that, due to the filtering, the resulting sound is noticeably/completely different from the original (to the ears, I mean). The quality doesn't matter anymore if the result differs (so much). Even bare 8-bit quantization, with all its noise, sounds more faithful to the original.

Judging from the resampled output, you are right. Something has gone wrong here.

Some may be due to the differences in A500 vs. 1200 low-pass. This needs to be addressed.
I actually would not apply any equalizer or compensation for the Amiga low-pass just yet. This can be done a a very last stage, when everything else is working. Otherwise it is hard to tell apart what is doing what..

The dithering is also overdone in my opinion and should be scaled down - it shouldn't be that noticeable but barely audible. Just used a a trick to assist the noise shaping step.

[pandy71]

With more aggressive noise shaping filter, better tuned to Amiga sampling frequencies (i mean not standard like 44.1, 48 etc) you can achieve IMHO very good result.

Made some filter for Sebastian G. application (noise.zip in the Zone) and results are more than 16 bits dynamic (approx 112dB so closer to 18 bit).

Providing filter (added txt extension to upload filter - sos is just set of coefficients for filter), small script and spectrogram on dynamic test (signal level sweep). Sadly to say but requent.jar not accepting 24 bit files (i failed to requantize 24 wav so forced to use 16 bit as source - seem java has some issue with opening 24 bit files - perhaps there is issue with some wav structure as author mention support for 8..24 bit support), detected this recently - seem also there is issue with size of file - even few ten MB seem to be no go for this software - as such there is PITA unless someone will write software doing requantization in similar fashion - i'm not coder...

Script accepting source filename as first parameter and noiseshaping filter as second parameter - there is possibility to change gain (to prevent dither clipping when signal is close to 0dBFS - this is visible on spectrogram) and also allow to change dither type (float in range 0..2) - based on author description:

        0   = off
        0.5 = rectangular (+/- 1/4 LSB)
        1   = full rectangular (+/- 1/2 LSB)
        1.5 = half-way between 1 and 2 (morphed)
        2   = full triangular  (+/- 2/2 LSB)

@setlocal
@echo off

@SET file=%1
@SET fname=%~n1

@SET GAIN=1.0
@SET DITHER=1.0

@java -jar requant.jar %file% %fname%_%~n2.wav 8 %GAIN% -d%DITHER% -s@%2

@endlocal
@pause

[saimo]

@Gorf

Originally Posted by Gorf

I actually would not apply any equalizer or compensation for the Amiga low-pass just yet. This can be done a a very last stage, when everything else is working. Otherwise it is hard to tell apart what is doing what..
The dithering is also overdone in my opinion and should be scaled down - it shouldn't be that noticeable but barely audible. Just used a a trick to assist the noise shaping step.

Previously I had made tests with just dithering and noise shaping, using various tools (see post 33 and post 63) and in no case the results were satisfying.


@pandy71

If I'm not asking too much, could you to provide directly the sample processed with your latest script, please? I don't fancy installing Java just to make a test.


@all

I just finished making other tests. I think that they prove that:
* Paula resets the volume counter also for the second sample;
* volume works better than expected (although not perfectly) also at periods < 64.

I'll provide pictures and files in the evening, after work.

[pandy71]

Originally Posted by saimo

If I'm not asking too much, could you to provide directly the sample processed with your latest script, please? I don't fancy installing Java just to make a test.

No problemo Saimo - here are files: https://eab.abime.net/zone/ForSaimo.7z

[saimo]

Here are the findings from my latests tests.
Please keep in mind that all the equipment I have this humble ASRock J5040-ITX PC with its anonymous on-board audio chip, my A1200 and a cheap audio cable connected to the A1200 RCAs and the PC line-in. I sampled the output of the Amiga at 192000 Hz, the maximum allowed by the PC (initially I had used 55420*3 = 166260 Hz, but that was too little for period 24, it wasn't exact because 3546895/64 is not 55420 precisely anyway and because there is no way to sync the start of the playback and of the recording).
Click the pictures to see them in full size.

I generated a 13855 Hz square wave, sampling it at 13855*4 = 55420 Hz (so that, for each cycle, there were 2 "up" samples and two "down" samples ) and used it for both the left and right channels:



Below is the output sampled with various volumes and periods.

First off, output at volume 64 and period 64, to see if Paula's output is sane with such "perfect" values:



(The output got clipped a little bit, so it was a bit more than 60 dB, contrary to the previous measurements made with a sinusoid.)

Then, output with halved period (32) - given that the volume remained 64, the amplitude was expected not to (and did not) change:



Another test to check that the output is sane in normal conditions - here the left channel is played at volume 32 (with the period restored to 64 for both channels):



Now the first "real" test: what happens with volume and period both set to 32? If one looked at just the counters, the amplitude should always be maximum, as the volume counter never reaches 32; instead, Paula presents a nice surprise:



Paula adapted to the period! However, the adaption was not perfect:



Consider especially the more wavy contour and the spikes, as the initial part is irregular also when the parameters are perfectly "normal": it's as if Paula needs a settling period (sorry, I was too tired to investigate this specific aspect; EDIT: it occurred to me later that this might be caused by a bug in the buffering initialization of the player; EDIT2: nope, the code is fine; also, checking the waveforms, it turns out that the settling period is about 2000*period/64 samples long - maybe Paula uses an internal buffer of 2 kB per channel? EDIT 3: as reported later, the problem was in the source test sample - please ignore these ramblings about the settling period); here's a zoomed in view of the samples beginning:



Also, the amplitude attenuation was inferior (I measured the amplitude where there are no spikes; the absolute values are for sure totally unreliable, but their comparison is definitely meaningful):
* period 64: -6.128 dB (keep in mind that this is affected by clipping; the AHRM says it should be -6 dB)
* period 32: -5.180 dB

Such surprise suggested me to check the output also at period 50 (the one used by the demo) and 24 (a very small one; I didn't go further because the corresponding frequency is already 147787.292 Hz, i.e. a bit too much for the 192000 Hz sampling limit):



I measured the attenuation also in these cases; putting them together suggests a (linear? the measurements are way too imprecise to say) relationship between period and volume according to which the smaller the former, the less accurate the latter:
* period 64: -6.128 dB
* period 50: -5.331 dB
* period 32: -5.180 dB
* period 24: -4.883 dB

Then I repeated the same procedure for volume 1, to get an idea of how (much more) distorted is the 14-bit playback at a period < 64:



* period 64: -34.921 dB (the AHRM says it should be -36.1 dB)
* period 50: -33.672 dB
* period 32: -32.291 dB
* period 24: -31.772 dB

(Wishful thinking follows) With precise measurements, maybe it could be possible to determine the (linear? logarithmic?) equation to recalculate the amplitude of samples to compensate the distortions introduced by periods smaller than 64. (Note: my bet is that some distortions happen also with periods greater than 64 but not multiple of it, but I'm just too tired to make other tests - this stuff is taking too much of my time and limited energies...).

Finally, while at it, I thought I'd make a test to see whether the volume counter resets for the second sample of each word. I shifted the left channel by 1 sample and played the channel at volume 32 and period 32: if the counter did not reset, the left and right channels would look definitely different; instead they look "perfect":



(Note: maybe the volume compensation mechanism of Paula plays a role?)

I suggest that you have a closer look at the data yourself: WAVs and pictures archive.

[saimo]

Originally Posted by pandy71

No problemo Saimo - here are files: https://eab.abime.net/zone/ForSaimo.7z

Thank you!

I tested just the last 10 seconds or so, as I don't have a quick way to transfer large files to the A1200 and, anyway, those 10 seconds are enough for quality evaluation. Here are the results: 8-bit 16-bit

Unfortunately, in both cases the noise is pretty audible (especially in the 8-bit case), but, more importantly than that, the problem is that the sound is still just too different: the lower frequencies got lost, so both sound much thinner than the original.
A consideration: I don't think that for 14-bit playback any (heavy) dithering/noise shaping is necessary, as the sound is already clean as it is - there, the problem is the distortion due to volume regulation as discussed in the previous posts. Check out the demo music again: that's been exported from Audacity without any dithering or noise shaping (and the output is also affected by the lesser volume attenuation due to the period 50, as reported in the previous post). Afterwards, I tried also those options, and I have a really hard time distinguishing the results.

After all these tests, I think that the best solution to achieve high quality are:
* sticking to period 64 to avoid volume distorsions;
* 14-bit playback with or, to avoid calibration issues...
* ... 8-bit + amplitude modulation (hmmm... isn't also AM subject to calibration issues, after all?)

[pandy71]

Originally Posted by saimo

Thank you!

Unfortunately, in both cases the noise is pretty audible (especially in the 8-bit case), but, more importantly than that, the problem is that the sound is still just too different: the lower frequencies got lost, so both sound much thinner than the original.

During process mid and high frequencies are boosted by +24dB (almost) as such signals lower than 4kHz are "weaker" - audio may sounds thinner - low pass filter present in almost all Amiga models (except A1200) - attenuate those boosted frequencies so as result should be flat response - you can activate CIA controlled filter to get similar result - overall lower signal level can be compensated by increasing sound level (at the amplifier).

Of course i can prepare samples without such boost and also sample length to 10 seconds to address your setup specification.

Originally Posted by saimo

A consideration: I don't think that for 14-bit playback any (heavy) dithering/noise shaping is necessary, as the sound is already clean as it is - there, the problem is the distorsion due to volume regulation as discussed in the previous posts. Check out the demo music again: that's been exported from Audacity without any dithering or noise shaping (and the output is also affected by the lesser volume attenuation due to the period 50, as reported in the previous post). Afterwards, I tried also those options, and I have a really hard time distinguishing the results.

Noise shaping is beneficial always - in 14 bit samples you can compensate 2 bit loss also as overall quantization error has significantly lower amplitude then noise shaping itself will be inaudible (8 bit quantization error is still pretty high but from 10..12 bits is inaudible).

Originally Posted by saimo

After all these tests, I think that the best solution to achieve high quality are:
* sticking to period 64 to avoid volume distorsions;
* 14-bit playback with or, to avoid calibration issues...
* ... 8-bit + amplitude modulation (hmmm... isn't also AM subject to calibration issues, after all?)

Spectral analysis shows that 8 bit + NS may work especially if higher oversampling ratio is possible and can be viable alternative.
That's why i provided 24 bit files with "play" in name - this is 24 bit produced from 8 bit samples - only lowpass filtering (to simulate Amiga audio signal path) and resampling (to deal with PC audio card specify).
Except this two steps you can hear 8 bit sound. Amiga need no resampling and automatically apply lowpass filtering so in analog domain this should work fine.


Btw about 14 bit with noise shaping - probably this description will the best argument for noise shaped 14 bit:

Originally Posted by https://www.dutchaudioclassics.nl/the_evolution_of_dac_the_digital_filter/#philips-tda1540

Philips TDA1540

The evolution in the use of digital filters and DACs, which greatly influence the sound quality of Super Audio CD/CD players, is also a fascinating topic. In the Red Book, which defines CD standards, the resolution is quoted as being 16 bit.

However, the first DAC used in the first Marantz CD player, the Marantz CD-63, and the Philips LHH-2000, was the 14-bit resolution Philips TDA1540. Since Japanese CD players all employ 16-bit resolution DACs, the Marantz/Philips specification would seem inferior, but according to data taken from actual performance situations and among audiophiles with a good ear for sound quality, the performance of the TDA1540 CD player with its 14-bit DAC is rated outstandingly high.

In actual fact, the secret to the excellent sound produced by the Philips SAA7030 digital filter incorporated into the TDA1540 is not obvious at first. The SAA7030 with a 4x over sampling filter capacity utilizes a top quality noise shaping circuit called a secondary noise shaper. When the Marantz CD-63 went on the market, the effect of noise shaping (imposing frequency characteristics onto distribution of quantized noise in order to shift noise component to an ultra-high ranges that are difficult to hear with the naked ear) was lagerly unknown. The combination of the SAA7030 and TDA1540 used noise shaping to achieve a resolution equivalent to 16-bit.

[Karlos]

I don't know if you guys have seen this, but there's a lot of interesting info here: http://bax.comlab.uni-rostock.de/dl/...mTheoretic.pdf

[Gorf]

Originally Posted by Karlos

I don't know if you guys have seen this, but there's a lot of interesting info here: http://bax.comlab.uni-rostock.de/dl/...mTheoretic.pdf

Yes - I linked to it already here in the thread.

[saimo]

@all

Originally Posted by saimo

... as the initial part is irregular also when the parameters are perfectly "normal": it's as if Paula needs a settling period (sorry, I was too tired to investigate this specific aspect; EDIT: it occurred to me later that this might be caused by a bug in the buffering initialization of the player; EDIT2: nope, the code is fine; also, checking the waveforms, it turns out that the settling period is about 2000*period/64 samples long - maybe Paula uses an internal buffer of 2 kB per channel?); here's a zoomed in view of the samples beginning:

Please disregard this part: today I didn't have any time to sit at the Amiga, but now I had a quick look and found out that it was the source tone to be broken! That first part was as at 55420/4 = 13855 Hz as intended, but the rest was at 55420/10 = 5542 Hz!I'll have to redo the tests. Now I just did quick and different one that seems to indicate that things are worse than I reported in my previous post, but I'm too exhausted and I might just be messing things up once again. Tomorrow or in the next days I'll return to it.