Amiga "14-bit audio" refuted

[chb]

Originally Posted by roondar

I guess I'm asking if it's:

• Paula amplitude modulation affects digital value of sample, staying inside of the -128/+127 range
• Paula amplitude modulation affects channel volume, staying inside of the 0-65 range
• Paula amplitude modulation is a separate process from the above (and therefore can change sample output outside of the normal 8-bit limits/outside of normal 0-65 volume range)

AFAIK it is No. 2 - when choosing AM or FM, the word-sized valued fetched by DMA are just written to the volume or period register:

Originally Posted by http://amigadev.elowar.com/read/ADCD_2.1/Hardware_Manual_guide/node00E7.html

When a channel is used as a modulator, the words in its data set are no longer treated as two individual bytes. Instead, they are used as "modulator" words. The data words from the modulator channel are written into the corresponding registers of the modulated channel each time the period register of the modulator channel times out.

[bloodline]

Originally Posted by buggs

The question about the A600 has been settled earlier in here, but I'd like to add some words. I extracted the previously stated numbers for the A600 filter from a schematics PDF that's circulating, which in turn actually contained A1200 Rev1d specs in the audio filter.
As luck would have it, I was just handed an A600 for fixing. And unsurprisingly by now, that A600 has the same filter as A500/2000, 100nF+390Ohms.

That’s interesting, so the A1200 really is the odd one out! Has anyone got the CD32 schematics to check what the output filter is on there?

The A1200 cap rating is something like an order of magnitude out, is it possible a typo in the design?

[Gorf]

Originally Posted by Thomas Richter

Findings:
1) The phase drift on the Amiga is approximately one magnitude larger than on the PC. In other words, the Amiga cannot really reproduce the target frequency precisely, less precise than the PC. That is probably no surprise given that the PAULA frequencies cannot match precisely the target frequency (44100 Hz).

Therefore the whole test setting is biased:
To really make a useful comparison, you need an analog audio wave a staring point.
Each machine must be allowed to sample and playback that tone at a sample frequency that best matches it’s inner workings.

[ross]

Originally Posted by Gorf

Therefore the whole test setting is biased:
To really make a useful comparison, you need an analog audio wave a staring point.
Each machine must be allowed to sample and playback that tone at a sample frequency that best matches it’s inner workings.

http://eab.abime.net/showpost.php?p=...8&postcount=55

But I think that cannot be never compensated enough, Paula quantization levels are too raw (the PCM part) and the PWM need to be handled as a separate effect from pure sampling.

If the original wave is sampled at CCK submultiple then probably all you need is a good calibration to get from Paula her best, probably not much but a little better .

[pandy71]

Originally Posted by Thomas Richter

I don't care what people claim. I'm a physicist. I care what I can measure.

Originally Posted by Thomas Richter

I'm not even sure what you want to say here.

Originally Posted by Thomas Richter

Findings:
1) The phase drift on the Amiga is approximately one magnitude larger than on the PC. In other words, the Amiga cannot really reproduce the target frequency precisely, less precise than the PC. That is probably no surprise given that the PAULA frequencies cannot match precisely the target frequency (44100 Hz).

My point is that you should be more familiar how to measure digital audio... Amiga cannot create 44100Hz sample rate - Paula can create only $4F 44897.41Hz, $50 44336.19Hz, $51 43788.83Hz.
FFT should be performed coherently, you should apply proper windowing and few other things. You should avoid signals with frequencies being "nice" numbers (to test each DAC code word, stimulus shall be uncorrelated with sampling rate - that's why audio industry use 997Hz instead 1kHz - with proper testing time and averaging you can test whole DAC codebook i.e. get proper DAC dynamics and SNR as dynamics is not SNR).

Can't upload few other pdf's due size limitations for attachments.

[no9]

Actually... that begs the question who and how measured Paula 14 bit method as 14 bit in the first place. If ever...

[ross]

Originally Posted by no9

Actually... that begs the question who and how measured Paula 14 bit method as 14 bit in the first place. If ever...

It was not measured at all, 14 bits of a 16-bit sample were used with a separation between 8 high (max volume) and 6 low bits (min volume)

[Gorf]

Originally Posted by ross

http://eab.abime.net/showpost.php?p=...8&postcount=55

But I think that cannot be never compensated enough, Paula quantization levels are too raw (the PCM part) and the PWM need to be handled as a separate effect from pure sampling.

True ...here the question would be: how to really take the PWM effect into consideration - it probably would require a more complex mathematical operation to turn a raw 16bit sample into a *propper* 8bit PCM + 6bit PWM representation.
(Plus things like sound dithering by using high sample-frequency)
But that step would in a fair comparison be part of the sampling-process.

So the question still remaining is, how close can Paula match a certain analog tune (using all possible tricks) in comparison to standard 16bit PC audio.

[no9]

Originally Posted by ross

It was not measured at all, 14 bits of a 16-bit sample were used with a separation between 8 high (max volume) and 6 low bits (min volume)

You are right. I wondered if this was ever confirmed what we effectively get at the end of Amiga audio signal path.

[nikosidis]

Originally Posted by Gorf

True ...here the question would be: how to really take the PWM effect into consideration - it probably would require a more complex mathematical operation to turn a raw 16bit sample into a *propper* 8bit PCM + 6bit PWM representation.
(Plus things like sound dithering by using high sample-frequency)
But that step would in a fair comparison be part of the sampling-process.

So the question still remaining is, how close can Paula match a certain analog tune (using all possible tricks) in comparison to standard 16bit PC audio.

Put on some good headphones or good speakers and listen to the sound examples I posted in this thread. There are 2 different tunes with the reference and the one I output from my A1200. I would say it match 16bit PC Audio good.
Voices, instruments sound real. Even the dynamic range is good. The Amiga 1200 sound even better for real than the stream I put up. When Audacity join the Amiga L,R channels there are some info that get lost in the mixing.
I understand Thomas wanted to proof that Amiga can not do 14bit and I'm sure he have. It does matter little anyway. 10bit or whatever is good enough. Put down the kHz and the audio quality will suffer and not sound real. How much depends on different instruments etc. I would say it is very easy to hear on voices when the kHz is 22Hz instead of 44.1Hz. From what I understand bit is the noise floor. Difference between quiet and loud.

16-bit is insane. Only some orchestral recordings take advantage of something close to that. All pop, rock music are compressed and have little dynamic range. When there are quiet passages you will hear some noise from the Amiga, but with latest EaglePlayer he got that noise very low. Color me impressed!

Sorry to all that have other Amiga models cause Amiga 1200 sound by far the best.

[Bruce Abbott]

Originally Posted by no9

Actually... that begs the question who and how measured Paula 14 bit method as 14 bit in the first place. If ever...

Originally Posted by ross

It was not measured at all, 14 bits of a 16-bit sample were used with a separation between 8 high (max volume) and 6 low bits (min volume)

By definition, a DAC that uses 14 bits to produce its output is 14 bit. However the result may not be be accurate to 14 bits.

Originally Posted by Thorham

Yes, i was wrong about the audio being 14 bit

Actually you are both 'right'. In 14 bit mode Paul has a resolution of 14 bits, but the output is typically not monotonic and has other artifacts that reduce the measured bit depth.

Other DACs can also have these issues. For example the Burr-Brown DAC56, a 16 bit DAC commonly used in early CD players, has '12-bit monotonicity guaranteed'. So are all the CD players using this chip actually only 12 bit? Of course not.

Another factor that 'scientific' tests often don't take into account is the response of the human ear, including psychoacoustic effects. For a fair comparison the signal should at least be 'A' weighted, so that components outside the normal hearing range are suppressed. Signals should also be evaluated at different levels corresponding to the dynamic range of music, rather than just at maximum volume where audio masking tends to suppress low level noise and distortion.

Depending on the type of distortion, a system may sound better or worse than scientific measurements suggest. In the old days of class A tube amplifiers an amp with 10% harmonic distortion at rated output sounded pretty good because the distortion was mostly low order harmonics that gradually increased as the volume level increased. Then Class B transistor amps with large negative feedback were introduced. The large negative feedback greatly reduced high volume distortion, but caused hard clipping on the slightest overload, and crossover distortion produced high order harmonics that were plainly audible even when measurements showed less than 1% distortion. What's worse is the crossover distortion became more prominent as signal level reduced. In response to this audio engineers demanded even lower distortion figures that would be way below audible levels in a class A amp.

If 14 bit Paula audio can reproduce sound with sample amplitudes below the minimum quantization level of 11 bits then it is more than 11 bits. There may be a lot of distortion at that level, but it still has the required resolution to reproduce the low level sound, and that gives it the dynamic range needed to play 'hi-fi' music (which is the biggest problem with standard 8 bit audio).

In the end it comes down to what it sounds like to the human ear, so while listening tests may be 'unscientific' they are just as valid or even more so than naive 'scientific' tests. Instruments have the advantage of being objective, but the data can be misleading or even irrelevant if not related to the real goal, which is our perception of the sound quality.

Unfortunately (or fortunately) many of us are getting on in years and our ears aren't what they used to be. I can barely tell the difference between the filter being on and off in my A1200, because my hearing is 30dB down above 3kHz and I have constant tinnitus which masks higher frequencies. The older I get, the better Paula sounds!

[malko]

Originally Posted by Bruce Abbott

[...] I have constant tinnitus which masks higher frequencies. The older I get, the better Paula sounds!

. A grandmother's hint : Rub a cotton bud with sweet almond oil and "brush" the inside of your ears with it.
Practice daily for a few months and the tinnitus will fade away .

[roondar]

Originally Posted by Bruce Abbott

Unfortunately (or fortunately) many of us are getting on in years and our ears aren't what they used to be. I can barely tell the difference between the filter being on and off in my A1200, because my hearing is 30dB down above 3kHz and I have constant tinnitus which masks higher frequencies. The older I get, the better Paula sounds!

Don't worry, many of the worlds best professional audio engineers & audio mixers are in their mid 50's so even bad hearing doesn't mean much

[nikosidis]

Bruce Abbott: Very well written. I'm sure distortion and noise to a degree can be a good thing, in many cases it will give warmer, fuller sound. At least if a recording is not very good and on the analytical side. Most pop, rock recordings sound best when the signal is not too clean. Hi-Fi equipment. setup that is to analytical is for most to cold and at loud levels are not nice on the ear, unless the recording is first class.
I think Paula sound warm and nice but can also be to dull on some models. My Amiga 500 rev. 3 sound very dull.

[grond]

Originally Posted by malko

. A grandmother's hint : Rub a cotton bud with sweet almond oil and "brush" the inside of your ears with it.
Practice daily for a few months and the tinnitus will fade away .

There is nothing funny about tinnitus. I also have it and it got quite a bit worse with my covid infection.

[Thomas Richter]

Originally Posted by Bruce Abbott

By definition, a DAC that uses 14 bits to produce its output is 14 bit.

Unfortunately, that is not the situation we have. We have an 8 bit DAC multiplied by a 6-bit DAC, added to another 8-bit DAC. This is how the output is produced, not by a 14-bit DAC. This type of setup *could* be equivalent to a 14 bit DAC if all DACs are calibrated well enough such that the voltage levels of the multiplied 8-bit DAC fit into the staircases of the full-scale 8-bit DAC, but it isn't.


In the end, you get a non-linear transfer curve with staircases not fitting together.



Or, to put it in different terms: An ideal DAC is equivalent to applying a mid-level uniform equi-quantizer to the signal (which is the ideal quantizer in the high-bitrate regime). However, what the Amiga produces is a non-uniform quantizer whose quanitzation levels are so distorted that you miss 30dB, which is equivalent to missing 5 bits.


Said quite simply: The staircases do not fit toghether at all.


The whole setup cannot work quite right since the calibration precision of all DACs involved is way above the limits required for 14 bit quality.


That is not surprising, of course. It was never designed or manufactured to such precise quality.

[ross]

Originally Posted by Thomas Richter

We have an 8 bit DAC multiplied by a 6-bit DAC, added to another 8-bit DAC. This is how the output is produced

Thomas, I understand that you want to simplify the matter, but it is not that by continuing to say this, it becomes true.
Only bypassing the PWM stage, ie at volume 64, you have an output signal comparable to a normal PCM converter DAC. For all other volume values it is never true!

So the calibration or signal management to have output levels with a dynamics comparable to x-bits is not trivial.

EDIT:
Unless you mean what Paula's PWM provides is a form of 'standard' DAC.
Instead I see it as a form of manipulation and change on the analog/time domain (coupled to the filter), from the original PCM signal, that give you a crude form of volume change.

So saying that you are adding up, is like saying that you are adding oranges to apples... to have a basket full of apples at the end.
And in fact by making a simple sum you certainly can't have 14 bit

[roondar]

Originally Posted by Thomas Richter

2) The SNR on the Amiga is approximately 30dB (!) worse than that of the PC. In particular, the PC quality is approximately 64dB, that of the Amiga "14 bit sound" approximately 33dB.

However, what the Amiga produces is a non-uniform quantizer whose quanitzation levels are so distorted that you miss 30dB, which is equivalent to missing 5 bits.

The SNR rating you've found is the only thing about your measurements I do find a bit odd. Several others (including some of the links in this very thread) have measured Paula SNR to be much better than 33dB. The paper I linked measured 42dB for 8 bit playback and buggs measured close to 46dB for that. He also measured close to 60dB for "14 bit calibrated" playback.


Do you have an idea why your figures are so much lower?

[nikosidis]

Maybe it is Thomas A2000 that is that bad. I guess my A500 rev. 3 board is that bad too.
A1200 is nothing near that bad SNR. At least around 60db I would say is correct for A1200, EaglePlayer, 14bit amp.

[Thomas Richter]

Repeated the experiment with EaglePlayer. The result is pretty much the same, it is 31.80 dB, whereas the previous play16 output was 31.87dB, which is identical within the expected tolerance. Video mode was productivity.

Also, despite some people claiming otherwise, there were dropouts during the playback. Of course, my recording was made without such drop-outs.