Amiga "14-bit audio" refuted

[nikosidis]

Here you have a nice one from Edison 2018 with a lot of bass and quiet passages. No compression.

Original stream.
https://files.scene.org/view/parties...lten_dune.flac

Paula A1200.
https://files.fm/u/uyak32ucy#/view/sucdmh6ae

I do not know why, but it sounds better when I playback the 2 Paula channels I recorded from Audacity than the file I saved to wav and uploaded. It is more clear, detalied. Maybe something happen when the 2 channels are mixed or is it simply the streaming?

[chb]

Originally Posted by daxb

I don't know how 14 bit sounds or should sound and I guess nearly nobody knows it. Do exits 14 bit playback devices? That might be interesting to compare Amiga 14 bit playback with a 14 bit reference device. Comparing with a bad 16 bit sound card doesn't make sense to me.

Easiest way: Put a 16 bit sample in a sample editor, reduce the volume in the editor to 25% (so it uses only a 14 bit range), and play it back, restoring the volume level using your amplifier or your desktop volume controls (which probably does it in the digital domain).

[nikosidis]

Originally Posted by daxb

Don't forget that 14 bit playback needs 16 bit source to make sense. On Amiga everything is 8 bit. So you need to copy uncompressed 16 bit 44kHz audio material to Amiga first. I don't see much use of it.

All the 8 bit sound on Amiga (8SVX, RAW, ... samples) used for music and soundfx will be played with native Paula. Sounds good enough and it is recommended.

I don't know how 14 bit sounds or should sound and I guess nearly nobody knows it. Do exits 14 bit playback devices? That might be interesting to compare Amiga 14 bit playback with a 14 bit reference device. Comparing with a bad 16 bit sound card doesn't make sense to me.

The point here is that Paula is more than capable of playing 16 bit sound. Most do not have good enough speakers or hi-fi system to justify better source. It is not very useful to most but to me it is. I have my Amiga 1200 in my bedrom and have lot's of music on it that I converted from FLAC to AIFF or WAV. I have some decent speakers connected to it. I enjoy the sound from it

[no9]

Originally Posted by daxb

Don't forget that 14 bit playback needs 16 bit source to make sense. On Amiga everything is 8 bit. So you need to copy uncompressed 16 bit 44kHz audio material to Amiga first. I don't see much use of it.

All the 8 bit sound on Amiga (8SVX, RAW, ... samples) used for music and soundfx will be played with native Paula. Sounds good enough and it is recommended.

Recommended for what? I wouldn't recommend using DigiBoosterPro in 8 bit mode to anybody if there is a hypothetical 14-bit available for Paula. It was quite common opinion amongst Amiga musicians that those '14-bit HIFI' AHI modes not necessarily sound so close to 16-bit as raw mathematical calculation would suggest, but it was still improvement over 8 bit. And Thomas measurements seems to confirm that notion.

I don't know how 14 bit sounds or should sound and I guess nearly nobody knows it. Do exits 14 bit playback devices?

There are 12-bit samplers. And there is software where you can reduce sound parameters according to your wish. Like that
[ Show youtube player ]

[alexh]

I think you're all going off topic slightly, in that personal opinion is coming into play rather than technical / empirical proofs.

If I read the thread correctly Thomas proved empirically that using his set-up it wasn't possible to replicate 14-bit accuracy at 44KHz?

[roondar]

Originally Posted by alexh

I think you're all going off topic slightly, in that personal opinion is coming into play rather than technical / empirical proofs.

If I read the thread correctly Thomas proved empirically that using his set-up it wasn't possible to replicate 14-bit accuracy at 44KHz?

Yup. That's about the gist of it.

Thinking about it some more, there are some possible issues (though he may have already taken these into account and merely not noted this in his OP) with this test.

The two main potential issues I can see is that AFAIK the Amiga's audio output quality is somewhat dependent on the age of the components and on what model of Amiga you actually use. I suppose the number of tested Amiga's can also play a role here. Suppose Thomas happened to have tested one that just had very poor output compared to the average (or conversely tested one that had very good output compared to the average). That would impact the results.

It's fairly well established that Amiga audio output quality can suffer if components on the board are old or on the verge of failure, especially with the capacitors. Similar audio quality issues can arise from a poor power supply.

It's also known that the audio output quality of the AGA Amiga's was significantly better (due to a slightly redesigned filter) than that of the OCS/ECS models.

There may be some other factors as well that could influence the outcome, but they're probably less important.

Note, I'm not trying to suggest that Thomas did a poor job or that his test results are wrong (in fact, I don't see them as unreasonable results at all). Merely that it's actually not that easy to test these kind of things correctly and that issues can arise. And for all I know he already took all of the above into account

[nikosidis]

I guess Thomas started the thread to make a discussion. That is what a forum is about. Sure he try to prove something but it might be difficult if not tested on different hardware. Roondar make many valuable points. I have recaped my A1200 and bought a new PSU. The A500 I have is very old and the sound quality is terrible.
From what I heard also the A600 should have the new audio filter as A1200 and possible A4000.

From my experience as a listener I think Paula audio does CD quality audio very good. There are some noise in quiet passages but does not annoy me. Like some other talked about there, vinyl is not exactly perfect either but people like it. Well, that is including myself I bought back a REGA turntable some years back.

[ross]

http://eab.abime.net/showpost.php?p=...0&postcount=44

In my opinion Paula sounds like a 12+ bits

That is great


EDIT: and about vinyl, yes for me also sound 'better' than CD, it is more colorful and full-bodied (and I'm not talking about the 'physical' or 'measurable' qualities of sound at all)

"To determine the effective bit rate necessary to fully capture the information on a vinyl LP, you need to know the bandwidth and the dynamic range. A typical vinyl LP has a bandwidth of about 18kHz (when it's brand new, it might get up to about 22kHz). There is certainly some audio stuff above that, but it generally bears no relation to meaningful programme material (ie. it's noise and distortion). So let's be generous and assume a bandwidth of 22kHz: you'll need to sample this at 44kHz. The dynamic range of a beautifully pressed LP on virgin vinyl can get to about 65-70dB on a good day with a following wind, which equates to slightly less than 12 bits. So the bit rate required is 44,000 x 12 x 2 (for stereo), giving about 1030kbs. A more typical LP (18kHz bandwidth, dynamic range of 55dB) needs a bit rate of about 650kbs. For comparison purposes, the CD bit rate is 1378kbs." [cliveb\hydrogenaudio]

So for me Amiga 14 bit vs 16 bit is like vinyl vs CD quality

[grond]

Originally Posted by Thomas Richter

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

Henryk has a PhD in the field of signal processing. Even though you are only a physicist, you should know the concept of preemphasis (considering your involvement with MPEG et al). If you didn't measure a case where you used preemphasis, you didn't debunk his approach.

[chb]

Originally Posted by grond

If you didn't measure a case where you used preemphasis, you didn't debunk his approach.

I do not think that pre-emphasis would have any effect on a 441 Hz signal, as it is well below the LPF frequency. But his improved calibration may.

[nikosidis]

I made a little video where I play some 16-bit audio files and show EaglePlayer.
It is important to use the 14-bit amplifier and Paula direct.
[ Show youtube player ]

[Thomas Richter]

It's not really a filter issue. There are multiple issues, as I have pointed out.

Point 1) is that it is unclear how precise the 8-bit DAC in Paula actually is, and how precise the volume voltage control is relative to the DAC. If the step sizes of the voltage DAC does not fit well to the stepsizes of the wave DAC, things do not fit together. I doubt Paula was calibrated carefully enough to ensure a consistent quality.

Point 2) is that the overall separation of the audio circuit to the rest of the system might be not ideal such that noise leaks from the digital part to the analog part. You have the same problem with cheap PC systems.

Point 3) is that the frequency control is not fine enough. That is, if you attempt to play 441Hz, you get something different. Currently, the test program tries to compensate for that by adjusting the phase, but it is also a distortion by itself.

In the end, I can only advocate to test things carefully. My experiment shows at least that my A2000 is *not* capable to reach 16 bit CD audio quality. Not even close. 30dB difference is a lot.

You may get closer with "slightly more modern hardware", but I have my doubts that you get above 40dB. For "true 14 bit output", with my setup, you would even need to get into the 50dB area. Anyhow, you can prove me wrong. I provided my sources.

[chb]

Originally Posted by Thomas Richter

Point 1) is that it is unclear how precise the 8-bit DAC in Paula actually is, and how precise the volume voltage control is relative to the DAC. If the step sizes of the voltage DAC does not fit well to the stepsizes of the wave DAC, things do not fit together. I doubt Paula was calibrated carefully enough to ensure a consistent quality.

There's most likely no voltage DAC for the volume, but PWM volume control - there shouldn't be linearity issues with that.

BTW, interesting experiment.

EDIT:

Originally Posted by Thomas Richter

Point 3) is that the frequency control is not fine enough. That is, if you attempt to play 441Hz, you get something different. Currently, the test program tries to compensate for that by adjusting the phase, but it is also a distortion by itself.

I think that is a debatable approach in your experiment. While the limited frequency control can indeed be an issue for many applications, it should not have an effect on a measurement of the dynamic range, as the reference sample rate can be considered arbitrary. For pure stand-alone playback there's no significant effect on sound quality if the frequency is off by a couple of cents. So IMHO the distortion introduced by your phase compensating algorithm does not represent an audible distortion.

If the Amiga audio is used together with other audio sources or synced to video, the frequency inaccuracies can pose a serious problem of course.

[grond]

Originally Posted by chb

I do not think that pre-emphasis would have any effect on a 441 Hz signal, as it is well below the LPF frequency. But his improved calibration may.

Yes, you are right, the preemphasis is only needed to get the output bandwidth up to what a CD can do. For a relatively low output frequency it should be more important to calibrate all available steps carefully and reorder them from their numerical order to their real physical output value. Aging components and power supplies will also impact the Amiga's output signal. The readme for eagleplayer quoted above also mentions noise shaping.

My main point is that it all depends very much on what signal preprocessing you do before you output the 14 bit audio. My impression is that Thomas didn't do any preprocessing for his measurements. Wer misst, misst Mist.

[roondar]

So I decided to Google "Amiga signal to noise ratio" for fun to see if anyone else had done some measurements so we had something to compare these results to. And as it turns out, some people at the university of Dalhousie (Canada) did indeed do so back in 1989 to see if the Amiga could be used for frequency systhesis. Edit: note that they did not try "14 bit mode", all their results are for straight 8-bit samples.

The things one can find with Google never cease to amaze me, I did not expect this gem

I'll provide the link to the full PDF here: https://www.researchgate.net/publica...emory_of_pitch
There are more links, but Google didn't want me copying those properly for some reason.

Perhaps the most relevant part is that they found a 42DB SNR using their measurements, but there's all sorts of goodies about the audio quality and limitation of the original OCS Amiga in there. Perhaps this shows the difference of using the machines while still "new" and using them today?

[Thomas Richter]

Originally Posted by chb

There's most likely no voltage DAC for the volume, but PWM volume control - there shouldn't be linearity issues with that.

BTW, interesting experiment.

I'm not so much concerned on the linearity, but on whether the levels of the two DACs working together fit. This mode more or less assumes that the steps of a lower-volume 2nd channel fit into the coarse steps of a full-volume channel, and this does not necessarily hold. In fact, this is why we need this "calibration" at all.

[ross]

Originally Posted by Thomas Richter

I'm not so much concerned on the linearity, but on whether the levels of the two DACs working together fit. This mode more or less assumes that the steps of a lower-volume 2nd channel fit into the coarse steps of a full-volume channel, and this does not necessarily hold. In fact, this is why we need this "calibration" at all.

And I call this linearity (for the the two DAC working as a quasi-one)

I would be a little more concerned about the channels phase (this might color the sound, that's not necessarily bad)

[Thomas Richter]

Originally Posted by grond

My main point is that it all depends very much on what signal preprocessing you do before you output the 14 bit audio. My impression is that Thomas didn't do any preprocessing for his measurements. Wer misst, misst Mist.

The test input is just a sine wave. The only effect a preemphasis (in frequency space) has is a change of the signal amplitude. However, the signal amplitude is part of the fitting parameters of the measurement script, so it doesn't matter. If you mean a "non-linear preemhasis", then you would also need a fitting non-linear postemphasis to get your sine wave back, but your average amplifier doesn't have that. It's all as close to linear as the manufacturer can do it, so that doesn't help in the application you use your Amiga for.

Note well, I'm doing signal processing all the day, albeit typically on images, not on audio.

[chb]

Originally Posted by Thomas Richter

I'm not so much concerned on the linearity, but on whether the levels of the two DACs working together fit. This mode more or less assumes that the steps of a lower-volume 2nd channel fit into the coarse steps of a full-volume channel, and this does not necessarily hold. In fact, this is why we need this "calibration" at all.

There's a very comprehensive document by Henryk Richter on Amiga audio and also on calibration (section 4.2). He also discusses Christian Buchner's approach on calibration and why he considers it to be non-optimal.

Originally Posted by Henryk Richter

It came as quite a surprise to me that the 16 Bit to 14 Bit conversion table example code produces neither type of quantizer. Zero is a valid output but any 16 Bit value <0 directly maps to the aforementioned maximum negative number in 14 Bit. This amplified the 16 Bit to 14 Bit quantization error amplitude by a factor of 4 (?12dB).

It may be interesting therefore to repeat the experiment with EaglePlayer (if you're interested, ofc).

[bloodline]

Originally Posted by Thomas Richter

Conclusion: If the PC precision is 16bit, and Amiga is 5 bit worse, then the Amiga "14 bit audio" is approximately worth 11bits. Not 14. That corresponds approximately to my estimate. It is better than the 8 bit output, but not close to 14 bit.

Excellent experiment! And I appreciate the time you have taken to test this.

I think you are being a little generous suggesting that the output is approximately to 11bit, I would suggest it is closer to 10bit.