MQA Deep Dive - I published music on tidal to test MQA

[DimitryZ]

Here is a combination of posts I wrote attempting to explain the functionality of MQA to my non-technical friends at PFM:

Let's say you want to grow tulips in your yard. They need the sun, but only one diagonal corner of your yard (triangular shaped) gets sunny, and the other half is all shade.

MQA gardener seeds only the sunny side of the garden with bulbs and gets 100 tulips. LPCM gardener seeds the entire garden, but the tulips only grow on the sunny side, so he gets 100 tulips. So MQA is lossless for tulips (music), but LPCM gardner also has lots of unrisen bulbs (zeros).

Does this make sense? This is the very basis of MQA functionality, their core process. They don't plant (decode) where tulips (music) don't grow (exist).

 

[dc655321]

No welcome?

Link to the plot of MQA/LPCM comparison added to above. I assumed you folks were up on this stuff.

Other stipulations are in my post. Analysis is simple dB math for amplitudes, which I assume all can do:

dB=20*log(A1/A2)

"Hollow post?" It's one of very few with actual numbers - audioSCIENCEreview?

Welcome.
Your link doesn't work.

 

[DimitryZ]

Welcome.
Your link doesn't work.

Thank you!

Permissions updated. I am not a savvy Google person.

 

[dc655321]

From graphs posted at PFM and elsewhere, "eyeball" MQA null against LPCM looks like -120dB, or 1 in a million (1E-6). This is clearly better than an excellent lossy codec but clearly not as good as mathematically lossless one.
https://drive.google.com/file/d/1o79kijUug1Rg3Ne0aHM35oibMMPbWRK6/view?usp=drivesdk
(Red and Blue curves used for conservatively deriving 120dB number)

The plot at the link looks suspiciously like something from @Archimago's site. If so you may want to acknowledge other's work

Also, can you explain how you get to a 120dB with your "analysis"?

 

[DimitryZ]

The plot at the link looks suspiciously like something from @Archimago's site. If so you may want to acknowledge other's work

Also, can you explain how you get to a 120dB with your "analysis"?

This was posted prior without attribution.

When I blow up the plot a lot it looks like there are a few locations where I can see some red to blue separations. When I look at the worst locations, and looking at the Y-axis scale, the differences look to be really small, a fraction of a dB. The "unrendered" comparison is even closer. I chose 0.2dB as representative. Taking the relatively low signal frequency of ~5000Hz, the level is ~ -24dB. 24/0.2=120. At higher frequencies the delta looks the same, but the dB difference will be much smaller - minus 140, 160, etc.

If Archimago took this data, he could have easily computed the A-weighted nulls between relevant curves at meaningful frequency ranges, like 20Hz-35KHz. Has he?

 

[amirm]

1) Modern ‘general purpose’ compressors are extremely efficient with low-information-density cyclostationary data. An 88kHz FLAC carrying 44kHz ‘plus one bit’ of PCM-modulated waveform satisfies that nicely. Resulting in the 24/88 music FLACs being comparable in size to 24/44 or 24/48 MQAs. Which was pointed at and illustrated by actual file sizes a few times within this thread.

What? What is "plus one bit?" Did you just make a new format? How is that going to deal with extra spectrum?

Here are the file sizes for L2's sample files for the first track in their benchmark list:

The CD at 16/44.1 is at the bottom. Go to 24/88 kHz and file size jumps by a factor of 3.8X. Go to 24-bit/176 kHz where there can't possibly be more music and now file size balloons by a factor of 8.2X. The information density is anything but low since we see doubling of the file size between 88 khz and 176 kHz.

The reason is simple: entropy is quite high. What is in ultrasonics is not only noise, but high frequency one. This is very hard to compress using "dumb" mathematically lossless algorithms. They see nothing but constantly fluctuating numbers and can't compress them much. Worse yet, their code books are not optimized for this kind of thing so efficiency sinks yet again.

Now look at the MQA version which is delivered as 24-bit/44.1 kHz. It is 2.4 X the size of 16/44.1 but it delivers the equivalent of the 176 kHz version because it knows not to even bother to encode the difference between 88 and 176 kHz. It is even smaller than 24-bit/88 kHz because it figured out there is no need to encode all of that.

The MQA version by the way is also suffering from poor performance of Flac lossless encoder because it is stuffing its bits in the high frequency spectrum of the 44.1 kHz file and at 24-bit. Flac sees a lot more random data there at high frequency and can't compress them well. Without this inefficiency increase, MQA would have done even better.

And what do you mean the file sizes are "similar?" MQA is 45 megabytes and the smallest high res file at 88 kHz sampling is 72 megabytes. Those are not close or even remotely so. Since MQA decode upsamples, the "fair" comparison is actually to the 176 kHz one and there, we are talking about 45 megabytes against 156 megabytes.

Bottom line, MQA has knowledge of what it is encoding and uses that to get higher efficiency. Flac encoder doesn't so even when it is encoding very low level audio data, it still generates massive files. I just downloaded all of the above and the 176 kHz one took forever to download.

 

[amirm]

2) ‘Noticing the statistical aspects of music‘ while ignoring ‘hearing aspects of the listener’ - ie not being able to hear ultrasonics, and especially within 6dB above noise, is hardly ‘neat‘ in my playbook. [If i remember, you pointed this yourself some posts earlier.]

It could have trivially done that as well. But the market doesn't want that. The market wants the files as created in the "studio" as that sounds like it is better. What you are saying is why the hotel customer doesn't drink tap water instead of paying $5 for the Fuji bottle on the table. Even if they taste the same, some want the Fiji water thinking they are in a paradise for a minute or two. Don't change the value prop and then ask why they did this. They did it because there is market demand.

Also, they conducted controlled test showing that filtering of high-res audio 44.1 khz can have audible effects:

“The audibility of typical digital audio Filters in a high-fidelity playback system,” [peer reviewed] Convention Paper, Presented at the 137th AES Convention 2014

The dotted line was p < 0.05:

“The dotted line shows performance that is significantly different from chance at the p<0.05 level [5% probability of chance] calculated using the binomial distribution (56.25% correct comprising 160 trials combined across listeners for each condition).”

Not a home run as far as night and day but passes the standard used in research for audibility. This paper was published a year or two before MQA as a format was announced so it laid the foundation for why this format should exist.

 

[dc655321]

When I blow up the plot a lot it looks like there are a few locations where I can see some red to blue separations. When I look at the worst locations, and looking at the Y-axis scale, the differences look to be really small, a fraction of a dB. The "unrendered" comparison is even closer. I chose 0.2dB as representative. Taking the relatively low signal frequency of ~5000Hz, the level is ~ -24dB. 24/0.2=120. At higher frequencies the delta looks the same, but the dB difference will be much smaller.

Uh huh

 

[Raindog123]

Here is the ‘new one-bit format’:

3. MQA (or really MQL as they are calling it now). This type of compression is different in that it only codes music above the baseband 44.1/48 kHz based on information/noise level it has. If its max range is 6 dB above noise floor, then all it takes is 1 bit to represent it, not 24. To do this, an analysis of the content must be made in the encoder and decisions made about what is noise and what is content above or below noise.

Since MQA decode upsamples, the "fair" comparison is actually to the 176 kHz one and there, we are talking about 45 megabytes against 156 megabytes.

…and this simply does not deserve a response. Though I am thankful you picked 176kHz and not 352kHz for your example, it - ‘the “fair“ comparison’ - would be as far from truth but would look even more impressive to an average consumer.


It’s not fun anymore. And I am starting to think maybe it is the goal all alone. Which is an extremely sad realization, but even the most patient of us have finite amount of the benefit-of-the-doubt…

You won. MQA is the best thing since sliced bread.

 

[voodooless]

Also, they conducted controlled test showing that filtering of high-res audio 44.1 khz can have audible effects:

“The audibility of typical digital audio Filters in a high-fidelity playback system,” [peer reviewed] Convention Paper, Presented at the 137th AES Convention 2014

The dotted line was p < 0.05:

“The dotted line shows performance that is significantly different from chance at the p<0.05 level [5% probability of chance] calculated using the binomial distribution (56.25% correct comprising 160 trials combined across listeners for each condition).”

Not a home run as far as night and day but passes the standard used in research for audibility. This paper was published a year or two before MQA as a format was announced so it laid the foundation for why this format should exist.

I found quite some critical comments on this paper here on a lot of aspects of the paper. I can’t corroborate any since I don’t have access to it, but the comments look genuine enough to make me apprehensive to trust the conclusions.

 

[DimitryZ]

Uh huh

So this means you don't agree? Or you have a different interpretation of the plot?

Your logo says you are addicted to learning.

Well?

 

[UKPI]

They will be entirely indistinguishable from one another to the listener and are, therefore, identical. The big debate about MQA lossy/lossless status is entirely meaningless for consumer reproduction.

The entire business of MQA mainly hinges on hi-res audio being relevant. And hi-res audio is relevant among audiophiles not because of audibility. It is mainly about owning "the" master or at least something better than a CD. The problem is undermined trust. This current fiasco happened because of MQA Ltd playing with words ("lossless"), continuing to promote dubious claims ("time smearing", "filters that have tons of imaging but are better than brickwall", "authentication system that fails to check file integrity"), and showing hostile reactions when their claims are questioned.

By the way, why do MQA CDs and 16-bit MQA encoded streams exist in the first place?

 

[UKPI]

“The audibility of typical digital audio Filters in a high-fidelity playback system,” [peer reviewed] Convention Paper, Presented at the 137th AES Convention 2014

The dotted line was p < 0.05:

Isn't TPDF non-subtractive dither mainly used in audio?

 

[DimitryZ]

The entire business of MQA mainly hinges on hi-res audio being relevant. And hi-res audio is relevant among audiophiles not because of audibility. It is mainly about owning "the" master or at least something better than a CD. The problem is undermined trust. This current fiasco happened because of MQA Ltd playing with words ("lossless"), continuing to promote dubious claims ("time smearing", "filters that have tons of imaging but are better than brickwall", "authentication system that fails to check file integrity"), and hostile reactions when their claims are questioned.

By the way, why do MQA CDs and 16-bit MQA encoded streams exist in the first place?

I don't know how to answer that, other than I do think musically relevant near ultrasonics (up to maybe 40KHz) content does seem to provide better overall sound.

As to the MQA-CD, I actually bought a few of the Japanese imports, as I have a CD player (Oppo UDP-205) that can properly decode them. They do sound legit excellent, but are expensive.

MQA is backward compatible to CD and carries its' hires information hidden in the 16KHz + noise floor of the standard CD.

 

[dc655321]

So this means you don't agree?

It means I think you need to do some learning.

 

[DimitryZ]

Here is the ‘new one-bit format’:




…and this simply does not deserve a response. Though I am thankful you did not pick 352kHz, it would look even more impressive.


It’s not fun anymore. And I am starting to think maybe it is the goal all alone. Which is an extremely sad realization, but even the most patient of us have finite amount of benefit-of-the-doubt…

You won. MQA is the best thing since sliced bread.

I think that what anyone that actually listens to MQA is arguing is that it has a right to exist and can succeed or fail based on the market decisions.

This should not be a controversial position.

 

[amirm]

You won. MQA is the best thing since sliced bread.

It is not and never has been. Market for high-res audio is small and the subset that are convinced MQA provides value there is much smaller still. MQA actually copies some of the techniques in HDCD which provided backward compatibility just the same by creating a channel in baseband audio to encode 20 bits into 16.

What I hope has been "won" is a more clear understanding of the topic so we don't say things that can be effectively countered by someone like me, or officially as Bob Stuart did in his blog post. Best to not step into things and then have to dig out of a hole.

 

[amirm]

This should not be a controversial position.

Well said. Nothing kills objectivism more than emotion and emotion is what we have been dealing with on topic of MQA.

 

[DimitryZ]

It means I think you need to do some learning.

So that's a very snarky reply to a new member.

You asked for the data reference and I provided it, with an explanation of my methodology.

Would you mind pointing errors or discrepancies in my analysis? Perhaps you can offer an alternative/competing one? As with numbers, since this is ASR?

 

[amirm]

Isn't TPDF non-subtractive dither mainly used in audio?

Nobody knows what dither (if any) is used in any (16/44.1) content we consume. My analysis of high-res audio master files shows incredibly poor practices in creating said content. So no, we don't know if TPDF dither is used.