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

[KSTR]

Can the 'dither' (the 3 bits' in our cartoon) be added by the decoder _after_ the 'ultrasonic information' bits are stripped away? Or does it carry some information (shaping?) all the way from the encoder?

The latter.

The relevant graph IMHO is this one:

The red line -- threshold noise-spectral-density -- is the important one. It's the level of noise vs. frequency that is just below audible threshold, "the loudest signal we can not hear" as Bob Stuart words it.

The space below the line (in the MQA-encoded distribution file which is half the sample rate of the original and compatible when played unencoded) is the additional coding space available (with a scale factor, though) that must be made pseudo-random to be proper noise-shaped dither.
In this pseudo-randomness is encoded the "subchannel" info, which comprises of the folded content (to appear later after added into the 2x upsampling) and meta-data and control information for the unfolding.

That's at least what I think I have understood ;-)

 

[DimitryZ]

• When you write, "Correct" and fail to note that my being correct also means that your prior comment was incorrect, you are gaslighting.
• When you write that MQA is "justified" in downsampling 352.8k content to 88.2k, but fail to note that this fact means your repeated attempts to drive the conversation into a comparison of MQA specifically with 352.8k DXD are flawed and misleading, you engage in a method of discussion that is indistinguishable from trolling.
• And when you fail to note that MQA does not simply throw out some ultrasonics, but also compromises bit-depth by retaining other ultrasonics that are just as unnecessary as the ultrasonics it threw out, you are once again engaging in selective discourse that is indistinguishable from trolling.

I do appreciate your acknowledgment that the final upsampling is “kind of fakey.” Not quite how I’d characterize it, but close enough.

I added a thank you about a minute after initial posting.

The relevant question is "compared to what?" Since a CD resolution downsamples to 44, MQA has a starting advantage. As for using DXD, I stand behind this comparison, since it is the modern lossless master from which all consumer destributions originate CD, downsample LPCM and MQA. Also increasingly for sale "as is."

And one absolutely prefer a 88 downsampled LPCM and consider it superior to 88 downsampled MQA. However, in @Archimago blind tests early on, SQ preference were split.

I don't actually understand your last paragraph, except the part where you attack me unnecessarily.

 

[AdamG]

Behave as adults or leave the Thread. Please and thank you .

 

[q3cpma]

The latter.

The relevant graph IMHO is this one:
View attachment 133424
The red line -- threshold noise-spectral-density -- is the important one. It's the level of noise vs. frequency that is just below audible threshold, "the loudest signal we can not hear" as Bob Stuart words it.

The space below the line (in the MQA-encoded distribution file which is half the sample rate of the original and compatible when played unencoded) is the additional coding space available (with a scale factor, though) that must be made pseudo-random to be proper noise-shaped dither.
In this pseudo-randomness is encoded the "subchannel" info, which comprises of the folded content (to appear later after added into the 2x upsampling) and meta-data and control information for the unfolding.

That's at least what I think I have understood ;-)

Best graph I've seen about this. On a purely technical level, it is then a quite ingenious solution to a problem that doesn't really exist.

 

[Hayabusa]

The intent was quite obvious. The poster asked me why I do one type of testing for hardware and am advocating another for MQA. In other words, he was challenging me to see if I am consistent. So I showed how the OP could care less about measurements yet all of a sudden has gotten religion when it comes to MQA. To the extent his reviews of hardware is all about random listening tests, then that is what he should have done with MQA. If he is not, he is the one that is hugely inconsistent.

As to rest of your post, you are being purely argumentative with no value added to the thread. One more like them and you will get a Reply Ban. Think hard before hitting post.

Really?

 

[KSTR]

Best graph I've seen about this. On a purely technical level, it is then a quite ingenious solution to a problem that doesn't really exist.

Yep. FLAC is good enough.
Maybe using a Gaussian "Apodizer" or other filter above 20k to bring the out-of-band noise down and to retain good impulse response, applied to 88.2k/96k 24bit files (or 176.4k/192k). Then packed in FLAC, which probably would compress better than with full-spectrum source (net benefit might still be questionable). Done...

 

[Phos]

Once again I have to ponder why this encoded information is being stored in the audible band like it’s the 1970’s. You can put more than two tracks in a digital file!

 

[raistlin65]

I am sorry, that's a tough place to be.

You can go over the @Archimago assessment and maybe adopt his language? Works and sometimes sounds better than FLAC to some listeners, but advertising is over the top...

That's a good text.

Does help, but doesn't solve the problem. The genre convention for Reddit tends to be short explanations, which get in the way of even summarizing more involved arguments like that. So still makes it difficult.

And for every person seeking help on Reddit (or even here at ASR), there's gotta be another 10 to 100 who just go with MQA because of that misinformation. Or go with it because a friend, who fell for the advertising, told them it's great.

They might be choosing Tidal because of the misleading MQA marketing which makes them think it has the best SQ, where they actually like the catalog or user experience of Amazon HD or Quboz better. Or maybe they bought a Dragonfly to get MQA when they could have saved money or have gotten a better DAC for their money.

Consequently, what's needed is more people speaking up about the misleading advertising prose to help counter that misinformation, as well as to counter those who want to silence criticism of MQA.

 

[Raindog123]

The red line -- threshold noise-spectral-density -- is the important one. It's the level of noise vs. frequency that is just below audible threshold, "the loudest signal we can not hear" as Bob Stuart words it.

The space below the line (in the MQA-encoded distribution file which is half the sample rate of the original and compatible when played unencoded) is the additional coding space available (with a scale factor, though) that must be made pseudo-random to be proper noise-shaped dither.
In this pseudo-randomness is encoded the "subchannel" info, which comprises of the folded content (to appear later after added into the 2x upsampling) and meta-data and control information for the unfolding.

That's at least what I think I have understood ;-)

Thanks! … yet still more questions than answers. Will do some digging tonight “after come back from work” So, if anyone still has any additional pointes/words of wisdom - do not be shy!


…and I keep forgetting this “backward compatibility with legacy CD players” point. While some might argue it for a virtue, in my book it’s a considerable inhibitor!

 

[tmtomh]

I added a thank you about a minute after initial posting.

Thank you - I appreciate that.

The relevant question is "compared to what?" Since a CD resolution downsamples to 44, MQA has a starting advantage.

This statement reflects a fundamental error in the application of digital sampling theory, and/or a series of unsupported (and I would say disproven) beliefs about the capabilities of modern reconstruction filters.

CD/redbook's 44.1k sampling frequency can encode frequencies up to and slightly beyond the range of human hearing. In terms of accurately sampling audible musical sounds, doubling the sample rate to 88.2k gets you nothing. If sampling any frequency more than the 2 times necessary for proper reconstruction provided an increase in audible "resolution" (as the term is often misused in audiophile circles), then we'd all be in a lot of trouble: No matter the sample rate you use, a bass drum at 60Hz is sampled 30-85 times more than a musical sound in the 2-5kHz range where human hearing is most sensitive, and therefore by the "more sampling = higher resolution" fallacy, lower frequencies of music would always be more refined and hi-fi than higher ones (which is of course not true). Once we reach the minimum 2 samples per frequency, the number of additional times we sample is irrelevant - and it's a good thing because different frequencies get oversampled by vastly different degrees, no matter the sample rate.

Now, if your point is that 88.2kHz is better because it allows for a gentler-sloping reconstruction filter between 20kHz and Nyquist than CD/redbook does, the problem there is that one can find plenty of data all over these forums, primarily in @amirm 's reviews of DACs, that illustrate the ability of relatively steep filters to function properly in the 2.05kHz space that redbook provides. Conversely, he's shown measurements of many pieces of equipment that use filters that have such poor out-of-band attenuation that even an 88.2k sample rate won't save you - and in fact it can be worse because there could be more ultrasonics in the file, which leads to more aliasing into the audible band. So yes, in an ideal world we'd have more consistency in filter quality, and in an ideal world the physical size and scan rate of the video tape format Sony first used for PCM encoding would've allowed for a 48 or 50kHz sample rate instead of 44.1k. And on filters it's worth noting that Bob Stuart is a major reason so many slow filters and "apodizing" filters are out there to begin with, as he's been evangelizing them for decades based on provably incorrect claims and assumptions.

Bottom line, MQA's sample rate is not a "starting advantage" over redbook.

As for using DXD, I stand behind this comparison, since it is the modern lossless master from which all consumer destributions originate CD, downsample LPCM and MQA. Also increasingly for sale "as is."

You stand by the DXD comparison - except when it doesn't suit you, as for example when you instead compare MQA with CD/redbook, as you did earlier in the very same comment.

Your claim that DXD "is the modern lossless master from which all consumer distributions originate CD" (italics mine), is false. Putting that aside,
it has already been explained, quite clearly and in detail, why the fact that a CD is downsampled from a higher-res digital master has nothing whatsoever to do with the question of whether it makes sense to compare MQA with DXD. The responsibility for not acknowledging the relevant points or taking them into account lies with you.

I don't actually understand your last paragraph, except the part where you attack me unnecessarily.

By my last paragraph, do you mean the final bullet point where I point out that you praise MQA's discarding of some ultrasonics while not acknowledging that in order to keep the remaining ultrasonics it has to compromise bit-depth? I'm not sure what to say in terms of understanding, since I think my point there is fairly clear. Sorry if it's not. I had no specific intention to insult you; rather, I tried to be as specific as I could, just as I did in the other two bullet points, about exactly what aspect of your prior comment I took issue with, precisely why, and how it landed to me as a reader.

If instead you're referring to the final bit of my comment, about your acknowledgement of MQA oversampling as "kind of fakey," I was agreeing with you there and had no critical intent.

More generally, I will modulate my language whenever necessary to conform to forum rules and mod requests. Beyond that, though, when it comes to critical comments about your mode of argument here, I'm afraid you've made your own bed.

 

[Raindog123]

Once again I have to ponder why this encoded information is being stored in the audible band like it’s the 1970’s. You can put more than two tracks in a digital file!

They run with this “backward compatibility with Redbook playback equipment” flag.

 

[TK750]

They run with this “backward compatibility with Redbook playback equipment” flag.

My technical understanding of this process is extremely limited so I'm probably saying something stupid here but wouldn't that also decrease the reduction in file size which seems to be one of the main benefits they give?

 

[tmtomh]

Once again I have to ponder why this encoded information is being stored in the audible band like it’s the 1970’s. You can put more than two tracks in a digital file!

Yes - they could have done the same lossy compression on the same portion of the ultrasonics and just not stored it (and the metadata/flags etc) in the audible band. It could have been stored in the PCM file header (as most PCM formats allow pretty much anything to be stored in the header), where it would be ignored by non-MQA-capable equipment, and recognized and used by MQA-capable equipment.

My technical understanding of this process is extremely limited so I'm probably saying something stupid here but wouldn't that also decrease the reduction in file size which seems to be one of the main benefits they give?

Yes - the problem with the above scenario is that it would have reduced the file-size savings, and while it would have worked with digital redbook files, it would not have worked with CDs, as it would have been really cumbersome to figure out where to put the extra data on the disc, and with any CD that had a very long program of music on it, there wouldn't be enough room on the disc to write all the extra data.

So the solution of storing it in the audible band is a more elegant and compatible solution - the issue, of course, is that it's a solution to a problem that doesn't exist. Or to be more generous to MQA, it's a cure that's worse than the disease.

 

[Raindog123]

My technical understanding of this process is extremely limited so I'm probably saying something stupid here but wouldn't that also decrease the reduction in file size which seems to be one of the main benefits they give?

Yep, That’s what people also pointed out, even recently. One can have “FLAC’ed” 20/96, or 17/96, or even 17/48… without the need to carry the pseudo-randomized dither+info from the source, all these will be pretty compact, the last one incredibly compact… while still being incredible “music to your ears”

 

[tmtomh]

Yep, That’s what people also pointed out, even recently. One can have “FLAC’ed” 20/96, or 17/96, or even 17/48… without the need to carry the pseudo-randomized dither+info from the source, all these will be pretty compact, the last one incredibly compact… while still sounding as “music to ones ears”

Yes! Or alternatively, as many folks have already noted, just go with 24/48 FLAC and be done with it once and for all.

 

[oursmagenta]

Yes - they could have done the same lossy compression on the same portion of the ultrasonics and just not stored it (and the metadata/flags etc) in the audible band. It could have been been stored in the PCM file header (as most PCM formats allow pretty much anything to be stored in the header), where it would be ignored by non-MQA-capable equipment, and recognized and used by MQA-capable equipment.



Yes - the problem with the above scenario is that it would have reduced the file-size savings, and while it would have worked with digital redbook files, it would not have worked with CDs, as it would have been really cumbersome to figure out where to put the extra data on the disc, and with any CD that had a very long program of music on it, there wouldn't be enough room on the disc to write all the extra data.

So the solution of storing it in the audible band is a more elegant and compatible solution - the issue, of course, is that it's a solution to a problem that doesn't exist. Or to be more generous to MQA, it's a cure that's worse than the disease.

Out of curiosity.

Do they simply erase the lower bits entirely? Or they add the mqa information packed as as pseudo noise to the lower bits? This last one seems pretty far fetch, especially regarding how to recover this packed data among the noise at decode time, but I prefer to ask anyway, we never know maybe this codec is very clever.
Or is it something else?

P. S: no need to answer with a lenghtly post, a link to some data will suffice.

 

[pjug]

The latter.

The relevant graph IMHO is this one:
View attachment 133424
The red line -- threshold noise-spectral-density -- is the important one. It's the level of noise vs. frequency that is just below audible threshold, "the loudest signal we can not hear" as Bob Stuart words it.

The space below the line (in the MQA-encoded distribution file which is half the sample rate of the original and compatible when played unencoded) is the additional coding space available (with a scale factor, though) that must be made pseudo-random to be proper noise-shaped dither.
In this pseudo-randomness is encoded the "subchannel" info, which comprises of the folded content (to appear later after added into the 2x upsampling) and meta-data and control information for the unfolding.

That's at least what I think I have understood ;-)

If the red line is the audible threshold, shouldn't 20KHz or thereabouts be an asymptote that isn't crossed (or at the least, not so easily)? I think they must be meaning something more than that.
Edit: I found the curve or similar in a post by Amir in another thread: https://www.audiosciencereview.com/...-mqa-empirical-measurements.22969/post-779255
Apparently the reason red curve crosses into ultrasonic is to make room for noise shaping, if I am relaying what Amir said correctly.

 

[TK750]

Yes - they could have done the same lossy compression on the same portion of the ultrasonics and just not stored it (and the metadata/flags etc) in the audible band. It could have been stored in the PCM file header (as most PCM formats allow pretty much anything to be stored in the header), where it would be ignored by non-MQA-capable equipment, and recognized and used by MQA-capable equipment.



Yes - the problem with the above scenario is that it would have reduced the file-size savings, and while it would have worked with digital redbook files, it would not have worked with CDs, as it would have been really cumbersome to figure out where to put the extra data on the disc, and with any CD that had a very long program of music on it, there wouldn't be enough room on the disc to write all the extra data.

So the solution of storing it in the audible band is a more elegant and compatible solution - the issue, of course, is that it's a solution to a problem that doesn't exist. Or to be more generous to MQA, it's a cure that's worse than the disease.

Yep, That’s what people also pointed out, even recently. One can have “FLAC’ed” 20/96, or 17/96, or even 17/48… without the need to carry the pseudo-randomized dither+info from the source, all these will be pretty compact, the last one incredibly compact… while still be incredible “music to your ears”

Got it, thanks for the replies

 

[Phos]

The problem I have with the CD explanation is that MQA was launched in 2014 and the format has exclusively been used for streaming.

 

[tmtomh]

Out of curiosity.

Do They simply erase the lower bits entirely? Or they add the mqa information packed as as pseudo noise to the lower bits? This last one seems pretty far fetch, especially regarding how to recover this packed data among the noise at decode time, but I prefer to ask anyway, we never know maybe this codec is very clever.
Or is it something else?
P. S: no need to answer with a lenghtly post, a link to some data will suffice.

I'm the wrong person to ask in terms of data, as I don't have any specific MQA data at hand that would show what you're asking for.

I can provide a brief answer, though: My understanding is that an MQA file in a 24-bit container will use both methods: it will take over several of the lowest bits to store the encoded ultrasonics; and it will mix certain encoding/decoding info into pseudo-noise in the lower 3 bits of the first 16 bits (in other words bits 14, 15, and 16). When people write that even 24-bit MQA files only have an effective bit depth of about 17 bits, this two-level method of putting stuff into the original is what they are talking about - they are saying that these two methods combined end up sacrificing about 7 bits' worth of depth in a 24-bit file.

@mansr could explain it more completely and definitively were he able to post in this thread. Hopefully someone else with more technical expertise than I have - and there are lots of you here! - will jump in.