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

[restorer-john]

I do know better and have explained why it is totally inappropriate to test MQA with test tones. Codecs are designed with certain assumptions. You can't violate them. A voice codec for example is created to encode voice. You can try to encode music with it and complain that it doesn't sound good but folks that know better would just laugh at you.

The *science* of lossy codecs says you use music and controlled listening tests to evaluate them. It also says learn what the technology does and then test it the way it should. A music codec is a music codec. It is not designed to spit out full range ultrasonics when such content doesn't exist in any music whatsoever.

The multitone test is used because it can be argued it is more representative of music than single spot frequency sines. How do you think an MQA encoder will respond to that? Throw more errors perhaps?

How about a spectrally shaped multitone?

 

[amirm]

Is MQA lossless to the CD quality?

I don't see the application of encoding CD in MQA. Is this what you are asking?

 

[amirm]

The multitone test is used because it can be argued it is more representative of music than single spot frequency sines. How do you think an MQA encoder will respond to that? Throw more errors perhaps?

If it goes above 20 kHz, pretty horribly. Within 20 kHz, it still violates statistical distribution of bits in music but it may not be too bad.

 

[JohnYang1997]

I don't see the application of encoding CD in MQA. Is this what you are asking?

What I was trying to ask was: Was MQA encoding process make the final audio output worse than CD quality?
And by asking this way, it seems we already made agreement that MQA is nowhere near lossless.

 

[KSTR]

A music codec is a music codec. It is not designed to spit out full range ultrasonics when such content doesn't exist in any music whatsoever.

That's the whole problem: Define what music is. Experimental "artsy" Music may be just anything, and that includes high level ultra LF and ultra HF content (HF meaning 10..20kHz here when released in CD format. During production it might have gone higher -- typical example is up-shifted samples of other music, sometimes several octaves). The music synthesizers built in the company I work for easily spit out sustained 20kHz++ at full scale (at 48k sample rate) and that happens all the time.

 

[amirm]

I don't really understand the argument that the test files were not suitable, or that encoder warning would be an issue. You have to keep the hypothesis of the video in mind: MQA is lossless.

SO he is testing a market claim. That he accomplished.

MQA is countering saying that if you feed the codec music instead of test tones, they do indeed preserve its fidelity using that triangular encoding method (as opposed to rectangular encoding of PCM). OP's test did nothing to test this. To the extent MQA is able to do that with music, then OP didn't accomplish anything other than getting a few lines changed in a Wiki/marketing pages. Content owners will still be able to be convinced that with real content, the process is "lossless."

 

[amirm]

What I was trying to ask was: Was MQA encoding process make the final audio output worse than CD quality?

Based on what they are explaining, no. It gives you wider bandwidth and more depth if you start with 24 bits/high res as opposed to straight encoding to 16/44.1.

The above by the way is a numerical answer. A perceptual answer would involve listening tests which I don't have. They don't have. No one else either.

 

[restorer-john]

Within 20 kHz, it still violates statistical distribution of bits in music but it may not be too bad.

Yes, that's why I added spectrally shaped. So with say a pink characteristic, but multiple discrete tones. That could be a somewhat fair way of determining what it adds or distorts?

 

[amirm]

That's the whole problem: Define what music is. Experimental "artsy" Music may be just anything, and that includes high level ultra LF and ultra HF content (HF meaning 10..20kHz here when released in CD format.

Post a spectrum and let's see. MQA says that they have accumulated data across very wide range of content and they have yet to find such content. I only know of one clip that has unusual characteristics and that is the "key jangling" clip.

Note that unusual content is still encodable but may not be optimal. Controls in the encoder for example could band limit, etc. If you feed random noise to a video codec it also barfs on it. You can choose to release it that way and rely on no one caring if there are artifacts in that random noise.

All lossy codecs can suffer from so called "codec killers." This doesn't stop them from being distributed and subjected to such content. Key is whether across vast amount of content does the codec perform. If it does, then it is ready for market. What it does on some crazy content here and there is not going to stop anyone from using it.

 

[badboygolf16v]

You also talked about MQA's wording. If that applied to MQA same applies to you. You can't have double standard. Be in the clear that's what we were talking about.

At no point has @GoldenOne stated as a fact that MQA removed it. He has reiterated repeatedly it is his opinion.

MQA are [proven liars|economical with truth|wonderful marketeers] call it what you want, it's the same result.

Healthy suspicion of bad faith actors is a good approach to take.

 

[voodooless]

MQA is countering saying that if you feed the codec music instead of test tones, they do indeed preserve its fidelity using that triangular encoding method (as opposed to rectangular encoding of PCM). OP's test did nothing to test this. To the extent MQA is able to do that with music, then OP didn't accomplish anything other than getting a few lines changed in a Wiki/marketing pages

They now even claim it's better than lossless... Whatever that means... I think this statement is even worse.. And you know perfectly well what lossless means in the audio world. You complain when others have other definitions of DRM.. now you're heading in the same direction.

Content owners will still be able to be convinced that with real content, the process is "lossless."

So you say that as long as one can convince the content owner that something is lossless, it is lossless? I'd love to hear more about that reasoning, please?

 

[JohnYang1997]

Based on what they are explaining, no. It gives you wider bandwidth and more depth if you start with 24 bits/high res as opposed to straight encoding to 16/44.1.

The above by the way is a numerical answer. A perceptual answer would involve listening tests which I don't have. They don't have. No one else either.

From my own non-disclosed testing, MQA attenuates high frequency over 20khz and introduces 16khz standing tone. The test was performed using actual music.
So imo maybe more bandwidth but I'm not sure about depth.

 

[amirm]

Yes, that's why I added spectrally shaped. So with say a pink characteristic, but multiple discrete tones. That could be a somewhat fair way of determining what it adds or distorts?

I don't know what you do with that data. That is why I put quotes on multitone emulating "music." It isn't music even if we tailor its spectrum as you suggest.

I think a proper set of codec killers can be developed for MQA which we can then perform listening tests on. We can search for real music that has higher high frequency spectrum for example and see what distortions get created, etc. Test tones are not going to get us there.

 

[restorer-john]

The music synthesizers built in the company I work for easily spit out sustained 20kHz++ at full scale (at 48k sample rate)

Tweeter killers when played live...

 

[JohnYang1997]

I don't know what you do with that data. That is why I put quotes on multitone emulating "music." It isn't music even if we tailor its spectrum as you suggest.

I think a proper set of codec killers can be developed for MQA which we can then perform listening tests on. We can search for real music that has higher high frequency spectrum for example and see what distortions get created, etc. Test tones are not going to get us there.

Well. The issue from here is how on earth did they define the word lossless? Is that even valid to begin with?

 

[restorer-john]

We can search for real music that has higher high frequency spectrum for example and see what distortions get created, etc.

I've always found any music with bar chimes somewhere in it, well recorded, is a perceptual encoder killer. Nothing seems to cope with them.

 

[KSTR]

Tweeter killers when played live...

Sound guys can handle this. And any PA worth the name has proper limiters, as do studio monitors.
I don't see many (hardly any, actually) reports of keyboarders frying tweeters or subwoofers.

 

[amirm]

They now even claim it's better than lossless... Whatever that means... I think this statement is even worse.. And you know perfectly well what lossless means in the audio world. You complain when others have other definitions of DRM.. now you're heading in the same direction.

Not at all. There is for example "perceptually lossless." This is when you tell a lossy encoder to produce the highest quality it knows, and have the bit rate vary. This is as special VBR mode that is available for some lossy codecs. We had this at Microsoft and in vast amount of content it was transparent and about half the size of mathematical lossless.

MQA proposes another form of lossless that says it music follows statistical distribution, they are able to fully preserve its actual informational content including what is in ultrasonic. If they are able to achieve this fully, what is your beef with it? That it doesn't spit out a bunch of bits for a rectangular channel is not used in music?

When we want to say the same bits come out that go in, we clarify with the term "mathematically lossless."

Remember, there is no lossless codec that works for all content. Here is the wiki on that: https://en.wikipedia.org/wiki/Lossless_compression

"By operation of the pigeonhole principle, no lossless compression algorithm can efficiently compress all possible data. For this reason, many different algorithms exist that are designed either with a specific type of input data in mind or with specific assumptions about what kinds of redundancy the uncompressed data are likely to contain."

The code books in lossless codecs for example are trained on a specific dataset which in this case would be music, not random test signals. Lossless codecs can actually make the output larger than input! When this happens, they cheat and just pass the input to output. MQA can't do that because it explodes the output so it complains with error messages.

 

[KSTR]

I've always found any music with bar chimes, well recorded, is a peceptual encoder killer. Nothing seems to cope with them.

Full Ack. Wind chimes is really hard stuff for anything.

 

[abdo123]

Well. The issue from here is how on earth did they define the word lossless? Is that even valid to begin with?

they could easily define it as perceptually lossless, which is the goal of any audio codec in the world that isn't called FLAC.