Invite: MQA vs Hi-Res blind test...

[Archimago]

Hey guys, just wanted to drop by to let folks know that I'm running a blind test of MQA Core Decode vs. original hi-res...

 

It's a way to compare what MQA decoding sounds like with your own DAC especially for those of you who don't have TIDAL and may be wondering about the difference between a stream that's MQA Core decoded compared to say a non-MQA 24/96 stream.

 

After you listen, I would of course love to know your perception of differences. The tracks are randomized and "blinded". Tell me which you prefer and how much difference you perceive!

 

Test instructions and files can be found here:

INTERNET BLIND TEST: MQA Core Decoding vs. Standard Hi-Res Audio

Thanks all, have fun...

Arch

 

[fas42]

Thanks for putting this up - second time I've compared an MQA versus "normal", also from 2L - and again the differences were very clear in two cases, less so in the third. There was a clearly superior version, for me, in each pair - that first time I compared, the MQA lost out - will be interesting to see how this pans out ...

[plissken]

I'll try this in Foobar ABX comparator and post the log file.

[Archimago]

Thanks guys for the submissions so far!

 

Hopefully I can get as many test results as possible... The more ears and comparisons, the stronger the data set which I think will give us a good idea of the magnitude of difference out in the "real world" compared to claims made by the company.

 

[Miska]

Where did these tracks come from, because there are so much alias images? I don't remember seeing such on 2L's original DXD material.

 

[austinpop]

Before I devote time and effort to this, just to be clear ... 

 

By "Core decode" you mean the 1st unfold software decode that the Tidal app or Audirvana + 3 do, correct? Not the full decode, which only an MQA DAC can do, if I understand right.

 

If so, I don't understand the files contained in your zip. For example, the Arnesen is a track I know well. The MQA file is distributed as 24/44.1. The fully decoded version is 24/352.8. A "first unfold" should only yield 24/88.2. Yet the zip file contains this track at a resolution of 24/176.2.

 

How did you go from 88.2 to 176.4? Was that using a non-MQA oversampler?

EDIT: I see in your blog you say "decoded with MQA Core to 24/88 and 24/96, then "rendered" with a digital filter typical of an MQA DAC"

 

Can you say more on what you mean by "rendered" here? What SW did you use?

[austinpop]

Also, regarding that same Arnesen track, how did you get the non-MQA track at 176.4?

 

If I look at the 2L Test bench site, they only offer downloads of this track at these resolutions:

• 24/352.8 (DXD)
• 24/192
• 24/96
• 24/44.1 (MQA)
• 16/44.1

 

Could you clarify? Thanks, @Archimago

[Archimago]

1 hour ago, austinpop said:

Before I devote time and effort to this, just to be clear ... 

 

By "Core decode" you mean the 1st unfold software decode that the Tidal app or Audirvana + 3 do, correct? Not the full decode, which only an MQA DAC can do, if I understand right.

 

If so, I don't understand the files contained in your zip. For example, the Arnesen is a track I know well. The MQA file is distributed as 24/44.1. The fully decoded version is 24/352.8. A "first unfold" should only yield 24/88.2. Yet the zip file contains this track at a resolution of 24/176.2.

 

How did you go from 88.2 to 176.4? Was that using a non-MQA oversampler?

EDIT: I see in your blog you say "decoded with MQA Core to 24/88 and 24/96, then "rendered" with a digital filter typical of an MQA DAC"

 

Can you say more on what you mean by "rendered" here? What SW did you use?

 

Yes, correct.

 

Although I will go into more details with the testing done to describe the procedure in full, in order to control for the digital filtering effect, what I have done is compared the MQA Core Decode at 24/88 (or 24/96) with the actual original file downsampled (eg. DXD source) to 24/88 using the same kind of upsampling filter that MQA utilizes as demonstrated previously through the Mytek Brooklyn and Dragonfly Black.

 

64-bit iZotope RX 5 was used to model the MQA-like impulse response. If anything, the 64-bit resampling algorithm is better than what I have seen when I tested the Dragonfly Black with less overload.

 

This should be a good apples-to-apples comparison since the upsampling from 88 --> 176 using the same filter will isolate the quality of MQA Core decoding compared to the "real thing" (i.e. an actual 24/88 or 24/96 derived from the source) by standardizing the filtering being used. Also, because the MQA-like filter is the default for some MQA devices like the Mytek Brooklyn, the original hi-res upsampled to this filter would be similar to the effect from such a DAC when played back on your own device.

 

[scan80269]

Very interesting questions, @austinpop!  I've had the same ones in my head.

 

I hope @Archimago can clarify but I recall reading that he will disclose only after the end of the test period.

 

Meanwhile, here's a bit of speculation on my part...

 

1. The "24/192K" and "24/176.4K" tracks representing MQA sound were created by 2x upsampling with a close approximation of one of the "dumped" MQA filters (gentle roll-off slope and very short post-ringing), modeled using iZotope Rx (??), applied to the "unfolded" 24/96K and 24/88.2K data streams ripped from an MQA DAC (before it got sent to the DAC section for "rendering").

 

Should the above turn out to be correct, then it's basically MQA "unfolding" by MQA DAC, followed by a close facsimile of the MQA "rendering" done externally (outside of the MQA DAC).

 

2. For the tracks representing non-MQA:

  A. Arnesen & Mozart tracks have DXD (24/352.8K) masters, so down-sample the DXD downloads from 2L site to 24/176.2K.

  B. Gjeilo track has 24/96K master, so up-sample the 24/96K download to 24/192K.

 

[austinpop]

Thanks guys! Interesting experiment. I will give it a listen.

[Miska]

6 hours ago, Archimago said:

This should be a good apples-to-apples comparison since the upsampling from 88 --> 176 using the same filter will isolate the quality of MQA Core decoding compared to the "real thing" (i.e. an actual 24/88 or 24/96 derived from the source) by standardizing the filtering being used. Also, because the MQA-like filter is the default for some MQA devices like the Mytek Brooklyn, the original hi-res upsampled to this filter would be similar to the effect from such a DAC when played back on your own device.

 

Unfortunately, half of the MQA story is about the upsampling filters. Using same kind of crappy filter to upsample from 2x rate does a disservice for the comparison. I would have kept the original DXD from 2L as-is, and just upsample the MQA version using such crappy filter to the same DXD rate. That would be better comparison. Now also the "hires" version is spoiled or "MQAified"...

 

(going the path you went, closer comparison would have been to also noise shape the "hires" version down to 15-bit same way as MQA does)

[crenca]

14 hours ago, Archimago said:

 

This should be a good apples-to-apples comparison since the upsampling from 88 --> 176 using the same filter will isolate the quality of MQA Core decoding compared to the "real thing" (i.e. an actual 24/88 or 24/96 derived from the source) by standardizing the filtering being used. Also, because the MQA-like filter is the default for some MQA devices like the Mytek Brooklyn, the original hi-res upsampled to this filter would be similar to the effect from such a DAC when played back on your own device.

 

 

 

I am confused by this as well (like Miska).  What is the point in processing the DXD back through a (partial) MQA treatment?  In the case of Mytek (or similiar product), it would only prove that partial MQA is worse (or better) than full MQA...what am I missing?

[scan80269]

I concur with Miska.

 

For the tracks with DXD masters (Arnesen, Mozart) the comparison should be between the DXD masters (directly from 2L) and the fully unfolded/rendered MQA also at DXD sampling rate, to simulate the output of MQA DACs like Meridian, Mytek, etc. playing 24/44.1 encoded MQA. 

 

For the track with 24/96 master (Gjeilo) the comparison should be between the 24/96 master (directly from 2L) and the unfolded MQA at 24/96.  This is assuming an MQA DAC will only perform the first unfold to get back to the original 96 sampling rate, and not do further up-sampling to 192 or 384 before doing D-to-A conversion.

 

In this blind test, the tracks representing non-MQA are not representative of the hi-res PCM sound coming out of "legacy" non-MQA DACs, since these DACs will never "benefit" from the leaky aliasing-prone filters employed by MQA.

 

Also, both the A and B tracks appear to carry elevated ultrasonic levels capable of upsetting analog equipment (e.g. preamp or headphone amp) downstream from the DAC.  The DXD or 24/96 masters from 2L sounded cleaner than either A or B tracks with my headphones setup.

 

[austinpop]

4 hours ago, scan80269 said:

I concur with Miska.

 

For the tracks with DXD masters (Arnesen, Mozart) the comparison should be between the DXD masters (directly from 2L) and the fully unfolded/rendered MQA also at DXD sampling rate, to simulate the output of MQA DACs like Meridian, Mytek, etc. playing 24/44.1 encoded MQA. 

 

For the track with 24/96 master (Gjeilo) the comparison should be between the 24/96 master (directly from 2L) and the unfolded MQA at 24/96.  This is assuming an MQA DAC will only perform the first unfold to get back to the original 96 sampling rate, and not do further up-sampling to 192 or 384 before doing D-to-A conversion.

 

In this blind test, the tracks representing non-MQA are not representative of the hi-res PCM sound coming out of "legacy" non-MQA DACs, since these DACs will never "benefit" from the leaky aliasing-prone filters employed by MQA.

 

I must agree 100%!

 

4 hours ago, scan80269 said:

Also, both the A and B tracks appear to carry elevated ultrasonic levels capable of upsetting analog equipment (e.g. preamp or headphone amp) downstream from the DAC.  The DXD or 24/96 masters from 2L sounded cleaner than either A or B tracks with my headphones setup.

 

In fact, I was just going to post about this, when I saw scan's post. 

 

I happen to have all the resolutions that 2L published for all 3 of these test tracks already, so it was easy to compare. Both A and B tracks provided by @Archimago have a harsh, fatiguing edge to them, that was puzzling, until I listened to the originals from 2L.

 

[hmartin]

Sounds like a very strange comparison. The MQA file is supposed to be optimized for upsampling with the simple interpolation filter while the DXD version if properly made will suffer a lot from that abuse.

 

If MQA keeps any of its promises the MQA version should win hands-down.... waiting for the results.

[Archimago]

There is method to the madness.

 

Remember guys, there are many potential variables. Must make sure that one is not confounding too many things. This is a test of *MQA Core decoding*. This is not a test of filters. In fact, I want to keep the filtering the same while providing people with an opportunity to hear what an MQA-like filter output sounds like. If this were about filters (eg. an MQA decoded and upsampled with MQA-like filtering vs. unmolested downsampled DXD), it would be way too easy for people to identify the 2! Anyone and their dog could just open up the files and see the differences rather than listening...

 

As an Internet test in an age of technological sophistication, I want to keep this as difficult as possible to differentiate "by eye". Just use your ears and tell me which sounds better.

 

If there is any advantage at all in the MQA Core decode to make these files sound "better" than just simple 24/88 or 24/96, then we should see the results regardness of the upsampling filtering. If there is no evidence of improvement, then all MQA and its "time domain" improvement is as far as these 2L tracks are concerned is just the different upsampling filters identified...

 

In any event, thanks to those who have already submitted results. Just short of 2 weeks to go!

 

[Archimago]

Results are being posted now...

 

Procedure (Part I)

Results (Part II)

[lucretius]

Too bad only three tracks were tested.  It would be nice to see the results over a couple dozen tracks and with various styles of music.

[hmartin]

Thanks Archimago!

 

That is interesting, I picked the Hi-res in 3 out of 3. I was expecting the poor upsampling to kill the hi-res version.... Lucky the test was blind I guess.

 

But when I have tested high-res 96/24 and downsamples 48/16 against MQA core-decoded, MQA has lost always. So maybe the high-res was "killed" but only to the extent of compensating for the lower quality of the MQA core-decoded version.

 

Any plans for a follow-up test?

 

Edit: Just a clarification, I found the difference much smaller in the blind test than I have previously experienced in my own tests.

[esldude]

Thanks for doing the test, and all the effort involved. 

 

I am sure there will be criticisms of how you did this.  No such online test is perfect.  You are doing something to get some data on how this works.  If the hype around MQA were 10% true, you would have had nearly unanimous choices for MQA, and instead the results are something not at all far from 50:50.  Good show mate!