MQA spectrum plots

[comedus06]

I don't mean to be rude but I don't think you're following:

 

The file that Mansr picked is from the 2L file collection which has the same DXD recording in multiple formats/sampling sizes (including MQA) so it is the same master ("actual input of (and before) the encoder) which would then be encoded using MQA (since MQA is not a recording format, nothing can be recorded in MQA, it can only be encoded as such after the fact). So, the series of spectrographs in the first are exactly what you suggesting - comparing the DXD which is the actual input of (and before) the encoder and the output after the decoder (Bluesound) has done its thing.

 

Then read post #50 on his assertion as to why he is using an unEncoded sine wave or sweep sine to analyze performance of the decoder independent of the encoder.

 

Hope that helps.

 

 

I get it now, thanks for clarifying.

 

But I still think that to be able to really analyse what MQA is doing one should in fact be able to compare the actual input of (and before) the encoder with the output after the decoder. I understand now this is hard or impossible to achieve and therefore a DXD file has been used for comparison.

 

Sent from my HTC One_M8 using Computer Audiophile mobile app

[audiventory]

[ATTACH=CONFIG]32298[/ATTACH]

 

Thank you, Mansr. Now we can see full band picture.

 

I see at the figure the alias harmonic with amplitude from -70 dB at zero frequency.

 

Possibly about 10 kHz the decoder overloaded (filter's amplitude-frequency). If it is so, the decreasing of level of the sweep sine can help.

 

 

 

Impulse response of the resampler:

 

 

[ATTACH=CONFIG]32299[/ATTACH]

 

Minimum phase. No surprise there.

 

 

Yes. I suppose, it is that called as «anodising filter».

[Nikhil]

Yes. I suppose, it is that called as «anodising filter».

Did you mean an apodizing filter?

 

Found this via Google: https://mrapodizer.wordpress.com/2011/08/16/technical-analysis-of-the-meridian-apodizing-filter/

[audiventory]

Did you mean an apodizing filter?

Yes. Apodizing filter. Thank you cor correction.

[Vincent1234]

I don't mean to be rude but I don't think you're following:

 

The file that Mansr picked is from the 2L file collection which has the same DXD recording in multiple formats/sampling sizes (including MQA) so it is the same master ("actual input of (and before) the encoder) which would then be encoded using MQA (since MQA is not a recording format, nothing can be recorded in MQA, it can only be encoded as such after the fact). So, the series of spectrographs in the first are exactly what you suggesting - comparing the DXD which is the actual input of (and before) the encoder and the output after the decoder (Bluesound) has done its thing.

 

Then read post #50 on his assertion as to why he is using an unEncoded sine wave or sweep sine to analyze performance of the decoder independent of the encoder.

 

Hope that helps.

 

Thank you, you're not rude at all and I got all that already. I also read post #50 before which clearly shows some pretty strange behavior. I also fully get it that the same 2L master has been used for both the original DXD and MQA encoded files.

 

My remaining point is however that it might be the case that Bluesound's MQA decoder isn't doing exactly what it's supposed to do. Is that a possibility? I don't want to rain on your parade and I do appreciate all these measurement efforts! But lots of brands will be handling MQA decoding (possibly making the system more vulnerable btw..!) and it might be hard to draw overall conclusions based on only one decoder from a not so high-end brand in a very early stage of MQA..?

Or am I seeing ghosts here and is every decoder 'MQA-certified'? I know dCS is writing their own code too, for instance..

Would it be possible to do similar tests with a Meridian DAC to stay 'closer to the source'?

 

Sent from my HTC One_M8 using Computer Audiophile mobile app

[Jud]

Yes. Apodizing filter. Thank you cor correction.

 

Not all minimum phase filters are apodizing (not sure about whether all apodizing filters must be minimum phase - don't think so?). Someone who actually knows something can say for sure.

 

 

Sent from my iPhone using Computer Audiophile

[matthias]

Not all minimum phase filters are apodizing (not sure about whether all apodizing filters must be minimum phase - don't think so?). Someone who actually knows something can say for sure.

 

 

Sent from my iPhone using Computer Audiophile

 

From WBF:

 

Apodizing filters - a fad that has passed or standard these days? "Revolutionary"?

 

Matt

[audiventory]

Not all minimum phase filters are apodizing (not sure about whether all apodizing filters must be minimum phase - don't think so?).

 

As far as I know, apodization is removing (full or partial) of ringing in general.

 

Minimal phase is secondary feature of filter that have oscilations behind the front of input signal.

 

Oscillations behind the front of input signal is demanded feature of the filter.

 

Minimal phase is еру feature of such type of FIR filter that we can't avoid, but can try approach to linear phase.

 

There no pre-ringing, but post-ringing 2 times more.

[adamdea]

The MQA decoder will gladly upsample any input to twice the rate, even if it is not an MQA stream. It stands to reason that this uses the same resampler as when actually decoding (much of the same code is executed in both cases). This means we can use synthetic test signals to probe some of the behaviour.

 

First, white noise at -3 dBFS:

[ATTACH=CONFIG]32286[/ATTACH]

The dashed vertical line marks 22.05 kHz. From the graph, we can see that the filter is down by only 3 dB at fs/2 and arrives within 3 dB of maximum attenuation at roughly 31 kHz. The stop-band attenuation is also quite poor.

 

Next, a 1 kHz sine wave at various amplitudes:

[ATTACH=CONFIG]32287[/ATTACH]

WTF, how is this even possible? As expected, there is a strong alias component at 43.1 kHz. Less expected is all the low-level distortion even with a -60 dBFS input. There is also significant noise modulation based on input level. Note that the squiggles immediately around 1 kHz are an artefact of the FFT windowing. Everything else is caused by the MQA resampler.

 

Here's a 10 kHz sine wave:

[ATTACH=CONFIG]32288[/ATTACH]

 

And a 20 kHz sine wave:

[ATTACH=CONFIG]32289[/ATTACH]

 

The distortion products seen in the last two graphs are readily audible.

I'm a little confused about how these synthetic test signals can meaningfully be applied. Presumably the MQA decoder is expecting an MQA encoded signal. If fed a 24/48 1/10/20 kHz test tone, which will contain pure pcm (presumably dithered at the 24th bit) it is surely receiving an "illegal" signal ie one which could never be the result of MQA encoding.

 

It would hardly be surprising that the MQA filter "reading" the bottom bits would result in the creation of spuriae in the upper band based on reading baseband information in the test signal as encoded upperband information. It seems to me that it may also create baseband spuriae (for example if it results in the first 15 bits or so being read as an undithered 15 bit baseband signal, perhaps for other reasons.)

 

Apologies if I have misunderstood: I'm very interested in your work to try to understand what is going on, but I'm not sure that it is meaningful to put an unencoded file into a decoder. Perhaps I have misunderstood were the test tones you put in ordinary 24 bit pcm or were they made to at least conform to what an mqa file might look like in the lower bits (something like 13 or 14 bits plus pseudo noise)?

 

 

I've been thinking about this for a while since looking at the stereophile measurements for the mytek brooklyn's MQA filter at figure 6 here Mytek HiFi Brooklyn D/A processor–headphone amplifier Measurements | Stereophile.com

 

These look similar to your 19.1khz test tone. But I'm not sure what they mean. I did put a query on the stereophile website but received no reply.

[mansr]

Thank you, Mansr. Now we can see full band picture.

 

I see at the figure the alias harmonic with amplitude from -70 dB at zero frequency.

 

Possibly about 10 kHz the decoder overloaded (filter's amplitude-frequency). If it is so, the decreasing of level of the sweep sine can help.

 

Yes, lowering the level by 3 dB leaves only the fundamental and the alias frequency one would expect from the frequency response plot. Although an actual recording typically has some headroom, I see no reason for the filter to break down like this with a full-range input.

[Jud]

As far as I know, apodization is removing (full or partial) of ringing in general.

 

Minimal phase is secondary feature of filter that have oscilations behind the front of input signal.

 

Oscillations behind the front of input signal is demanded feature of the filter.

 

Minimal phase is еру feature of such type of FIR filter that we can't avoid, but can try approach to linear phase.

 

There no pre-ringing, but post-ringing 2 times more.

 

 

Hi Yuri -

 

I don't know what the word "epy" means above (I bolded it) - can you tell me what you meant?

 

 

This is something I found with Google, in case it explains anything - I don't know enough to know whether it does: H I F I D U I N O: Apodizing vs Non-Apodizing

 

 

Edit: Also, here - http://www.whatsbestforum.com/showthread.php?13391-Apodizing-filters-a-fad-that-has-passed-or-standard-these-days-quot-Revolutionary-quot&p=248796&viewfull=1#post248796

[mansr]

My remaining point is however that it might be the case that Bluesound's MQA decoder isn't doing exactly what it's supposed to do. Is that a possibility?

 

There is always the possibility that I made some mistake running the decoder. I'll dig a little deeper and see what I find.

[PeterSt]

J.A. on the Mytek Brooklyn :

 

Following publication of this review, I reexamined the behavior of the Brooklyn's MQA reconstruction filter at lower signal levels. The spectrum with a 19.1kHz tone at –0.1dBFS, –0.4dBFS, and –0.7dBFS was identical to that at 0dBFS (fig.6, blue and cyan traces) but with data at –1dBFS, the spectrum looked dramatically cleaner (fig.6A). This suggests that the MQA filter overloads with high-frequency signals that approach 0dBFS. As Bob Stuart has discussed, real musical spectra does not have full-scale content at the top of the audio band, meaning that the filter's resolution can be optimized for low-level signals.

Mytek HiFi Brooklyn D/A processor–headphone amplifier Measurements | Stereophile.com (thank you adamdea)

 

Mansr, this seems to imply that you must be able to show the same, when at -3dBFS (all clean except for the clear aliasing). But I don't think this happens in your case. So your decoder is different or something else is going on. Maybe it is data dependent ? (what about dither vs no dither ?)

[audiventory]

I've been thinking about this for a while since looking at the stereophile measurements for the mytek brooklyn's MQA filter at figure 6 here Mytek HiFi Brooklyn D/A processor–headphone amplifier Measurements | Stereophile.com

 

These look similar to your 19.1khz test tone. But I'm not sure what they mean. I did put a query on the stereophile website but received no reply.

 

Looks like, at the Figure 6 noise level limited by noise floor of spectrum analyzer. MQA should have reserve, as shown at unpacked MQA graph by Mansr above.

 

Also 19 kHz close to 0 dB.

 

Yes, lowering the level by 3 dB leaves only the fundamental and the alias frequency one would expect from the frequency response plot. Although an actual recording typically has some headroom, I see no reason for the filter to break down like this with a full-range input.

 

Thank you for new information. Overload generate only harmonics of higher order for single sine. But there 2 sine (basic and alias), hense overload picture is more sophisticated.

 

I don't know what the word "epy" means above (I bolded it) - can you tell me what you meant?

 

Hi Jud,

 

It is «the» in Russian keyboard layout Sometimes Mac don’t switch language (I use dictionary sometimes), and some words need re-enter several times.

 

 

This is something I found with Google, in case it explains anything - I don't know enough to know whether it does: H I F I D U I N O: Apodizing vs Non-Apodizing

 

 

Edit: Also, here - Apodizing filters - a fad that has passed or standard these days? "Revolutionary"?

 

«Both filters are minimum phase, (thus eliminating the pre-ringing), but the Apodizing removes aliasing distortion and the soft-knee reduces post ringing»

 

Would be interesting look to impulse response both filters.

 

Minimum phase don’t define ringing suppression as itself.

 

Pre-ringing eliminated there by using minimal phase filter.

 

Post-ringing of the anodizing filter in the article is decreased by sloper transient band.

 

If we apply software filter with -200 dB stop band there will more post-ringing.

 

I.e. border between apodizing and non-apodizing filter is blurred.

[mansr]

J.A. on the Mytek Brooklyn :

 

Following publication of this review, I reexamined the behavior of the Brooklyn's MQA reconstruction filter at lower signal levels. The spectrum with a 19.1kHz tone at –0.1dBFS, –0.4dBFS, and –0.7dBFS was identical to that at 0dBFS (fig.6, blue and cyan traces) but with data at –1dBFS, the spectrum looked dramatically cleaner (fig.6A). This suggests that the MQA filter overloads with high-frequency signals that approach 0dBFS. As Bob Stuart has discussed, real musical spectra does not have full-scale content at the top of the audio band, meaning that the filter's resolution can be optimized for low-level signals.

Mytek HiFi Brooklyn D/A processor–headphone amplifier Measurements | Stereophile.com (thank you adamdea)

 

Mansr, this seems to imply that you must be able to show the same, when at -3dBFS (all clean except for the clear aliasing). But I don't think this happens in your case. So your decoder is different or something else is going on. Maybe it is data dependent ? (what about dither vs no dither ?)

 

This matches what I've seen. I just don't think it's a competent way to design a filter.

[Vincent1234]

There is always the possibility that I made some mistake running the decoder. I'll dig a little deeper and see what I find.

Thanks, much appreciated! I do have the gutt feeling that there is more to it than we have discovered until now.

 

Sent from my HTC One_M8 using Computer Audiophile mobile app

[Jud]

Does looking at test signals versus the 2L test bench music give some indication of what the encoder must do?

[audiventory]

Does looking at test signals versus the 2L test bench music give some indication of what the encoder must do?

 

Test signals don't hide subtle detais of the decoder reaction.

2L signal show how must look coded signal.

 

It is 2 complementary approaches to understanding how it work.

 

Also the articles may be interesting http://www.audiomisc.co.uk/MQA/origami/ThereAndBack.html

[Miska]

Test signals don't hide subtle detais of the decoder reaction.

2L signal show how must look coded signal.

 

It is 2 complementary approaches to understanding how it work.

 

Music signal difference testing is good for lossy codecs where there is some form of dynamic bandwidth reservation going on and codec needs to encode complex signals in limited available storage bandwidth. This is why I chose that approach long ago for comparing different lossy codecs like MP3/AAC/Vorbis codecs. For example different MP3 and AAC encoders behave in different ways.

[adamdea]

Yes, lowering the level by 3 dB leaves only the fundamental and the alias frequency one would expect from the frequency response plot. Although an actual recording typically has some headroom, I see no reason for the filter to break down like this with a full-range input.

So, broadly, what we were seeing with the 10 and 20 kHz test tones at 0 db was clipping?

[asdf1000]

Wow great thread @mansr. I don't know enough to understand all the reasoning and testing methods but I like that people are questioning what is being sold by MQA. It's good healthy technical discussion, which I don't understand but I like following the conclusions/interpretations and recommendations for better testing.

 

No doubt the MQA guys are now reading this thread too, even if they're not responding. I don't mean that in a negative way either - it's good discussion by passionate audiophiles.

[mansr]

So, broadly, what we were seeing with the 10 and 20 kHz test tones at 0 db was clipping?

 

There's no clipping evident in the output. It is possible that some internal calculation clipped or overflowed.

[adamdea]

There's no clipping evident in the output. It is possible that some internal calculation clipped or overflowed.

I was speaking loosely; but it appears that the really horrible output with a test tone is limited to tones over -1dB. does it do that for an ordinary music signal (comprising a broad spectrum) reaching >-1dB. I wonder what it would do with a cymbal crash (?).

[audiventory]

I was speaking loosely; but it appears that the really horrible output with a test tone is limited to tones over -1dB. does it do that for an ordinary music signal (comprising a broad spectrum) reaching >-1dB. I wonder what it would do with a cymbal crash (?).

 

DSD may has -6 db headroom for encoder stability.

[scintilla]

I see a business opportunity here for an enterprising digital filter designer to make plugins that finish the "unfolding" and "de-blurring" for software decoders such as Roon. Given that the Blusound decoder was designed with the NAD PWM DAC architecture, it should already have that transfer function considered into it's final (apparently upsampling) filter design. I'd pay for that to use with my M51 DAC out of HQP from Roon. Just saying. Anyone?

 

I mean, even if MQA is a botch, I'd rather hear it fully decoded to hardware spec accoring to their purported end-to-end strategy.

 

Good thread. Keep the reverse-engineering going. It's only a matter of time before it's cracked.