MEASUREMENTS: AudioQuest Dragonfly Black 1.5 - PART 2 (On "MQA Rendering")

AudioQuest Dragonfly Black - "rendering" MQA playback. Dragonfly logo turns purplish...

For those keeping track, the story of MQA has slowly evolved over the last few years. Initially, it was supposed to be just a technology incorporated into DACs that would take an encoded file (typically 24/44 or 24/48) and almost "magically" transform this data into the "original" high resolution (equivalent to 24/192+) sound while maintaining standard PCM playback compatibility. This was the case with the first products like the Meridian Explorer 2. Over the years, it appears that concessions have been made. Software decoding rather suddenly was introduced in early 2017 with TIDAL's support for MQA; branded as "MQA Core". At around the same time, DACs were classified as either "MQA Decoders" for components that can handle these 24/44 or 24/48 files start to finish (Meridian, Mytek DACs), and "MQA Renderers" which require the computer software perform initial "unfolding" to high-res samplerates (88 or 96kHz) followed by some kind of final processing to the sound performed by the DAC presumably to allow the sound to be more "accurate" to the studio production, or "authentic" (whatever that means).

With the recent firmware upgrade to version 1.06, the Dragonfly Black DAC now belongs to this latter category of "Renderer". As far as I am aware, it is the first of these kinds of devices to be released without "Decoding" capability.

If we take a step back and look at the big picture, there are implications to this two-step technique:

1. The Dragonfly DAC does not have enough horsepower to decode the "MQA Core" algorithm itself. I've already written about this back in January 2017 suggesting that the Dragonfly 1.5's Microchip PIC32MX microcontroller wasn't a particularly powerful device. Therefore, it has to offload the processing step to the computer. In the image above, I have Audirvana+ 3.0.6 running on a recent USB-C MacBook which is able to perform software decoding. At this time, the only other general computer software that performs MQA Core decoding is the aforementioned TIDAL app.

2. The 24/88 or 24/96 "origami unfolded" output then must contain within the bitstream instructions for the MQA Renderer (whether it be the Dragonfly or another render-compatible DAC like the Mytek Brooklyn) to perform the "proper" final processing.

Other than those who have signed NDA agreements with MQA, how exactly all this works is a bit of a mystery. Technically, the presumption is that the MQA Core step is more complicated as it takes the encoded standard sample rate data and expands it to "fill in" what's needed to create the higher samplerate output up to 96kHz. The "Render" step however would have to be simpler for low-power devices like the Dragonfly. It would involve the DAC examining the bits to look for the MQA fingerprint, decode embedded instructions, and using whatever it finds to presumably tell the DAC to output the intended analogue signal.

Based on the comments from MQA over the years, especially their insistence that digital filters make a huge difference in sound (see their PR video for example), it is literally a given that whatever the "Renderer" does would involve the use of "special" filter parameters. And given their apparent dislike of linear phase filtering (perhaps fear of the the "acausal" pre-ringing bogeyman?), it's pretty well also a given that whatever MQA is doing with the sound, minimum phase filtering will be the technique of choice.

In preparing for this post, I reached out to Måns Rullgård on Computer Audiophile as he has been digging around the MQA firmware since earlier this year. In his research, he has been able to discover enough about the decoded stream such that an arbitrary WAV file can be encoded with the MQA parameters to explore how a device like the Dragonfly Black performs when responding to these instructions!

To be clear, obviously while this "reverse engineering" can shed light into the device, it's of course not necessarily complete. Also, although certain features may be uncovered, it doesn't mean these features are actually being used in real MQA-encoded music out there.

So... With Måns' gracious assistance, let's have a look at some of the main parameters that an "MQA Renderer" like the Dragonfly Black appears to respond to...

1. DITHER BIT DEPTH

This is relatively straight forward. The DAC can be instructed to perform dithering on the 24-bit MQA Core unfolded data to a certain bit-depth. The options appear to be 15, 16, 18, and 20-bits.

Here's what the noise floor looks like on the Dragonfly Black with a "digital silence" track played back and recorded with my Focusrite Forte ADC (note that the recording was done to demonstrate relative differences, not measure the absolute noise floor):

The 18-bit and 20-bit options didn't seem to make a difference in my testing. As usual, ignore the 37kHz noise which is originating from the ADC, not the Dragonfly and also the ADC modulation noise is obvious from 55kHz onward.

2. DITHER NOISE SHAPING

In the graph above, I noted "noise shaping off". Well, the MQA stream can specify if noise shaping should be turned on and of what variety.

As you can see, there are 4 options from 0-3. In the Dragonfly, settings 0 and 2 appear to be the same. Setting 1 is quite different. One interesting comment when I showed Måns this result was that it appears the Dragonfly is applying the dithering before upsampling from 96kHz with the first peak around 35kHz. In his examination of another device, the noise shaped dithering seems to be performed after upsampling to 192kHz with the 35kHz peak pushed up to ~70kHz. Interesting how different devices can be performing the technique differently.

I have also asked my friend with the Mytek Brooklyn to examine this phenomenon. From preliminary results, with the recent Booklyn 2.34 firmware, the device doesn't seem to be responding to the noise shaping parameter. Another interesting observation is that Meridian's Explorer 2 DAC (latest 1717 firmware) doesn't seem to recognize the "MQA Render" embedded data at all. I suspect this is a result of the evolution of MQA since the Explorer 2 was one of the earliest MQA-decode DACs and the firmware was not programmed to acknowledge these "Render" instructions. Basically it looks like different MQA devices could behave differently!

3. RESAMPLING FILTER SELECTION

Embedded within the MQA stream is another parameter that tells the Renderer DAC the "original" samplerate of the encoded music. This can range from 44/48kHz all the way to samplerates no audio device currently supports :-).

With the Dragonfly, by default, if we just send a non-MQA 24/96 signal to it, it uses a filter with the following impulse response:

As you can see, this is the normal sharp minimum phase response similar to what we saw last week when looking at the non-MQA measurements.

However, if the device is told that this is "originally" a 24/96 signal and the MQA DAC should "render" it as such, the filter changes and the impulse response looks like this:

A very short minimum phase filter is used. Hmmm. OK, no surprise given the MQA PR material.

Where it gets interesting however is when the Renderer is told that the "original" resolution requires further upsampling. For example, if the "original" samplerate is specified as 192kHz, the Dragonfly will then pick from up to 32 pre-specified filters selected within the data stream to use. I don't know if all 32 potential settings are defined or not, and it really don't matter for the sake of this article. For illustration, I'll just show you the impulse responses from settings 0-8 along side the standard filter shown above:

I have no idea how or why MQA selected these filter variants. I'm not even sure if there's some kind of systematic order to them! At least we can say filters 0-8 are all minimum phase, of very short duration with minimum post-ringing. Of these first 9 filters, clearly "MQA 2" looks most different. According to Måns, analysis of MQA files "in the wild" suggests that filters 4, 6, and 8 appear to be the common ones with 4 being somewhat of a default.

My assumption is that if MQA somehow is "authentic", then all MQA certified DACs should utilize the same set of filters... It will certainly be interesting to see if other DACs produce the same table of filter variants!

Thoughts...

Måns has found a few other interesting "features" in MQA which I'll leave for him to describe as he sees fit in the future :-). But for the purpose of our exploration of the Dragonfly, I think the above captures most of what a "Renderer" does. As you can see, it's not really that complicated. Other than upsampling with the various filter settings, there is no evidence that the Dragonfly is able to "unfold" any more data from the MQA Core's 88/96kHz output into the 192kHz or 384kHz realm for example. The obvious question is therefore: just how much difference can this make in the sound!?

To answer this question somewhat, we have to go back to remembering why we want digital filters in the first place. For more, I refer you to the article from last year: Digital Interpolation Filters and Ringing among others I'm sure you'll find online. The bottom line is that a reconstruction filter allows us to remove imaging distortions during playback. Presumably, if we are interested in "high fidelity" and what is "heard" in the studio, we would want to reproduce the frequencies intended to be included in the audio file within the limits of the samplerate (ie. below Nyquist) and not want distortion components seeping through.

Remember that an impulse response is the waveform of what an isolated 100% amplitude sample looks like when passed through the filter. It shows us the coefficients used for the FIR "taps". With a larger number of taps (longer filter), one can achieve better stopband attenuation with good alias suppression. But longer filter lengths doesn't implicitly mean that there's more "temporal smearing" of audible frequencies nor would there automatically be more ringing in the audio output just because it's present in the impulse response! If you look at a well recorded piece of music, low-pass filtering will have already been done in the studio and there will not be any distortions in the signal to trigger "ringing" when upsampled. The only time you will see (and perhaps hear) ringing will be with very poor quality, typically extremely loud, compressed music with clipping and square wave anomalies.

As expressed in my "Digital Interpolation Filters" post, it's not really good enough to just look at the impulse response... We have to consider the Fourier transform pair and understand what's going on in the frequency domain! Notice how MQA keeps harping on the idea of "smearing" in digital audio, and emphasizing "time domain" performance as if this is some kind of idol which everyone should bow before simply because of these short impulse responses while neglecting the importance of what's happening in the frequency spectrum with these filters.

Well folks, let me show you... First look at the normal Dragonfly filter (presumably standard ESS filter setting for the DAC chip) as it plays 24/96 data. It does a pretty good job filtering out audio >48kHz with fairly steep attenuation such that harmonics such as the cluster around 60kHz is down by about -60dB compared to the primary signal at 20kHz:

Here's what happens if we use the 96kHz MQA filter:

Not bad still when it comes to the 19 & 20kHz tones. But it is a weaker filter and the main difference is as you can see, the steep filtering around Nyquist at 48kHz is gone and imaging is allowed to creep into the 50-60kHz frequencies with the wideband noise signal.

And suppose now we tell the Dragonfly that the "original" resolution was 192kHz and it should pick MQA Filter 2 for further upsampling:

Hmmm. Clearly we're now seeing frequencies extending all the way to 96kHz (using a 192kHz recording) even with the -4dB wideband noise signal. We're now also seeing imaging of the 19 & 20kHz tones up at 76 & 77kHz!

Here's a look at Filter 4 which as I mentioned above is a very common "default" choice in the music released thus far in MQA:

Again, like in Filter 2, the wideband noise is allowed to leak all the way to 96kHz. Also, there's the 76 & 77kHz "image" again. If you look a little closer, you will also notice that the Filter 2 and 4 graphs have slightly lower 19 & 20kHz sine peaks. This is because my usual -6dBFS peaks caused severe  overloading with these MQA filters whereas this was not a problem with the standard non-MQA filter and even the MQA 96kHz filter. I had to lower the 19 & 20kHz sine amplitudes to -9dBFS to avoid clipping.

Whether you "like" the look of the digital filter impulse responses or think that there's any substantial improvement in "time domain" performance, these filters used in the Dragonfly Black when "rendering" the MQA Core decoded audio are just plain poor when it comes to antialiasing performance! I might as well tell my TEAC UD-501 DAC to turn off the digital filtering altogether and let TIDAL software decode to 24/96 and I would have essentially achieved the kind of effect we're seeing.

Using RightMark, we can also appreciate that the filters do measure somewhat differently:

What we see here is a comparison done at 24/96 using test signals tagged with MQA and with different filters selected. We have the standard non-MQA filter, the MQA filter for 96kHz decode, then a series of 192kHz filters - the somewhat atypical appearing Filter 2, and the most similar and commonly used 4, 6, 8 trio found in most of TIDAL's albums out there. Remember that the numerical summary only analyzes the data from 20-20kHz so the values look very similar. However, the graphs show us where the differences lie... As you can see, the filters do change frequency response somewhat although none will dip more than -1dB at 20kHz. Also we can see that they create varying levels of distortion as shown with the IMD+N sweep. Not surprisingly, the stronger, more "orthodox" standard 24/96 filter results in the flattest frequency response, lowest noise, and potentially least distortion.

I appreciate that with real music, and real ears, the sound will not appear as noticeable as the diagrams above since naturally we can't hear all that noise above ~20kHz and recordings typically don't contain much ultrasonic high frequency content. But is there any rationale to think that this technique improves "temporal smearing" when we see all this frequency anomaly!? Remember that depending on the equipment itself, the imaging distortions above Nyquist can fold back into audible frequencies with varying severity and have unintended effects on what is heard (check out this YouTube video for a demonstration of what "aliasing"/imaging presents like). I think it's more likely that all those who hear "huge differences" between MQA and non-MQA Studio Masters are hearing the effect of the distortion than some kind of temporal domain improvement. Presumably these people like what the ultrasonic imaging is doing in their systems!

If you're wondering, yes, these anomalies do show up with actual MQA-encoded music. For example, here's a spectral plot at the same point in TrondheimSolistene's excellent recording of Britten's "Simple Symphony" which can be freely downloaded from 2L:

If we take the 24/192 download and play back through the SMSL iDEA DAC in GREEN as "reference" (remember, the Dragonfly Black cannot play 192kHz), we see that the Dragonfly Black playing the official 24/96 download is very similar in comparison (in RED). Likewise, without any special "rendering", the SMSL iDEA's playback using Audirvana+'s MQA Core decoding into 24/88 from the MQA file in YELLOW appears quite capable of reconstituting the source material up to ~44kHz (notice that cluster of frequencies above 35kHz is generally retained).

But look at what happens with an actual "rendering" using Audirvana+ 3 and the Dragonfly Black with MQA activated in BLUE (ie. the dragonfly LED is now that MQA purple color). Clearly, some details are being drowned out in noise like that cluster from 35-45kHz. Also, we can see all those frequencies above 60kHz which were not present in the original recording! This is the result of the weak filter and subsequent imaging distortions.

I had a listen to some of the MQA-encoded 2L tracks ("Simple Symphony", "Arnesen Magnificat") as well as some off TIDAL (Led Zeppelin, Madonna, Beyoncé, The B-52's, Eric Bibb) and compared the sound with "Studio Master" versions (typically from HDtracks). Whether I thought the MQA version was better or not is really hard to say subjectively. Sure, MQA-decoded playback sounds similar compared to the same "Stereo Master" download. But I can't in all honesty claim that MQA adds a more "relaxing" property nor did I hear anything wonderful like an expansion of the soundstage, added bass, or other such superlatives some have claimed. I've done comparisons before with software and hardware decoding and I'm not hearing nor "seeing" anything here with the Dragonfly Black that would change those opinions substantially. If anything, I'm actually more concerned about this MQA "Rendering" process of the Dragonfly Black deteriorating sound quality.

Conclusions...

Over the last few weeks, as I listen to the Dragonfly Black and try to wrap my head around what I'm hearing and seeing with MQA, I continue to ask myself the simple question... Do I want this?

As you know, over the last few years I've put up a number of blog posts exploring MQA. How it's supposed to do what it claims, what we're seeing empirically with the first MQA DACs, sonic effect with TIDAL software decoding, and more recently use of hardware for decoding... And today, exploring what's "under the hood" with the "MQA Rendering" function thanks to Måns and his work over the months.

The bottom line is that I am not surprised by what I'm seeing. As I had suspected back in January, rendering is basically the DAC looking out for an MQA "command" in the data stream, setting the DAC's digital filter and associated low-level processing (eg. dithering and/or noise shaping). As expected (and speculated on recently), the MQA folks like minimum phase and short FIR filters.

So, do I want this? Let me be blunt, while the blue frequency spectrum above with "Simple Symphony" might sound different, and some people might like it (the extra noise, like with DSD64 playback can be pleasing to some), I personally think this is not a good effect and antithetical to high fidelity playback. In fact, playing the MQA track decoded using Audirvana+ without "MQA Rendering" on the SMSL iDEA results in better low-level detail and objectively is closer to the original "master". Like I said not long ago, that term "authenticated" (MQA) appears completely meaningless at least when it comes to sound quality.

No. I personally do not want this. So far, I have seen no evidence that what this is doing actually improves quality. "Time domain" improvements, if any, appears to be contingent on the idea that one believes these short filters can achieve the goal and are thus desirable. I do not see where the company has so far provided evidence for their claims. If there are improvements that can be made by some kind of proprietary time-alignment DSP, then just offer it as such instead of wrapping it into this unnecessarily complex encoding scheme. As for the data compression being beneficial... Well, since the announcement of MQA in late 2014, my Internet download speed has already gone up from 25Mbps to 150Mbps for around the same price. 4K HDR video streaming with multichannel sound is a reality. And Qobuz is now offering Sublime+ lossless FLAC audio streaming up to 24/192. So why bother streaming partially lossy 24/48 when you know technically full resolution, unmolested 24/96 or 24/192 can actually already be done and will only get easier and cheaper in the years ahead?

As I end off, I want to bring up again the obvious dichotomy between what MQA is doing here with their weak filters and what Chord is doing with their DACs like the DAVE. They're like polar opposites! Chord is using a proprietary FPGA-based, very steep, "brick wall" filter with literally uncompromised antialiasing using a reported 164,000 tap linear phase filter, and on the other side we have Meridian/MQA with their insistence on these minimum phase, "porous", weak filters desiring that DAC manufacturers should go this route (for a licensing fee of course).

So which approach is right? I guess I can offer you the usual recommendation... "Go listen for yourself." :-)

As for myself, I believe Chord is "more right" than Meridian/MQA. My sense is that the MQA filters above are much too weak and sonically dirty (I believe some impressionable members of the audiophile press like dirty sound if they indeed hear what they claim they hear!). A very steep filter like the DAVE might not be necessary but I typically listen to upsampling with strong filters like that and enjoy the sound quality with no issues audibly or intellectually around time-domain performance.

Again, I would like to thank Måns Rullgård for all his work and stimulating ideas! He is certainly the brains behind allowing us all to have a peek under the MQA hood to see a substantial portion of how it ticks. Also, a big shout out to my buddy John for lending me the Dragonfly Black for the last few weeks of testing!

----------------------------

Recently, I came across this interesting YouTube post entitled "Tidal Masters MQA - Why the Frequency Spectrum Gap?":

Well, in light of what we're discussing here, you probably guessed that the cause for this "gap" is because of the MQA filter. Let's reproduce what the YouTuber showed...

Take Beyoncé's song "Love Drought" (off Lemonade) as used in the video (16/44 is fine); here's a convenient spot 40 seconds into the track:

We see that the 44kHz track rolls off starting at 20kHz.

Now, let's upsample with a filter setting that is modeled after MQA... Here's an iZotope RX 5 setting that approximates the 24/96 MQA filter from the Dragonfly:

Now let's have a look at what happened to the 44kHz track upsampled to 96kHz (it could of course just as well have been 88kHz, doesn't matter). Have a peek at the same spot as above, 40 seconds into the song:

Et voilà! There's your "gap". What you're seeing is a "mirror image" beyond 22.05kHz (Nyquist for 44.1kHz samplerate) due to imaging from the use of the MQA-like digital filter that has poor anti-imaging properties. These are of course false frequencies that were not part of the original song. Because this track has that significant roll-off of frequencies around 20kHz, it's easy to see this U-shaped anomaly.

Remember back in the day when HDtracks sold some fake high resolution tracks that were just upsampled 44/48kHz material as 88/96kHz? Well, this is an example of what happens when 44/48kHz audio is "unfolded" into 88/96kHz by the MQA Core decoder in TIDAL using their "leaky" filter of choice allowing the ultrasonic distortions to seep through. It's an example of "fake hi-res" MQA unfolding.

We have to really question what it means to be "Master Quality Authenticated" when this kind of stuff happens. Clearly that aliased image is NOT part of what Beyoncé and the engineers created, "heard" or likely intended to be present in the recording! I find it rather amusing that Lemonade in MQA should even be promoted as something worthy of "Master Quality".

Because of this kind of stuff happening, I applaud HiresAudio for taking a stand on not selling MQA files because of an inability to verify the content as true high resolution. The fake high frequency material is easy to see here, but if it were not for the 20kHz roll-off to create that ugly "gap", it would certainly be much more difficult to realize that there is something very wrong.

There is no magic in digital audio. And as far as I can see, there is no "gnostic" wizardry in MQA either when we look a little deeper. To all the "evangelists" out there... Beware of worshipping at the alter of the god of "time-domain" with salvation offered in the form of MQA and its weak filters. You'll likely find your faith challenged and ultimately recognize that it was misplaced despite a genuine desire for higher fidelity.

Remember, you can find significant levels of satisfaction in frequency & time-domain accuracy with room correction measurements and DSP. There is no free lunch, and it will likely take time and effort to get it just right for your DAC, speakers and sound room. For reference, here's the zoomed in 24/96 correction impulse response used for my DAC (TEAC UD-501), room and speakers (Paradigm Reference Signature S8) created with measurements using Acourate meant to be used with a convolution engine like with JRiver or BruteFIR. Notice that in real life, when customizing the impulse response for the system, the coefficients are more complex than those MQA filters and also varies for each stereo channel depending on arrangement and slight differences between speakers:

----------------------------

Finally we're seeing some great weather here in Vancouver. It's summer time and as I mentioned weeks ago, I've got some R&R coming plus exciting projects on the go at work. Hope the weather's good wherever you are...

I'll likely be "gone fishing" for a little bit :-).

Enjoy the music!

Addendum December 2017:
Check out Mans Rullgard's Audioquest Dragonfly Black Teardown. Great stuff!