Yes, I know. The last time I specifically addressed MQA was supposed to be "FINAL final" :-). But I got curious again. This is especially after the exploration of various filters recently suggesting to me that there actually is a case to be made that "blurring" can be seen with minimum phase filters. The question is, what do we "know" or "believe" MQA is able to do, and can we demonstrate that MQA even "de-blurs"?
Let's talk and think about this for a bit...
I. Intro...Let us mull over the question "What is blurring?" as a problem promoted by MQA needing to be resolved. MQA appears to suggest that "blurring" is ubiquitous in the world of PCM, so presumably as consumers we should be able to notice the problem and I would imagine at least we should have some general concept of the cause. Alas, as far as I can tell, the answer to that question is not so clear.
In the last couple of years, I have been in touch with some knowledgeable individuals whom I would have thought should be "in the know"... Alas, I notice a collective shrug when I bring up this question. Of course, nobody claims that every microphone, ADC, DAC, DSP plug-in, playback system operates at the same resolution. Comparatively, some would have poorer time-domain performance demonstrable with recording, manipulating or playing back signals with rapid transients. But is there some kind of clear "blurring" that is inherent in PCM as a digital representation of analogue waveforms? And is there some kind of fundamental "correction" that could be applied to studio productions for consistently "better" playback? I certainly have not heard of an affirmative to these questions from the people I spoke to...
But yet here is MQA claiming that this is possible.
While I know many of us have already read Bob Stuart's Q&A on "Temporal Errors in Audio" and "Temporal Blur in Sampled Systems" from 2016 where he lays out some facts, figures, and claims, a clear outline of what the "de-blur" algorithm actually does is rather obscure.
MQA accepts that "temporal blur" is not "quantised at the sample rate" and that 16/44 already resolves to 220ps. Rather, what "temporal blur" is seems much harder to define. Notice the first sentence of his "What we mean by temporal blur..." section: "There is no standard measure for temporal blur but we believe our use of the term is clear and intuitive." No standard measure yet the use of the term is "clear"? Is that even possible in audio engineering these days?
From there, the Q&A article proceeds with claims about how audible "fine details" seem to be "on timescales as short as 5µs". But since we already agreed that 16/44 resolves to 220ps - is there a problem!? Then he starts talking about limitations of the audio chain and how temporal errors within one component can limit the ultimate resolution of the rest... Sure. But he seems to be telling us that in post processing the loss of microsecond temporal definition can fully be repaired and you can overcome the resolution limit of that link in the audio production/playback chain after the fact!
[As an aside, I have heard of variations to that "220ps" figure. Perhaps the best laid out explanation of "time uncertainty" in a 16/44.1 system is this calculation in Kahrs & Brandenburg (Applications of Digital Signal Processing to Audio and Acoustics, 1998). Their final figure is 110 ps. The point is that we're talking picosecond values here!]
For the hardy reader who turns the page and considers the section on "Temporal Blur In Sampled Systems", the article series then meanders into idealism like how perfect low-pass "brick wall" filters are impossible (is perfect anything possible!?), appears to try to disparage fs/2 (Nyquist frequency) as either "tight engineering or lazy thinking" (???), and later on lays the groundwork for why he thinks the linear phase filter "smears" due to the "ringing". While he agrees "the human listener may not be able to hear the ringing frequency", he still insists "we are nonetheless sensitive to the overall envelope" referencing his own article on MQA. Within that paper, I do not see any testing referenced regarding "ringing" with actual listeners showing that this "overall envelope" is audible.
As best I can understand, the rest of that Q&A section jumps back and forth between criticizing "brick wall" filters, asserting that current systems are imperfect, at times seeming to justify that aliasing is okay ("The existence of aliasing is not evil—it is after all the right signal reproducing at the wrong frequency..."), eventually ending in vague discussions around the importance of "complimentary" filters in the encoding and decoding which presumably is what MQA thinks they can do "better" than current systems. Within this mash-up of incompletely argued ideas, is it any wonder then that "blurring" remains a topic of controversy?
Let us now spend a few moments exploring potential meanings of what "blurring" could be referring to, and if possible, explore whether MQA helps fix the problem...
II. Potential attributions and meanings for the term "blurring"...
A. Blurring as ringing, especially long "acausal" pre-ringing as "seen" in an impulse response waveform.
Let's not spend much time on this. As I've already discussed recently and made suggestions about using an intermediate phase filter, concerns about pre-ringing when looking at impulse responses is at best neurotic (as in obsessive-compulsive) and at worst irrational. MQA of course obliges with an impulse response that looks like this with no pre-ringing and of short tap length:
Remember that "ringing" is not a problem with properly anti-aliased audio recordings, and when it does happen during playback with a good quality linear phase anti-imaging reconstruction filter, it's typically because one is playing a hyper-compressed or clipped album that's not really mastered to sound "natural" or of "high fidelity".
In my opinion, to keep picking at ringing as an issue falls squarely in the "fear, uncertainty, and doubt" strategy of trying to persuade audiophiles/music lovers to accept something which superficially might look reasonable, but ultimately is of little consequence. This is in the same ball park as those "stair-stepped" representations of digital vs. smooth analogue waveforms; "intuitively" this might look correct, but in fact is a misrepresentation for a vast number of situations.
Given the linked articles already published, I really have nothing else to say about "ringing". It's not an issue inherent in properly mastered PCM, pre-ringing when the filter is supplied with a Dirac impulse is a response to being fed the "illegal" waveform, and I already talked about using a mild intermediate phase filter to suppress the ringing without much side effect if desired.
In the December 2017 Stereophile article by Jim Austin, it is claimed that the "de-blurring" is not simply about the filter used. Fine. But he still offers no details as to what it is!
B. Blurring as phase anomalies / group delays through the audio chain.
In the articles above, especially when examining the nature of steep minimum phase filters, we started discussions on another way of looking at "blurring". We discussed the fact that minimum phase filters create inherent phase shifts and this is something I want to focus on deeper this time in relation to MQA.
As a brief recap, remember that minimum phase filters (the "Tested filter" in the diagram below) results in typical graphs like this one off the src.infinitewave.ca SRC comparison page:
We see that there is a non-linear phase shift (relative temporal relationship) as the frequency increases.
These temporal shifts (they can be called phase deviations measured in degrees or group delays measured in microseconds) will create time-domain effects which can be seen in the waveforms of actual music as I presented previously:
To drive this point home, here's what happens when we "stack" 5 samplerate conversions using minimum phase as compared to the same manipulations in linear phase. I basically took a 24/96 150Hz square wave and ping-ponged this signal between 192kHz and 384kHz conversions in SoX using a very high quality 95% bandpass filter setting:
Every time a minimum phase DSP process is used (even something as straight forward as 192-to-384kHz sample rate conversion without attempting to affect frequencies), those temporal shifts become more prominent, eventually reducing the "slope" of the leading edge of the steep transient as you see above. This is why in general, it's better to utilize linear phase filters in audio processing (unless of course there truly is audible pre-ringing in extreme cases) because they do not cause "transient smearing".
You can imagine for example if an ADC used minimum phase filtering to capture the signal, in the studio the engineers used minimum phase EQ'ing, then performed minimum phase downsampling, and finally when the audiophile plays the signal at home with their minimum phase DAC filter, those temporal shifts are exacerbated with each step. Could MQA be trying to improve this form of "blurring"?
Who knows exactly what they're doing... But we do know MQA uses minimum phase filtering itself.
Recently, I spoke to a friend who works as a professional audio engineer with decades of experience and he was gracious enough to examine the digital filters on actual MQA DACs. He has access to measurement equipment and it's always good to have someone else independently review test results :-).
Using his Audio Precision gear, he was able to verify that both the relatively inexpensive Meridian Explorer 2 and higher end Mytek Brooklyn DAC utilized the same MQA filter design using the "Reis Test" also employed by Jon Atkinson in Stereophile to demonstrate the filter effects. Note the very long transition band. The upsampling algorithm is likely performed by the DACs' XMOS microcontroller (both 44.1kHz and 96kHz shown):
As discussed previously, MQA uses a weak anti-imaging reconstruction filter that will allow quite a bit of distortion to pass beyond the Nyquist frequency.
While we've seen graphs like the ones above already such as when Stereophile measured the Mytek, notice that we have not seen published measurements of group delay introduced by the MQA filter. Voilà, detailed measurements of each of the 4 filters available for the Mytek Brooklyn DAC using a combination of my friend's AP gear and FuzzMeasure on the Mac for group delay plots.
Fast Roll-Off (linear phase):
Slow Roll-off (linear phase):
MQA (minimum phase, slow roll-off) - notice the "overload" with the 20kHz 0dBFS tone:
I'm perhaps over-inclusive here with all these charts and graphs, but they do tell a rather complete story about the different filters available to users of the Mytek Brooklyn DAC.
The group delay plot confirm that with the linear phase filters, there is no delay between bass, mid, and treble - no "temporal smear". In comparison, both the minimum phase and MQA settings do indeed add a delay to the treble frequencies. Specifically, for the MQA filter, by 18kHz, there is about 40µs delay compared to <1kHz (not as bad as the sharp roll-off minimum phase filter with 100+µs). In other words, the MQA filter itself creates temporal smearing instead of "de-blurring"!
Also if you have a look at the square wave oscilloscope tracings in the center of each cluster of filter measurements, you can see the rise time measurement. For the linear phase filters (fast or slow roll-off), it's about 20µs, fast roll-off minimum phase has 34µs, and MQA sits at an intermediate of 24µs. Another indicator that from a time performance perspective, linear phase reconstruction filtering maintains faster transients and less smearing. Though not shown here, temporal smear is also present when the filter is applied to 88.2 and 96kHz playback (we can explore this perhaps another time).
I suppose MQA might argue that these temporal distortions introduced by the reconstruction filter are accounted for in the encoding step and therefore neutralized (they'll have to explain if this is true). What can I say? It seems like a lot of work to perform this kind of "Rube Goldberg"-esque manipulation when the bottom line is that a studio desiring to produce high resolution audio recordings could just use some high quality microphones, an excellent ADC, and maintain linear phase processing throughout to minimize transient smearing (remember that linear phase processing does introduce latency so I'm not saying this is always the best choice; the audio engineer/artist will need to decide what sounds best). On the consumer side, by using a high quality linear phase reconstruction filter (again, something like the Chord or high quality SoX settings), and you've ensured the highest quality time domain performance any human being could ask for in standard PCM capable of delivering down to a resolution measured in picoseconds!
C. Blurring as a more subjective phenomenon?
There is no standard measure for temporal blur but we believe our use of the term is clear and intuitive.
What is happening here is that the encoder (using system metadata and/or AI) resolves artefacts that are obviously different in each song according to the equipment and processes used. When these distracting distortions are ameliorated then the decoder can reconstruct the analog in a complementary way.Harman's "Clari-Fi"? Of course not, because audiophiles don't use stuff like this :-).
There is talk of a "white glove" approach to some audio tracks with impulse response measurements for each piece of gear in the production chain informing the algorithm in the "de-blurring" system. Maybe in these cases, they can use the impulse responses to determine how much group delay was introduced as the data was processed along the way. But if so, this would get very complicated very fast for multitracked recordings (all kinds of microphones, different ADCs used, etc...) with various kinds of EQ and other effects added. Hard to imagine studios putting much time and labor into doing this for a significant number of albums in their archives assuming it's even possible!
Rather, what's most likely is that MQA essentially "batch converts" masters given to them by the labels. Without the individual impulse responses and detailed studio information, the "AI" system then must guesstimate what the music "would" or "should" sound like if idealized (based on whatever "metadata" as per the statement above).
So how "accurate" would this be? Unless proven otherwise, I remain skeptical that any DSP technique would be that "intelligent" in deconstructing time-domain "artefacts" in each track of a complex recording, figuring out how many microseconds of "blur" each instrument/voice was subjected to and at what frequencies, and be able to reconstruct an almost magical version of the music at the end that accurately resembles the "studio sound" (whatever that was!).
[Of course, if batch encoding to MQA, claims of the sonic output being "what the artist/producer/engineer intended" would be rather meaningless! But that's nice that the blue/green light turns on :-).]
III. Concluding thoughts...While not fully satisfactory, these are some ideas as to what "blurring" could mean in the context of MQA's claims. Here are then a few conclusions we can draw and be quite confident in:
1. The whole "ringing" business in a linear phase filter is irrelevant and does not contribute to transient smearing or "temporal dispersion" (whatever technical term you want to call "blurring") in a practical sense. Yes, MQA doesn't use a filter with pre-ringing in the impulse response. So what? We cannot just look at a non-pre-ringing and short impulse response and make claims that it's "better" for time-domain performance because of appearance. In fact, taken to extreme those characteristics also limit both time and frequency-domain accuracy.
2. If we believe "blurring" is similar to "transient smearing" as seen in the use of minimum phase filtering (actually, any non-linear digital reconstruction filter), then the MQA filter worsens this - it actually "blurs" rather than "de-blur"! This was demonstrated using the Mytek Brooklyn DAC. The MQA filter at least is not as severe as a steep minimum phase filter when it comes to the amount of group delay it introduces, but isn't this still a bad thing when MQA is marketed as sounding "exactly how they were recorded in the studio"? Perhaps just like the fact that MQA is partially lossy, words like "exact" probably refers to "perceptually the same" rather than to a form of the Platonic ideal.
MQA could claim that somehow they have accounted for the group delay in the encoding step perhaps by purposely introducing an inverse temporal shift! But doesn't this then purposely "blur" the audio for those listening without an MQA decoder/filter? Yet they claim that even when not decoded, we're supposed to hear the benefits of "de-blurring" (as per the 2L Test Bench comments)! What a mess...
Speaking of the Mytek Brooklyn DAC and its filters again, my friend kindly also measured the distortion (THD+N) at 44.1kHz of the various filters using his AP gear:
Notice the MQA filter actually worsened the THD+N from the DAC! To demonstrate that this increase in distortion is not just with the Mytek Brooklyn, it's also present with the Meridian Explorer 2 although not as obvious due to the lower resolution of the device. The fact that the filter cannot be switched to another one in the Explorer 2 is rather unfortunate:
So... Not only does the MQA filter introduce phase shifts/group delays/temporal "blur", but it is "leaky" as a reconstruction filter allowing imaging/"upward aliasing" (as per BS), easily overloads (as can be seen with the 20kHz 0dBFS tone with the Mytek), and it also measurably worsens overall distortion!
Some objective results to consider regarding the MQA filter (@ 44.1kHz):
- Treble frequencies are delayed up to 50µs by 20kHz.I find it hard to believe that this is the kind of reconstruction filter an audiophile interested in achieving the highest playback fidelity should accept or at least feel confidant about; much less desires to see standardized across numerous devices and applied to large numbers of so-called "hi-res" albums from the big labels! Objectively, the MQA filter is the worst quality filter of the 4 offered by the Mytek Brooklyn. For this reason, I believe DAC makers must make sure the MQA filter is not being used except when decoding MQA data as intended!
- Transient rise time is slowed by 4µs.
- THD+N as measured by the Audio Precision gear goes up to 1% by 20kHz.
For example, it's good to see that Stereophile withheld a full recommendation when they reviewed the Aurender N10 recently because the device was inappropriately using the MQA filter. This is especially bad when applied to non-MQA 44.1kHz sample rate material.
3. At the end of the day, despite all the hoopla around "white glove" encoding and impulse response measurements of the production chain, "de-blurring" might just be a DSP process that is meant to subjectively "improve" sound. Who knows how well this all works. The algorithm is proprietary and under wraps. We don't know how "intelligent" it is. And as far as I can tell, MQA has never taken the time to explain things better nor truly demonstrated the potential for consumers to appreciate.
To MQA: Why not release 24/192 pre- and post- "de-blur" audio tracks for customers to have a listen to? Even better, have 2 "de-blur" versions - one using standard "batch" de-blur and the other with a "white glove" approach to show the level of improvement achievable. This will provide an opportunity for audiophiles to assess just the claimed time-domain improvements apart from worries about the MQA filter, bit-depth reduction of 24-bit material, and the whole "origami unfolding"/data compression pieces. Surely you must have access to high quality demo recordings with permission to distribute. This should not be difficult after all this time since much of the processing must already be done or highly automated (maybe just use one of the 2L test tracks)!
To close off this main blog topic in the hopes of not having to talk much more about MQA in the days ahead, I just want to address the "analogue to analogue" claim. It's as if involving the word "analogue" excuses all kinds of clear limitations evident in the digital analysis of MQA! Over the years I have heard MQA supporters parrot this claim as if significant. For example, the phrase "MQA starts with the analog signal in the studio and ends with the analog signal on playback" in this The Absolute Sound article from May/June 2016. Oooohhhh... Sounds impressive, right?
Anyone who's still unsure of what I'm talking about can watch this video from way back in RMAF 2015 (October 2015, Chris Connaker did a good job with the interview and moderation, allowing the guests to talk quite freely in those "early days" of MQA's introduction when much less was known about it):
I remain amazed by the steadfastness of the audiophile media in promoting this failure of a "philosophy" as if it's "the next big thing" towards higher resolution or can benefit the consumer. That's even without thinking about the DRM potentials. For a magazine like TAS recently to claim that a DAC that does not decode MQA is "obsolete" is both ridiculous given MQA's limitations (summarized here) and premature considering that there's barely any interest in MQA beyond the debates among a small group of audiophiles after more than 3 years since the introduction. I suspect many audiophiles at this point recognize that they're simply chasing after the wind with MQA and will eventually be disappointed if they entrust their faith in this myth of a "better" encoding format.
If you've been around the message forums, you'll know that the long-running thread on Computer Audiophile titled "MQA is Vaporware" started by Rt66indierock in early 2017 is now an unwieldy 300+ pages! Other discussions like a 50+ page thread on Steve Hoffman Forum was unceremoniously deleted from existence a couple weeks back with at least some dissatisfaction I suspect.
In regard to the Hoffman thread deletion, it's worth thinking about censorship and the lack of transparency when something like this happens. Even though a forum moderator might have the right to pull the plug, there is still something to be said about the fact that time and efforts were contributed by the membership of those on the forum. While it's within the rights, it might not be just. To close a thread to further comments if felt to be unfruitful or delete inappropriate, rude commentary is reasonable. But IMO it should not be because of disagreements, expressed unhappiness, or because someone argued a certain viewpoint. A society that values truth (even if it "offends" some) and the search for answers (even if the results contradict the agenda of others), must also fight for freedom of speech. No, MQA battles will not change the world at large, but even in our small corner of the Internet, my sense is that to completely wipe off many well reasoned thoughts and comments is obviously draconian and speaks to some kind of insecurity. Hopefully the thread can be reinstated at some point even if closed to further commentary for the sake of posterity.
Around the same time as the MQA thread deletion, the Hoffman forum also pulled the plug on the very interest poll question: "Do you think High res audio is an audible improvement over CD quality sound?". Here's a screen grab of the results up to January 12, 2018 with 813 votes (multiple selections allowed):
It is encouraging to see some sanity out there that mastering is recognized as being very important. Personally I would have voted a "Maybe" with "Only if mastering is different/improved" which I think is consistent with my previous blog commentary on this topic.
Years ago I wondered if music labels could "keep it as simple as possible" and release two mixes/masterings for albums to target different purposes and listeners (ie. a "Standard Resolution" dynamically compressed version for radio/lossy streaming and an "Advanced Resolution" master with retained dynamic range for those who value high fidelity). Obviously, not all music needs both versions. This would be similar to a movie studio releasing an SDR (Standard Dynamic Range) movie for Blu-Ray and a HDR (High Dynamic Range) remaster for UHD Blu-Ray these days. The difference is obvious especially with higher quality screens. I believe only when there is clear differentiation around what is an elevated mastering standard versus a typical release can there be any trust and "value" in the logos and certifications as representing a product that's "better". Only then will anyone be excited about owing the higher quality "definitive" version of their favourite artist's music in higher bitrate. Only then can the record labels be seen as doing something more than just selling the same thing yet again. Needless to say, IMO, MQA has no part in encouraging higher quality music production.
The way it is right now, depending on the music, I believe it may be possible to hear the difference with a recording captured and produced in 24/88+. These would have to be modern digital recordings. I stopped buying yet another version of Kind of Blue a long time ago - even if Peter Qvortrup thinks recording quality peaked in the 50's and 60's and loudspeaker performance peaked out in the 30's and 40's - LOL!
I'm not sure what degenerated in that hi-res Hoffman poll thread since by now, I suspect most audiophiles have had plenty of opportunities to hear high-res tracks and form their own opinions. Again, what a shame that open discussions were shut down.
Have a great week ahead everyone! Hope you're all enjoying the music...
Addendum - February 14, 2018
Looks like the Steve Hoffman MQA thread has been resurrected but closed to further comments as expected. Happy Valentine's Day.