COMPARISON: TIDAL / MQA stream & high-resolution downloads; impressions & thoughts...

As I mentioned last week, and I'm sure you've seen all over the audiophile news, TIDAL has started streaming MQA audio and has embedded a software decoder into the Windows/Mac desktop player. It will basically take a 24-bit 44kHz or 48kHz stream that's encoded by MQA and spit out an 88kHz or 96kHz data stream to send out to your DAC; whether the internal one in your laptop, or a fancy external DAC with options for "Exclusive" mode which allows changes to the appropriate samplerate (I know this works well on the PC, have not tried the Mac).

If you're not using "Exclusive" mode, you can tick "Force volume" to set it to 100% volume so the internal mixer/dither routine hopefully doesn't mess with it. "Passthrough MQA" should be ticked only if you have an MQA enabled DAC or want to purposely hear MQA undecoded (I'll say it now that this is not recommended). My assumption is that if you do have one of these MQA DACs and passthrough is on, you should either make sure "Exclusive" mode is ticked or if not, manually make sure the OS samplerate is correct (ie. 44kHz or 48kHz at 24-bit depth) and that the volume is 100% (either with "Force volume" or making sure the computer volume slider is 100%). Otherwise, it will not be "bit-perfect" and the DAC will not recognize the MQA encoding. I suspect this could be confusing for some.

Remember, over the last year, I have posted already on MQA - here and here. And summarized my feelings that it's essentially a smart, partially lossy encoder. My intent today is not to review the discussions previously or to rehash specialty hi-res audio files like those by 2L... The availability of a software decoder doesn't change those impressions nor general criticisms which we can still think about and summarize near the end of this post. However, it does give us a look at MQA in the context of real music that would be consumed by the average TIDAL subscriber.

What I wanted to have a look at today is a few of the "Masters" albums available and a peek at what the MQA encoding/decoding did compared to what I believe are the hi-res downloads they originated as. As you can see above, one of these "Masters" albums is Buena Vista Social Club, released back in 2000 and has been available as a 24/96 DVD-A for years.

Purely for the purpose of study and comparison, I digitally "extracted" the decoded MQA audio from TIDAL during playback (at 24/96). I see that these albums have been mentioned among some forum discussions as sounding good and examples of demonstrable improvement due to MQA. Ostensibly, these improvements in sound quality are due to high-resolution MQA encoding/decoding and the proprietary "deblurring" process promoted. I chose the following songs from various albums since I also have high-resolution download versions to compare:

1. "Good Times Bad Times" - Led Zeppelin, from Led Zeppelin (1968).
2. "Chan Chan" - Buena Vista Social Club (2000).
3. "A Case of You" - Joni Mitchell, from Blue (1971).
4. "Material World" - Madonna, from Like A Virgin (1984).

The Led Zeppelin seems to be the 2014 HDtracks Studio Master, Buena Vista probably originates from the 24/96 DVD-A, the Blue album could be the 2013 HDtracks 24/96 remaster download, and Like A Virgin is probably the 2012 HDtracks download. Here are the DR analysis results of each track comparing the TIDAL/MQA decode with the album version I have:

As you can see, the peak and RMS amplitude values are almost all within 0.1dB for each of the tracks. This suggests that indeed the MQA stream likely originated from the same mastering as my 24/96 downloads or DVD-A rip.

Limitations of the source material are also the same. For example, the 1984 Madonna appears to be an upsampled old digital recording run through an analogue stage to get the 24/96 digitization:

Notice the 29kHz noise likely from the analogue transfer is retained in the MQA.

That's just an FFT capture of the signal 15 seconds into the song. Very close, eh? We basically see a steep frequency cut-off just past 20kHz. According to discussions here, Like A Virgin was a multi-track 16/44 recording mixed to 14/44; so resolution-wise, it's not even of CD quality to begin with.

However, it's also quite clear that indeed the MQA process is not a "bit-perfect" reproduction. A digital subtraction reveals the difference. Here again is "Material World", first 30 seconds with the MQA and HDtracks versions aligned and run through the Audio DiffMaker program (null depth reported as 87.5dB left, 86.2dB right - for context, remember that an MP3 at 320kbps typically achieve ~60-65dB correlation depth):

Spectral-Frequency difference plot - "Material World".

Whatever the MQA algorithm is doing in the Madonna track, it obviously is very minimal in effect. This is good I suppose. It means the TIDAL/MQA Madonna sounds just like the HDtracks "studio master". You can see in the amplitude analysis that the "Maximum RMS Power" for the difference is down at around -60dB, and on average the difference is way down at -80dB or so.

We know that the Madonna track is a good example of "obviously fake" high-resolution audio. How about something less fake (ie. could have been recorded in hi-res but did the recording actually need to be?!) where there are actual frequencies >24kHz? Here's the digital difference output of the first 30 seconds of Buena Vista's "Chan Chan" (null depth reported as 73.8dB left, 70.9dB right):

Spectral Frequency difference plot - "Chan Chan".

Notice that there is greater difference this time with average RMS power in the low -60dB range. If you look at that spectral frequency plot, we see that the "difference" is actually quite low in the lower audible frequencies and increases as we get ultrasonic... We can actually plot the FFT at around 15 seconds and see this:

Difference FFT around 15 seconds into "Chan Chan".

Although differences are there in the lower frequencies below 20kHz, it's in the ultrasonic range that we see a bit more signal difference... I assume some of this could be a reflection of the "lossy" encoding of the signal above the baseband or even some imaging products above Nyquist from the upsampler?

For the sake of completeness, here's "Good Times Bad Times" (correlated null depth 76.9dB left, 74.8dB right):

Spectral-Frequency difference plot - "Good Times Bad Times".

The MQA file had an average RMS power difference of around -60dB compared to the 24/96 download. Again, if we look at the lower frequencies especially <5kHz, the difference is minimal.

Joni Mitchell's "A Case of You" did not align well on Diffmaker although the DR values looked very similar. Null correlation was reported as only around 22dB but in A/B testing it actually didn't sound too different comparing the MQA decode with my 2013 HDtracks version. Visual inspection of the waveforms suggest there is a small DC offset. Also, there seems to be mild samplerate drift based on some messing around with Diffmaker. I can't be sure but it seems like they're the same mix, but different digital transfers perhaps.

Oh yes, another thing to remind everyone - MQA does not solve the problem of highly compressed albums of course. Bruno Mars' Unorthodox Jukebox is just as peak limited as before (for this album, TIDAL/MQA decodes to 24/88):

Bruno Mars - "Locked Out of Heaven" decoded MQA, DR4 - "hi-res" indeed :-)

It always has been and will be about the quality of the recording/mastering. An encoder/decoder will always be limited by the data it's fed - garbage in will still be garbage out... (Note that I do like Bruno's music, just not the "modern" dynamic compressed mastering and sound so much!)

Since we do not have access to an MQA encoder, we cannot at this point do a full evaluation of the CODEC's capabilities with test signals. I can still say though that the non-decoded MQA file's noise floor tends to rise in the higher frequencies:

Not that big a deal since this looks like the noise we might see in a downsampled 16/48 file with noise shaped dithering from 24/96. However, remember that it's actually a 24/48 file being streamed so if we disable MQA-decoding, it'll be something like 16-bit quality while streaming in a 24-bit container. Moral of the story - keep software decoding on unless you have MQA-decoding in hardware in which case make sure that output is bitperfect and MQA is properly decoded (ie. the decoding indicator turns on). I noticed that the streaming software strips silence off the ends of the playback so I was not able to get a read on what an absolute silence noise floor looks like when encoded/decoded.

Lastly, let's have a look at the file sizes being sent over the internet. I used "Like A Virgin" for this comparison. What you'll see below is the MQA 24/48 file streamed over the internet compared to the HDtracks 24/96 "Studio Master" along with variants - downsampled to 24/48, and to 16/44 using iZotope RX 5 (with MBIT+ dithering). All tags were stripped, all converted to FLAC level 8 to keep this consistent:

As you can see, the MQA 24/48 data is about the same size as the HDtracks file converted to 24/48. However, to stream that MQA data through the Internet, TIDAL needs almost twice the bandwidth compared to a typical 16/44 stream based on the highest FLAC compression. It is of course significantly less data compared to sending the original 24/96 files which by the way is less than 70% larger despite twice the samplerate. The bottom line is that MQA does take substantially more bandwidth to stream compared to the standard TIDAL "HiFi". If you are having any buffering issues with "HiFi" already like sending the audio through marginal quality WiFi, be careful with "Master" streaming. Also worth thinking about if you have data caps especially when the TIDAL app becomes available for mobile devices over cell phone plans (again, the effective data speed might be a problem if the wireless connection varies quite a bit). My experience here with 150Mbps cable downstream speed over the home ethernet has been good; 3-4 seconds of buffering before it starts to play. Otherwise no buffering issues or stuttering thereafter.

Conclusions:

1. As shown by the Madonna and Bruno Mars tracks, there are obvious non-hi-res albums out there in the "Masters" collection (no surprise). It doesn't really make sense to be streaming MQA for those albums just as much as it makes no sense to pay good money for a 24-bit download of those files. There is materially nothing "high resolution" about them. As I said a few years back, true high-resolution demands both high quality material (hi-res recording, mixing, and master) and an ability to play the music back in an environment conducive to appreciation of the subtle difference over 16/44. Always be mindful of this. For me, truly high-resolution music is a rare cohort inhabited by only a few record labels who actually produce digital audio of that calibre (eg. AIX, 2L, Channel Classics, Reference Recordings,  Chesky at times, a number of other classical labels).

2. MQA achieves its purpose with "smart" compromises. Internet streaming bandwidth is limited (looks like they're aiming for ~1.5Mbps). Files needed to be compatible with a standard DAC. Bit-perfect replica of the "master" was never their intent (remember the "crown jewels" comment in this interview). And finally it wants to be seen seriously as a high-resolution alternative which labels can sign on for - all the talk about "original resolution" and big numbers like 192kHz and such achieves this for the purpose of advertising and optics (while not being an exact replica of the actual "studio master"). Throw in the good reputation of Meridian as a company and Bob Stuart as a respected engineer. Add in the exaggerated voodoo around the benefits of minimum phase "apodizing" filter, plus whatever "deblurring" algorithm to impress the audiophiles under the banner of being "revolutionary" and "sounds better". Finally add an indicator (pretty green / blue LED!) to show that it's working and call it "authenticated" as another selling point. All the while, the basic idea is about sacrificing absolute bit-depth to reconstruct a facsimile of the ultrasonic spectrum for the upsampling algorithm.

Folks, I can appreciate the compromises and I do believe MQA achieved what they set out to do in terms of the files sounding as good as 24/96 or 24/192+ "high resolution". Certainly there was intelligence put into this solution! However, achieving that sound quality in my opinion was never difficult nor did it require these contortions... Show me evidence that humans can successfully choose 24/192 vs. a 24/48 high quality downsample in a blind test using the same mastering. Would dithering and downsampling 24/192 to 18/96 with lossless compression not also sound good with decreased file size? This is why the technique brings to mind those overly complicated Rube Goldberg cartoons when a solution much simpler would have achieved the same sound quality, compression ratio, and compatibility.

3. If you're streaming TIDAL HiFi/Masters today without an MQA DAC, leave the software decoding of MQA turned ON. No point listening to a non-decoded MQA file with clearly inferior noise floor. To save yourself from hassles since output samplerate varies, it's best to just leave it set as "Exclusive" mode for the audio output, letting TIDAL tell the DAC what samplerate to use. Definitely make sure "Exclusive" is clicked with an MQA-capable DAC. I can imagine folks with MQA decoding hardware activating the "Passthrough MQA" box, forgetting to leave volume at 100% and not manually switching between 24/44 and 24/48 output without "Exclusive" mode thinking they're gaining benefits when it's actually not decoding properly!

4. Ultimately, subjectively, it's about how MQA "sounds". This is true of course with any lossy compression scheme - for example, it's not really just about the intellectual knowledge that MP3 is lossy that is the problem, it's whether the quality of the encoding/decoding compromised transparency for the listener. The same can be said of lossy video encoding - all DVDs and Blu-Rays and UHD Blu-Rays are lossy; but some look better than others. And in this regard, yes, TIDAL/MQA sounds great with the material I've sampled. Objectively with the songs I examined, the software decoder works well to reconstruct what looks like the equivalent 24/96 download. As for the sound itself, having both the MQA decoded files and an original 24/96 "Studio Master", I have been able to do A/B comparisons easily with the foobar ABX plugin. For me (and my 45 year old ears!), I was not able to score >70% when blinded using AKG 701 headphones with my ASUS Xonar Essence One DAC/amp for any of the 4 "mainstream music" samples I examined. This is corroborated by the "correlation null depth" measured with Audio Diffmaker where I'm seeing 70-85dB; much better than the typical 60-65dB null depths with MP3 320kbps using recent LAME encoder versions. The only exception to the high correlation null measurement was the Joni Mitchell track which sounds like it's the same mix, but I suspect it comes from a different transfer compared to the 2013 HDtracks version I have. Even with that one I was not able to ABX with impressive result. Sure, maybe if I sat down and meticulously listened for subtle differences, maybe did a bunch more A/B switching one very quiet night, nuances in certain passages may be evident... By the way, based on the findings, one would probably try to listen for slightly more high frequency "presence". I tried this with my wife in the main system downstairs for around 15 minutes before she got really bored. She was not able to consistently choose either.

But seriously, that's no way to "enjoy" music. Whatever difference there may be with the MQA encoding process (ostensibly due to "deblurring" or whatever) is subtle and not worth obsessing over based on the MQA albums I've listened to so far using the TIDAL software decoder.

By the way, I would love to run one of my "Internet Blind Tests" (like this and this and this) with all of you using these MQA-decoded files to see whether there's a statistical preference for the MQA or HDtracks version. Alas, unless permission is explicitly granted, I suspect I'd get cease and desist notices at this point or have Google block the test (like what happened last year when I criticized iFi's "USB Audio Gremlins Exposed" article). :-)

Bottom line: TIDAL/MQA streaming does sound like the equivalent 24/96 downloads based on what I have heard and the test results (not that I think much of the mainstream music out there is true high-resolution of course). I'd certainly be cautious about claims that the 24/44 or 24/48 files sound "better" when non-decoded compared to standard 16/44 CD or that the "deblurring" makes a significant difference. The real question is still whether the album itself sounds good and truly high-resolution to start off with.

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

To end off today's post, let's consider for a bit the whole idea of software decoding and think about how this all fits in with native MQA-enabled DACs. The big question of course is whether software decoding is good enough.

As I posted a few months ago, MQA is simply a partially lossy CODEC in terms of what the firmware/software decoding is about. We pretty well know there was not going to be any "revolutionary" change just from firmware update because:

1. To make it compatible with standard PCM, the "high res" portion was squished into something like an 8-bit by 44/48kHz container under the noise floor which was then supposed to fill out the final 24/88-192kHz or whatever "original resolution" for playback. Regardless of the benefit of retaining >20kHz frequencies, it's not enough to reconstitute the original recording's resolution to account for every "bit" of detail. At best it's "perceptually lossless" and if there is "deblurring" DSP applied, hopefully it will be "euphonic" and will not result in objectionable sound quality.

2. As a firmware upgrade, the algorithm cannot be too complicated! How much power do we think a Meridian Explorer 2's embedded CPU has? Or the announced MQA firmware upgrade for the Dragonfly**? Sure, they might tweak a little here and a little there, but this is nowhere on the level of DSP processing like using HQPlayer on a desktop PC/Mac/Linux for transcoded DSD128+ playback! At best, these little processors hopefully can do a decent job upsampling 48kHz input to 192kHz with a reasonable anti-imaging filter and fill in a facsimile of the ultrasonic content.

With regards to point 2, I recommend folks have a look at this thread on Computer Audiophile started by Måns Rullgård. He has been able to examine the quality of the software decoder implemented in the Bluesound Node 2 firmware! There is/was a sine sweep on page 3 of that thread suggesting suboptimal quality of the upsampler. It looks like Mans will be running the test again and I certainly hope the output from this Node 2 firmware isn't that poor [new link here to follow along with the decoder evaluation]. My assumption is that if there is only one "authentic" sound quality to MQA decoding, all devices would be expected to implement the same quality upsampling. (What a nasty conundrum if it sucks!)

[By the way, for those curious about digital filters, you can review my article from 2015 looking at the various Linux filters and objective differences.]

Remember that firmware capabilities are limited by the underlying hardware. Therefore, a device like the Bluesound Node 2, although said to have a 1GHz "multi-core ARM Cortex A9" processor (I'm guessing dual-core A9 like many of the WiFi routers out there), is a slower machine than a Raspberry Pi 3 which is a quad core 1.2GHz Cortex-A53 with newer 20nm processor architecture. I don't think it'll take a creative coder too much effort to figure out what the algorithm is doing and writing their own decoding/upsampling routine now that the software is out there. In fact, a high quality decoder done on a much more powerful computer with better upsampling algorithm and arithmetic precision would be more accurate than the wimpy processing on essentially any DAC. Decode everything to a good 24/384 DAC or even transcode to DSD128/256/512 if you want and experience the best the MQA CODEC is capable of; maybe even stream at this high-resolution to a device like a Pi around the house.

Furthermore, software decoding includes the opportunity to manipulate the sound with DSP in playback which IMO is way more useful than what MQA claims to do in the time domain by taking into account your speakers and your room - not just your DAC. This is the logical future for MQA decoding if we truly want the best possible sound quality utilizing the power and flexibility of computer audio! Understandably, Meridian/MQA will not like this idea and will fight against this unless decoders are all licensed appropriately of course.

Watching the TIDAL player software (2.1.1.160) in Windows 10 on my Intel NUC 6i5SYH with a bunch of background tasks running but not active like a browser, Skype and Twitter app, streaming MQA "Master" over WiFi only used about 13% total activity with the i5 CPU throttled down at 0.86GHz.

i5 NUC playing a software-decoded MQA stream at 24/96 to USB DAC. Probably some buffering at the start (left) followed by steady low CPU utilization.

Obviously, there are plenty of CPU cycles available if anyone wanted to use the resources of the i5 NUC, itself not a particularly powerful computer these days.

I've always said that the most interesting part of the MQA claims for me is the "deblurring" algorithm they supposedly use in the production side to make the masters sound the best they could - taking into account the recording equipment like ADC used and presumably creating a master with minimum temporal distortion. I don't know how realistic this is beyond acoustic recordings like live classical. Nor do we have indication that the potential has been utilized in these albums I examined (for all we know, these albums were run through an automated "black box" encoder). These samples sounded fine already and I'm not sure I "hear" a difference compared to standard 24/96 from HDtracks. Certainly nothing to get ecstatic about if you already have purchased high-resolution downloads before...

Let's see what the next few months bring.

With that, I think I've said enough about MQA for awhile. My curiosity is satisfied and it's time to just listen and enjoy... I see there's now talk about MQA "Core", "Renderer", "Decoder" designations to stratify the different levels of processing. Folks at MQA, I think you should keep it simple. Speaking for myself rather than as a generalization I expect anyone else to accept, after all these years of listening, measuring, and thinking about it, I could care less whether the lossy "unfolding" goes out to 192kHz or 384kHz; all smoke and mirrors as far as I am concerned since it's inaudible plus lacking in decent recordings. Furthermore, even inexpensive high-resolution DACs are remarkably accurate these days so whatever tweaking is being done for individual DACs is highly dubious and more than likely perceptually meaningless. Nonsense complexity like this especially if gushingly embraced by the audiophile press will only serve to confuse the public and remind the typical music lover what they've suspected for a long time; "extreme audiophiles" are emotionally insecure obsessive-compulsives cognitively weak in performing reality testing and prone to the use of superlatives in advertising and media. Dear readers, IMO don't be an "extreme audiophile" and end up like that :-).

Remember that TIDAL does have a free trial if you want to have a listen to these software-decoded MQA albums and decide for yourself. Hope you're all enjoying the music... It's going to be busy for me over the next couple of weeks, will check in when I can.

PS: Here's a funny article about TIDAL / MQA. The first line says "You may have never heard Madonna's Like A Virgin the way she heard it when she recorded it in the studio 33 years ago, but now you may." LOL. What a joke! It's an old sub-CD resolution 14-bit 44kHz master made from ADCs and synthesizers long since retired for crying out loud... What in the world does the author think Madonna "heard" when producing it that can somehow be resurrected using MQA 33 years later? There's some seriously poorly educated "press" out there! :-)

** Addendum on the AudioQuest Dragonfly: I see some folks are getting excited about this firmware upgrade for the Dragonfly Black and Red. Recent news seems to suggest that the update would allow these DACs to function as MQA "Renderers" which is a lower level of processing than as a "Decoder". Supposedly this means you still need the software upsampling to 88/96kHz then the DAC will perform some final corrections internally on playback. Obviously these Dragonflys are not able to accept samplerates higher than 96kHz anyway.

As you can see from this PDF of model comparisons, the newer Black and Red are using the Microchip PIC32MX embedded microcontroller. The CPU on this low cost device can only run up to 120MHz and typically have a small amount of RAM, only 32KB up to 128KB. Based on the product page, the speed even of the fastest 120MHz product is rated at only 150 Dhrystone MIPS of integer performance! Guys, that's about the speed of a first generation Pentium 120MHz from 1995 unless there's something else inside the Dragonfly.

Hmmmmmmmmmmmmm... (Notice the extra long period of contemplation.) If it is true that there is an upgrade coming for the Dragonfly, what are we anticipating this firmware to do for a signal coming in at stereo 24/96? What kind of DSP can be performed with so little computational potential? Doesn't seem to be near enough power to even upsample 96kHz to 192kHz decently much less actually provide any sophisticated signal processing to do time domain corrections! Well, let's see what kind of magic AudioQuest can pull out of this kind of hardware. I certainly would not put a bet on anything sophisticated going on unless I've neglected something major. I suspect all it might be doing is detecting the presence of an MQA decode and just sending some different settings to the ESS DAC (like changing filter parameters?). If so, this would be reminiscent of HDCD which incidentally was also released to consumers in 1995 :-).