|Image from here.|
The song selected for testing/listening was from 2L (of course) - the track "Blågutten" from the Hoff Ensemble's album Quiet Winter Nights (2011, MQA downloadable here). The recording ADC device is the excellent professional quality RME Fireface 802.
Of course re-digitizing degrades the signal very slightly but the Fireface is an excellent ADC so whatever slight change is likely very minimal though arguably it'll add some "blurring" to the signal if we accept Meridian's claims.
Based on my friend's tech notes, the Explorer2 puts out a peak 2.1Vrms and he recorded this signal with the Fireface 802 set at +4dBu allowing for -5dBFS headroom. No further volume adjustment was made between the PCM playback and MQA. We can see this when we analyze the amplitude statistics for the track:
We see that both the 24/192 PCM and MQA decode are almost exactly the same peak and average amplitudes (<<1dB difference). This is good, it means the MQA process isn't changing the output volume making comparisons like A-B testing possible without needing to do anything extra (ie. volume compensation).
What is most easy to see and as I had wondered about is the loss of resolution in the dynamic range due to the compression / "encapsulation" / "origami" process. In this sample, if we look at the noise level at very close to the same (within milliseconds) place in the song near the end as it fades out, this is what we see:
This is the Meridian Explorer2 playing the 24/192 PCM version (ie. the non-MQA version) of the track. Notice the excellent noise floor up to ~45kHz (BTW this is just with Adobe Audition 3, FFT size 8192). After this, there is a gradual rise in the noise floor which I believe is modulator noise primarily from the original recording although the Fireface ADC might be adding to it. The excellence of the noise level clearly speaks to the superb recording quality from 2L and their ability to capture >16-bits of digital resolution.
What does this look like out of the Explorer2 with MQA decoding?
What you see above is clearly a rise in the absolute noise floor as expected because the lowest bits of the 24-bit data is being used for the "encapsulation". That noise floor in fact doesn't get much better than what you see above suggesting that indeed the original 24-bit signal has been reduce and probably a fair amount of dithering noise applied (looks to me at best to be 16-bits with a fair amount, maybe 2 bits, of dithering).
Out of curiosity, I ran another ABX test to see if I could hear the difference, the foobar ABX 2.0 tool was used and here's the result:
Compared to last time when I tried this with the 16/44 source file, it looks like I was better able to identify the difference this time! As before, I used my desktop firmware upgraded ASUS Essence One with Sennheiser HD800 headphones in a very quiet room using a silent fanless computer setup. I was honestly a bit surprised at the result because I certainly did not feel that confident at the end of the ABX session. However, I did try to be more deliberate and "serious" than last time and took a number of minutes at the start to listen between the two files and pick a place to start where I thought I could familiarize myself with the slightly different sound. Like before, the trick for me was to listen to the tonality of the two files with instantaneous switching. Again, the telling component was that the MQA decoded recording did sound a little "clearer" and more "sparkly" - perhaps more accented with the percussion hits and cymbal trails. As previously mentioned, it was similar to a TV where the rendering seemed "sharper" in terms of local contrast enhancement with the MQA file. I think one could say there seemed to be more "air", maybe like adding a little 10kHz boost kind of effect. As I had expressed before, I'm not sure whether this subjective effect would be universally considered to be "better". The PCM 24/192 sounded very smooth and natural already.
Note that even though I was aware that the MQA decoded file had a higher noise floor, I did not purposely take advantage of this by listening to a quiet portion of the file with volume pumped up to listen to the hiss. That would be cheating :-).
Another way to look at the difference between the two files is using the Audio DiffMaker program to create a "difference" file. So, running the decoded MQA and subtracting the PCM playback through it, we get a pretty good ~74dB average correlated null depth using the first 15 seconds of the audio (with compensation for sample rate drift from the ADC):
If we look at the "difference" FFT, at any one instance, it looks like this:
I can hear the musical residue but shifted to a higher pitch when playing the "difference" file. If I pump up the volume, the noise tonally is higher with a rising noise level above 5kHz as shown in the FFT. Perhaps this accounts for the tonality change I heard when ABX'ing. If we show the "spectral frequency display" using Audition 3:
|Spectral frequency display. Ignore the signal in the first 200ms - likely just DiffMaker not quite achieving the time compensation yet as it aligns the samples.|
|Amplitude statistics of the "difference" file (first 250ms discounted from analysis)... Although on average the difference is low, down at -70dB RMS, there were differences up to -30dB peaks.|
Conclusions:I think what is interesting about the listening experiment is that the difference was noticeable even with the ADC (Fireface) / DAC (Essence One) steps in between. This tells me that the "de-blur" DSP processing implemented by Meridian/MQA can in fact be captured through the analogue output though presumably the effect may not be as strong as listening directly (this is what my friend has noticed). Nonetheless, I can reasonably prove to myself that I was able to hear the difference in ABX testing.
Bottom line is this... So far, what I have seen of decoded MQA remains consistent with suspicions I had previously.
1. First, there is a "de-blur" DSP being applied which supposedly improves time-domain accuracy (presumably based on measured parameters of the equipment / processing steps used in the studio). I think this is the primary audible component. This is why they claim even an undecoded MQA file played through a standard DAC sounds "better"; the DSP effect is already "baked into" the sound even without the decoding process to expand the ultrasonics.
2. Second, the "encapsulation" process takes over the lower bits in the 24-bit data for data "compression". Hence we see a higher noise floor in the MQA decoded file. Based on the sample in this post, it still looks like MQA is not capable of >16-bit dynamic range when decoded.
3. Thirdly, a decoding algorithm reconstitutes the ultrasonic frequencies for final playback on the DAC to 192kHz or higher if the DAC is capable. I assume for the Explorer2, Meridian would be trying for their typical apodizing minimum phase digital filter. There is of course no way to know for sure unless we had access to tools to encode test signals. I doubt if this makes any audible difference in any case. The example here basically shows a bunch of noise added to the ultrasonic frequencies even in places where there wasn't any noise in the original 24/192 PCM file. Clearly the decoded MQA file is short of a true "lossless" reconstruction!
Given that I believe Step 1 above is the most likely to change sonic quality, I really wonder what standard 24/96 or 24/192 "digital masters" sound like with the DSP applied to "correct" time-domain parameters in the studio before "encapsulation" takes out the lowest bits.
Realizing that Meridian/MQA has provided essentially no technical details or objective results, if I am correct about what is going on as described above, I am personally not interested in MQA as a format I feel I would want. There’s no “magic” here and there are evident compromises when trying to be everything to everyone as MQA seems to be aiming for. It's aimed at data compression, being “compatible” with regular DACs, able to deliver “high resolution”, and “better sounding” even with regular 16/44. I wish Meridian/MQA success in presumably the target niche (high-resolution streaming like TIDAL) even though I think there are better ways to do this (ie. simply stream 24/48, or something like FLAC compressed 18/96 as Miska and others have discussed).
I honestly hope no studio implements this as some kind of archival alternative nor does this supplant digital downloads where true 24-bits and high samplerate are already within the grasp of consumers through places like HDTracks, Pono, etc... A "flat", standard PCM 24/88+ is compatible with essentially any DAC and capable of full sonic quality for one's music library already if one believes there is a need to even go beyond 16/44. As is often the case, when something tries to be the "jack of all trades", generally "master of none" is a typical outcome.
A big thank you to my friend with the Explorer2 and studio quality ADC for taking the time to do this and being so meticulous with the whole recording process!
[One last thing: Meridian/MQA - please stop using the description "revolutionary" for this product. I'm quite sure there are many of us who cringe every time we hear this...]
Have a great week ahead everyone! Enjoy the music...
Well, well, well... It looks like maybe the accolades for MQA isn't universal after all! Here's PS Audio's Paul McGowan's impression.
For the purpose of further clarification of what I'm seeing in the noise floor, here's the composite image of the various releases of the same track analyzed at the same place (3:49.000 into the track) with DSD64 and undecoded MQA played back using an ESS9018 DAC, compared to the Meridian Explorer2 playing back the "studio master" 24/192 PCM and the MQA track but this time decoded. Each sample recorded with the same RME Fireface ADC.
Though we can perhaps look at other samples, it's quite clear from this example that by 10kHz, the decoded MQA signal is unable to achieve the same resolution as the 24-bit PCM; in fact, not surprisingly, it follows the undecoded MQA which itself is subject to digital filter parameters for 44kHz playback. In comparison, we can look at DSD64 (SACD resolution) which follows 24/192 until about 24kHz then deviates with the large amount of ultrasonic grunge from noise shaping.
BTW: The reference was set at -20dB so as to show the relative noise floors a bit better (as I circled at the bottom of the graph).