- Joined
- Mar 6, 2021
- Messages
- 214
- Likes
- 289
As a newer member of ASR, I've learned a ton from reading posts on a variety of topics that use relevant empirical data in order to cut through the snake oil of audio marketing. This is a great service to audiophiles, a way to provide some ground truth that can focus our energy on getting the best sound quality at a reasonable cost without being distracted by all the gimmicks. Special thanks to @amirm for doing this (I know that you take a lot of flack from people for bursting their bubbles, I completely understand and empathize).
That being said, I've been a bit frustrated by some of the discussions surrounding MQA on this forum, and I've been distracted from ground truth a couple times and followed discussions down a tin foil hat rabbit hole to nowhere. We live in distrustful times, and people are often too quick to question or judge the motives of others, devolving into ad hominem arguments that distract us from the empirical facts. There are many claims about MQA encoded audio, both from MQA and from critics, but the only way to get hard facts is to make measurements...followed up by thoughtful analysis that focuses on those facts alone and doesn't jump to conclusions that aren't supported by data.
I am mostly concerned about the practical end product, which is the sound that actually comes out of my speakers and into my ears. We need to measure this by recording what comes out of the speakers directly, in as uniform a manner as possible. And I would like to encourage other people to do the same thing. When we say that there is "no audible difference" we should back up such claims with empirical data instead of hypothetical arguments. When we say that there is a difference, then let's see the measurement to prove it.
To this end, I did a simple experiment: I compared recorded playback through my loudspeaker system of MQA tracks streamed through the Tidal App vs. CD-derived (AIFF 16/44.1) tracks played using the latest version of Audirvana using its standard (factor of 2x) upsampling algorithms. The audio system is simple: MacBook Pro -> USB -> Topping D90 MQA -> March Audio P252 -> LS50 Meta speakers. The D90 roll-off filter is in mode 3 with typically -20 dB attenuation. The recording is made by a Sony PCM-D100 at 24/192 mounted on a heavy tripod at speaker level with the stereo microphones individually aimed at the speakers' focus points.
All components are kept in the same positions and settings as I play back different tracks. This means that, even if the measurement equipment is imperfect, correlated artifacts and imperfections will cancel out in the relative comparison. In terms of an equation:
Recording_1 = Signal_1 + Correlated_Imperfections_1 + Uncorrelated_Imperfections_1
Recording_2 = Signal_2 + Correlated_Imperfections_2 + Uncorrelated_Imperfections_2
where I define:
Correlated_Imperfections_1 = Correlated_Imperfections_2.
If I subtract these first two equations and use the third to simplify,
Recording_1-Recording_2 = Signal_1 - Signal_2 + (Uncorrelated_Imperfections_1 - Uncorrelated_Imperfections_2)
then the correlated imperfections go away. Correlated imperfections include any artifacts in the sonic character or quality of the entire chain of playback equipment down to the speaker itself. Uncorrelated imperfections, on the other hand, are artifacts that are not the same between recordings, such as stray sounds or noises in the room that appear in one recording but not the other. There are also ways to deal with uncorrelated imperfections, if they cannot be adequately controlled. If one assumes that these are normally distributed about a null mean then we can take many recordings of each signal and then average them together, in which case the amplitude of the uncorrelated imperfections decreases inversely proportional to the square root of the number of averaged samples.
I searched for a variety of tracks for which I have quality CD rips and which also stream as "Master" quality on Tidal and are explicitly recognized as MQA on the display of the D90 during playback. There were plenty of examples from my library (mostly rock music), however, subjectively I could not hear any differences in the tracks I compared. I opened the comparison recordings in Adobe Audition, and aligned the stereo tracks with one another for direct comparison. Here is an example excerpt from the track "Starshine" by Gorillaz, which has a lot of texture in the high frequencies and for which I expected to see some differences, if they exist. The top pair of waveforms (purple-ish) are the MQA/Tidal versions, and the bottom pair (orange-ish) are the CD quality/Audirvana versions:
The data seems to indicate that, apart from slight over all amplitude differences, the sound that I recorded from both versions is visually the same. At least upon visual inspection, the waveforms appear to be well-matched. (I showed some excerpts to my wife, a seismologist who studies wiggly lines and sniffs out differences, and she had the same impression.)
I scaled and overlaid some recordings on top of one another to see if any difference shows up. Here is an example from the above excerpt:
If you squint, you can see some very subtle differences in the top channel (less so for the bottom). However, it isn't clear if this is due to interpolation anomalies from the Adobe Audition display (which draws the solid line), caused by uncorrelated imperfections, or if it reveals some very slight differences in the MQA vs CD renderings.
I looked all through this track, and I found the same characteristics everywhere I looked. I've also looked at some other tracks (mostly RHCP tunes that have clean CD rips). So far, it looks as if playback (through my system) of Tidal "Master" MQA tracks and upsampled CD quality played through Audirvana are very close to identical.
Caveats: This work so far is done using visual comparisons of the waveforms, I've only looked for obvious differences and I might have missed more subtle cues. Also, any inferences made from this data are only relevant to my specific system and recorder set-up, as well as to the tracks I've tested. Other systems and tracks might reveal differences that my system does not.
Future work: Quantitative analysis is merited, cross-correlation to exactly align the pairs of recordings and normalization of the amplitudes. Once this step is completed, the aligned and scaled tracks can be subtracted from one another to reveal their measured differences. The difference can then be analyzed in order to see if it arises from actual differences in the tracks, or if it smells like uncorrelated noise. This is an obvious next step...it doesn't seem that Audition has the capability to do this, so I may need to write some code and run the analysis myself, which takes a little time (unless somebody else has some suggestions for software that performs this task).
Practical Conclusion: I cannot tell any difference subjectively and no compelling indications (yet) objectively that MQA and upsampled CD-quality tracks sound any different in my system. In this sense, at least for me personally, MQA is no better, nor is it any worse, than listening to CD quality tracks through a quality player.
I look forward to any comments on how to improve my method, if anyone sees any flaws or pitfalls, and discussion of the measurements (and techniques).
That being said, I've been a bit frustrated by some of the discussions surrounding MQA on this forum, and I've been distracted from ground truth a couple times and followed discussions down a tin foil hat rabbit hole to nowhere. We live in distrustful times, and people are often too quick to question or judge the motives of others, devolving into ad hominem arguments that distract us from the empirical facts. There are many claims about MQA encoded audio, both from MQA and from critics, but the only way to get hard facts is to make measurements...followed up by thoughtful analysis that focuses on those facts alone and doesn't jump to conclusions that aren't supported by data.
I am mostly concerned about the practical end product, which is the sound that actually comes out of my speakers and into my ears. We need to measure this by recording what comes out of the speakers directly, in as uniform a manner as possible. And I would like to encourage other people to do the same thing. When we say that there is "no audible difference" we should back up such claims with empirical data instead of hypothetical arguments. When we say that there is a difference, then let's see the measurement to prove it.
To this end, I did a simple experiment: I compared recorded playback through my loudspeaker system of MQA tracks streamed through the Tidal App vs. CD-derived (AIFF 16/44.1) tracks played using the latest version of Audirvana using its standard (factor of 2x) upsampling algorithms. The audio system is simple: MacBook Pro -> USB -> Topping D90 MQA -> March Audio P252 -> LS50 Meta speakers. The D90 roll-off filter is in mode 3 with typically -20 dB attenuation. The recording is made by a Sony PCM-D100 at 24/192 mounted on a heavy tripod at speaker level with the stereo microphones individually aimed at the speakers' focus points.
All components are kept in the same positions and settings as I play back different tracks. This means that, even if the measurement equipment is imperfect, correlated artifacts and imperfections will cancel out in the relative comparison. In terms of an equation:
Recording_1 = Signal_1 + Correlated_Imperfections_1 + Uncorrelated_Imperfections_1
Recording_2 = Signal_2 + Correlated_Imperfections_2 + Uncorrelated_Imperfections_2
where I define:
Correlated_Imperfections_1 = Correlated_Imperfections_2.
If I subtract these first two equations and use the third to simplify,
Recording_1-Recording_2 = Signal_1 - Signal_2 + (Uncorrelated_Imperfections_1 - Uncorrelated_Imperfections_2)
then the correlated imperfections go away. Correlated imperfections include any artifacts in the sonic character or quality of the entire chain of playback equipment down to the speaker itself. Uncorrelated imperfections, on the other hand, are artifacts that are not the same between recordings, such as stray sounds or noises in the room that appear in one recording but not the other. There are also ways to deal with uncorrelated imperfections, if they cannot be adequately controlled. If one assumes that these are normally distributed about a null mean then we can take many recordings of each signal and then average them together, in which case the amplitude of the uncorrelated imperfections decreases inversely proportional to the square root of the number of averaged samples.
I searched for a variety of tracks for which I have quality CD rips and which also stream as "Master" quality on Tidal and are explicitly recognized as MQA on the display of the D90 during playback. There were plenty of examples from my library (mostly rock music), however, subjectively I could not hear any differences in the tracks I compared. I opened the comparison recordings in Adobe Audition, and aligned the stereo tracks with one another for direct comparison. Here is an example excerpt from the track "Starshine" by Gorillaz, which has a lot of texture in the high frequencies and for which I expected to see some differences, if they exist. The top pair of waveforms (purple-ish) are the MQA/Tidal versions, and the bottom pair (orange-ish) are the CD quality/Audirvana versions:
The data seems to indicate that, apart from slight over all amplitude differences, the sound that I recorded from both versions is visually the same. At least upon visual inspection, the waveforms appear to be well-matched. (I showed some excerpts to my wife, a seismologist who studies wiggly lines and sniffs out differences, and she had the same impression.)
I scaled and overlaid some recordings on top of one another to see if any difference shows up. Here is an example from the above excerpt:
If you squint, you can see some very subtle differences in the top channel (less so for the bottom). However, it isn't clear if this is due to interpolation anomalies from the Adobe Audition display (which draws the solid line), caused by uncorrelated imperfections, or if it reveals some very slight differences in the MQA vs CD renderings.
I looked all through this track, and I found the same characteristics everywhere I looked. I've also looked at some other tracks (mostly RHCP tunes that have clean CD rips). So far, it looks as if playback (through my system) of Tidal "Master" MQA tracks and upsampled CD quality played through Audirvana are very close to identical.
Caveats: This work so far is done using visual comparisons of the waveforms, I've only looked for obvious differences and I might have missed more subtle cues. Also, any inferences made from this data are only relevant to my specific system and recorder set-up, as well as to the tracks I've tested. Other systems and tracks might reveal differences that my system does not.
Future work: Quantitative analysis is merited, cross-correlation to exactly align the pairs of recordings and normalization of the amplitudes. Once this step is completed, the aligned and scaled tracks can be subtracted from one another to reveal their measured differences. The difference can then be analyzed in order to see if it arises from actual differences in the tracks, or if it smells like uncorrelated noise. This is an obvious next step...it doesn't seem that Audition has the capability to do this, so I may need to write some code and run the analysis myself, which takes a little time (unless somebody else has some suggestions for software that performs this task).
Practical Conclusion: I cannot tell any difference subjectively and no compelling indications (yet) objectively that MQA and upsampled CD-quality tracks sound any different in my system. In this sense, at least for me personally, MQA is no better, nor is it any worse, than listening to CD quality tracks through a quality player.
I look forward to any comments on how to improve my method, if anyone sees any flaws or pitfalls, and discussion of the measurements (and techniques).