MQA has once again floated to the surface of the perfectionist-audio pond—not belly-up as some have hoped but forced there by relentless pursuit by anti-MQA predators posing as impartial jellyfish.
In case you’ve been living amidst the seaweed at the bottom of the pond, MQA is a thing that’s done to digital audio data—sort of a codec that is said to reduce the “blurring” repeated digital conversions cause. For high-resolution audio, MQA compresses data into a much smaller file or stream while retaining, it is said, the sonic benefits of the original high-rez audio (and also of “deblurring”). MQA appears to be genuinely clever and legitimately new, implementing post-Shannon developments in sampling theory that have not previously been applied to digital audio.
MQA is controversial, for good reasons. In the interest of making the sound better, it alters the sound the mastering engineer and musicians signed off on. It would replace open standards with proprietary ones—a big step backward, especially for internet-libertarian types who think information should be free (a perspective I respect but don’t fully share).
On the positive side, MQA partisans (including esteemed recording engineers Peter McGrath and George Massenburg, multi–Grammy Award–winning mastering engineer Bob Ludwig, and hi-fi and music critic Jason Victor Serinus of Stereophile) love how MQA sounds. What’s more, MQA takes small steps toward addressing provenance issues in digital audio (although in this respect it does not go far enough).
This new attack on MQA comes from “GoldenSound” on YouTube (footnote 1). His findings may seem damning, but there’s less than meets the eye. I am not a partisan, for or against MQA. If I’m partial to anything, it’s fairness, and GoldenSound’s critique is unfair. MQA isn’t perfect, but here it has been falsely maligned.
“GoldenSound” managed to get two music files accepted by Tidal and encoded in MQA. (Kudos; if I had known this was possible, I’d have done it myself a long time ago.) He presents results from two tracks: one with a sampling rate of 44.1kHz, the other at 88.2kHz. Inside these tracks, he embedded an impulse, white noise, a squarewave, a 32-tone test signal, “the entire RightMark Audio Analyzer (RMAA) test sequence,” and a 1kHz sinewave at –60dBFS. The 88.2kHz file added two ultrasonic sinewaves (with summed amplitude very close to full-scale) and a series of full-range sweeps (right up to the Nyquist frequency, at levels that appear to be about –6dBFS).
Let’s talk first about the high-rez (88.2kHz) data, since that’s the easiest to dismiss. The 88.2kHz file embeds high levels of ultrasonic information, including those two sinewaves, the full-range sweeps, squarewaves, and white noise right up to 44.1kHz.
Ever since its announcement in October 2014, Bob Stuart and the MQA team have been clear that MQA is intended for music. It is not intended to deal with high levels of ultrasonic information (which are not found in music)—certainly nothing close to full-scale. Also well-established—nothing new—is that MQA by design allows more aliasing than other methods. It’s hardly surprising that, in a file containing so much ultrasonic information, high-level aliased signals appear in the undecoded analog signal in the audible range. The high levels of noise seen in GoldenSound’s data, too, would be lower with data that more closely resembles actual music. (Apparently realizing this too late, GoldenSound submitted a third file with far less ultrasonic content, but by this time MQA and Tidal had grown suspicious; GoldenSound’s original files had been removed, and the upload failed.)
What about the 44.1kHz data? GoldenSound’s test-file–embedded music tracks still do not resemble real music. The test file has very high levels of information at the very top of and above the audible range—right up to the 22.05kHz Nyquist frequency. In the test spectrum shown in the video, at 6:30, you can see the white noise going right up to 22.05kHz at what appears to be about –6dBFS. Parts of the RMAA test include full-scale information, also up to the Nyquist frequency. The squarewave’s harmonics do the same, although with decreasing intensity as frequency increases, as squarewave harmonics do. Full-scale data at the Nyquist frequency is not something MQA, which is intended to optimize time-domain behavior and so is very careful about what it does near Nyquist, is intended to deal with; in contrast, a conventional DAC that employs a traditional brickwall filter would have little trouble with such data, although it would do the usual damage (if damage it be) in the time domain. Perhaps there is real, electronic music in the world that pegs the meter near 20kHz, but I wouldn’t want to listen to it.
GoldenSound’s –60dB, 1kHz sinewave is a low-level signal that is intended to test dynamic range. The result in his test file was high-frequency noise that was “only” 45dB below the signal level. This noise level is probably higher because of the issues I’ve already discussed, but even so, high-frequency noise 45dB below a –60dB signal—itself barely audible—won’t be audible at all. (The much higher level of high-frequency noise in the 88.2kHz track was surely a result of the excessive levels of ultrasonic information in the file.) What GoldenSound’s tests show, then, is that a bull in the china shop can damage the china. The solution is not to throw out the china but to keep out the bull.
GoldenSound’s tests are a missed opportunity.
Footnote 1: MQA’s (the company’s) rebuttal can be found at here Also see the excellent comments from Pablo Banñados of Santiago, Chile (“Mieswall”), here and elsewhere in the same thread.
Published at Sun, 06 Jun 2021 17:51:02 +0000