It’s reasonable to be concerned about MQA. It’s a big deal. There’s already much support from record labels and DAC manufacturers. It’s clear to me that MQA’s developers see it as an idealistic venture designed to fix what digital broke, sound-wise and within the music ecosystem. Others see things differently (footnote 1).
My goal for this series of articles, of which this is the first, is to subject MQA to a fair and thorough vettingnot as an expert, but as a science and technical writer. My role is not to make absolute judgments, but to do the hard work, struggling through dense technical articles, pestering people with questions, evaluating evidence in consultation with experts, and assembling what I learn into something coherent and accessible. I’ll present the evidence, and people can then decide for themselves.
This is a complex business, with far too much to look into all at once with any sort of rigor. So I’ll take on the issues one at a time, beginning here with an aspect of MQA’s time-domain behavior: its decoder/renderer’s impulse response.
A few months after my first request, Bob Stuart made available an MQA-encoded file containing a train of perfect impulses. Later, he sent me a non-encoded 24-bit/96kHz FLAC file containing exactly the same information, so that I could compare MQA’s performance directly with the performance of non-MQA DACs, on even terms.
An impulse is a very short signalthe shortest possible signal, in factso it’s tempting to think of a test of an audio system’s impulse response as a test of its response to very short signals. An impulse-response test is that, but because an impulse contains all the frequenciesfor band-limited systems, all the in-band frequenciesit’s a useful and commonly used measure of a system’s overall fidelity. The more closely an output impulse resembles the input impulse, the truer the output will be for any input (footnote 2).
MQA is, as Bob Stuart likes to say, an end-to-end technology: analog in to analog out. This test, though, starts in the middle: It skips the “analog in” stage by sending a digitally manufactured test signal directly to the DAC. We’re skipping half the MQA processthe encoding part, which Stuart says is responsible for some 70% of MQA’s claimed improvement in sound quality. The part we’re testingthe “renderer” (footnote 3)contributes about 10% to MQA’s performance, Stuart told me.
I measured an assortment of DACs that I have on hand. I recorded their analog outputs at 24/192kHz, so each sample is a little more than five microseconds (5µs) wide. I’ve expanded the view so that you can see the individual samplesthe little magenta dots. The major horizontal divisions are 100µs apart. To make comparisons straightforward, I’ve used the same vertical scale for all the plots.
John Siau, who designs the DACs made by Benchmark Media Systems, focuses on maintaining the signal’s frequency-domain integrity. This is why Benchmark’s DAC3 HGC includes a linear-phase reconstruction filter that rolls off very quickly in the frequency domain, and so produces a good bit of time-domain ringing (fig.1). This is typical of a linear-phase filter in that the ringing is symmetricit comes both before and after the main peak. One of the key notions on which MQA is based is that our ear/brain system regards pre-ringing as unnaturaland there’s plenty of it here. And yet, the DAC3 HGC is a brilliant-sounding DAC.
Fig.1 Benchmark DAC3 HGC, impulse response (one sample at 0dBFS, 96kHz sampling, 100µs/horizontal div.).
Next up is Mytek HiFi’s Brooklyn DAC, with its reconstruction filter set to Minimum Phase (fig.2). Again there’s lots of ringing, but it comes after the main pulseno pre-ringingand the post-ringing would normally be buried beneath the music’s reverberation.
Fig.2 Mytek HiFi Brooklyn, Minimum Phase Filter, impulse response (one sample at 0dBFS, 96kHz sampling, 100µs/horizontal div.).
Fig.3 shows the Brooklyn again, now with its slow-rolloff filter selected. This shows what you can accomplish in the time domain by using a filter that rolls off slowly in the frequency domain. The response is very short, but it’s still linear-phase (whether or how much this matters isn’t clear), with just a little pre-ringingabout 20µs total. That’s not much.
Fig.3 Mytek HiFi Brooklyn, Slow Roll-Off Filter, impulse response (one sample at 0dBFS, 96kHz sampling, 100µs/horizontal div.).
For the next tests, I sent the same data to the DACs, this time MQA-encoded. First I sent it to a non-MQA DAC, so that you can see what that looks like (fig.4). We’re now in the 48kHz domain, not 96kHz, so we expect a wider impulse. This response is mostly linear-phase, though the asymmetry suggests some nonlinearity in the phase response. The details of the response will depend on the DAC’s particular filter.
Fig.4 Benchmark DAC3 HGC, impulse response (one sample at 0dBFS, MQA-encoded, 48kHz sampling, 100µs/horizontal div.).
Fig.5 shows MQA proper via the Mytek Brooklyn DAC with MQA enabled, though the response should be the same for any MQA-enabled DAC. This is nearly ideal: There’s no pre-ringing, and the response is fast and shortclear evidence of MQA’s time-domain excellence, though the Brooklyn’s slow-rolloff, linear-phase response had very similar width and only a small amount of pre-ringing. Would such a small difference be audible?
Fig.5 Mytek Brooklyn DAC with MQA enabled, impulse response (one sample at 0dBFS, MQA-encoded, 48kHz sampling, unfolded to 96kHz, 100µs/horizontal div.).
Here’s a surpriseor it would be surprising, if there hadn’t been hints in John Atkinson’s measurements over the last couple of years: I’ve sent the PCM impulse filenot the MQA fileto the Brooklyn DAC with its MQA decoding turned on (fig.6). Same thing, right? Looks like it to me. Apparently, as long as the MQA decoder is enabled, the impulse response is basically the sameeven for non-MQA data. Stuart explained to me that, in some implementations of MQA, when MQA decoding is enabled, all data are sent to the DAC’s MQA module, which detects the file type and then does the right thing. In DACs that are built this way, including the Brooklyn, even non-MQA music is sent to MQA’s upsampling renderer. Don’t want MQA messing with your regular PCM data? Turn it off (footnote 4).
Fig.6 Mytek Brooklyn DAC with MQA enabled, non-MQA impulse response (one sample at 0dBFS, 96kHz sampling, 100µs/horizontal div.).
It’s important to consider what fig.6 doesn’t show. This is not MQA’s claimed deblurring. Deblurring, per MQA, is the removal of time-domain artifacts remaining from previous analog/digital conversions; here there are no artifacts, since this test file was built and delivered in the digital domain. I hope to find a way to demonstrate and test deblurringhow MQA handles imperfect filesfor a future article.
One of the challenges levied against MQA by its more knowledgeable critics is that while MQA’s approach may improve the shape of the impulse response, its sampling methodand the resulting, presumed increase in aliasingintroduce randomness in precisely when those impulses occur. If they’re right, this would offset any claim of time-domain advantage. I synchronized the MQA and non-MQA impulse responses: MQA in the left channel, non-MQA in the right. Over 30 seconds of impulses spaced 0.7ms apart, examined on a microsecond scale, I saw no random offsetsor offsets of any kindin where MQA’s impulses landed.
This is just one small piece of a large puzzle, but it’s a start. MQA’s filterthe one that in non-MQA DACs is called the reconstruction filteris apparently very well behaved in the time domain (footnote 5).
Next time: Sure, MQA’s compression has a lossy aspectbut how much does that really matter?
Footnote 1: See my “As We See It” in this issue.John Atkinson
Footnote 2: Strictly speaking, the signal we use to test DAC impulse responses is “illegal,” in that it violates the Nyquist/Shannon requirement for the signal to be band-limited to half the sample rate.John Atkinson
Footnote 3: The core decoder is in the circuit for the MQA test file, but other than routine unpacking, there isn’t much for it to do in this case. Jim Austin
Footnote 4: But see the review of the Aurender A10 elsewhere in this issue, where for non-MQA, regular-PCM files stored on its internal drive, the MQA filter can’t be turned off.John Atkinson
Footnote 5: However, as my measurements have shown, this filter is “leaky” in the frequency domain.John Atkinson