Audio Asylum Thread Printer
Get a view of an entire thread on one page
|For Sale Ads|
In Reply to: RE: Charles, Please tell me if this makes sense. posted by Jim Austin on May 31, 2017 at 20:21:45
>> This is a 44.1/24 recording. The gap seems centered on the Nyquist frequency. <<
I'm sorry to say that I don't know if this makes sense or not. It seems many of us are looking for additional information regarding the entire MQA encode/decode process, and this could be a helpful clue. I'm assuming that the light green trace is the CD - so far, so good. Then if the magenta trace is the MQA version, questions start to arise. All of the examples published by MQA I've seen only seem to address files that are multiples of 48kHz. I am as yet unclear on is how MQA handles files recorded at multiples of 44.1kHz.
Andreas Koch of Playback Design says in an a YouTube interview that the MQA process sample-rate converts files recorded at multiples of 44.1kHz to multiples of 48kHz (scroll forward to 6:25):
This seems to be supported by the patent application that can be found on the Benchmark Audio website, which shows a 96kHz input signal being fed to a noise shaper that separates the signal multiple times, both into high- and low-frequencies and also into upper- (first 17 and later 13) and lower-bits (included are links to the original MQA patent applications):
The next point that is unclear about the graph you posted is if the MQA version is supposed to represent the original 44.1/24 file or a higher resolution 88.2/24 file. If the former, it raises several questions, such as "What is the signal above the Nyquist supposed to represent?" and "Why not just encode it directly to FLAC for universal compatibility and a smaller file size?" If the latter, then other questions arise such as, "Does the MQA FLAC wrapper play back at 44.1kHz or 48kHz?" (you should be able to tell by the display on your DAC) and "What happened to the audio signal in the notch?" (a non-MQA processed 88.2/24 file would have no such notch).
At this point it appears that there is a bug in the MQA "unfolding" process, which is surprising to me. One would think that they have had both the resources and the time to work all of the bugs out. Yet the evidence seems to indicate that working with files based on multiples of 44.1kHz are not processed correctly, either in the encoding ("folding") or decoding ("unfolding") side. If in fact the audio is sample-rate converted, that could explain some of the variable reactions to different MQA titles. As Andreas Koch points out, sample-rate conversion is a lossy process - all of the original data is discarded and replaced by interpolated replacements.
I personally have never heard a sample-rate conversion that was transparent. The first time I ran into this was listening to a music video DVD with the soundtrack recorded at 48kHz. I liked the music so much that I bought the CD, apparently made by sample-rate converting to 44.1kHz. The sound of the CD was distinctly worse than the DVD even when played on the identical system, exhibiting a coarse and unpleasant graininness. I've heard similar effects from outboard sample-rate conversion boxes when making non-integer rate conversions (eg, 44.1kHz to 96kHz instead of 88.2kHz). In my experience, the less one fiddles with the raw data, the better the final results.
As always these posts reflect only my opinions and not necessarily those of my employer or fellow inmates.
A curiosity: When "first-unfold"ed in software, by Tidal, my Mytek Brooklyn indicates 88.2/24. When fully unfolded in hardware, by the same DAC, I see 44.1/24. Not sure what to make of that.
> > When fully unfolded in hardware, by the same DAC, I see 44.1/24. < <
My guess is that the DAC is showing the sample rate of the (singly "folded") incoming file. It started out as an 88.2/24 file and was "folded" during MQA encoding by taking the audio data in the frequency range from 22.05kHz to 44.1kHz, discarding its lower bits, compressing with lossless techniques, and then storing it in the lower bits of the "baseband" audio (from DC to 22.05kHz). This process discards the lower bits in the baseband (which is what allows for the reduced file size of MQA for high-res streams).
The lower bits in the dual-rate data are also discarded for two reasons. One is that the now lowered dynamic range signal can be compressed into a smaller stream that requires discarding fewer LSBs in the baseband audio, and the second is that there is no reason to have a lower noise floor in the dual-rate audio compared with the baseband audio. (Noise-shaped dither does exactly the opposite, *raising* the noise floor at high frequencies.)
Inside the DAC, the lower bits in the baseband are separated out, uncompressed to re-form the restricted dynamic range version of the dual-rate data (the discarded bits can never be retrieved), and spliced back to the baseband audio to re-create an 88.2/17 file.
When you set the Tidal app to perform the MQA decoding, it sends an 88.2/17 audio stream (possibly with zero padding to create 24-bit words). The DAC then displays 88.2/24.
The bug is that when the two frequency ranges are spliced back together with the Tidal software (as seen in the YouTube video), there is audio data missing from the original 88.2/24 file in the frequency range between 21kHz and 23kHz. What we don't know yet is if this is also true when decoding is performed in hardware. If you have Audacity, Adobe Audition, or some similar audio tools, you may be able to answer that question.
> > My guess is that the DAC is showing the sample rate of the (singly "folded") incoming file. It started out as an 88.2/24 file < <
I don't think so, for two reasons. First, the little blue light came on on the Mytek Brooklyn; on every other file I've ever played (and paid attention to), that's accompanied (on the display just to the left) by the fully unfolded (original "master") resolution--usually 24/96 or 24/192; this time it said 24/44.1. (Yes, it says the same thing when I turn off MQA decoding.) Second, I read somewhere--I cannot recall where--that the master is 24/44.1. I know that doesn't inspire a lot of confidence, but I'm pretty certain I'm right. Maybe someone else can provide a source for that information? Normally you'd just be able to locate a download (HD Tracks; Pro Studio Masters) but this is one of those recordings that's not available for download above CD quality.
If the original file is "only" 44.1/24, this raises two questions in my mind:
1) Why bother to process it through MQA at all? The file size would be smaller if just straight FLAC were used, plus there would be no loss in resolution.
2) If the original graph posted with the "notch" centered at Fs/2 really is just a 44.1/24 file, what is the content seen in the dual-rate band? Clearly it is not just an alias of the original baseband audio, or it would be a mirror-image. Even after passing through the MQA digital filter, the mirror image would only be down -3dB at ~38kHz. The graph does not look like that at all to my eyes. Instead it looks like the spectral content expected in a true 88.2/24 file but with two artifacts - the big notch at Fs/2 plus the mild hump between 40kHz and 45kHz.
As always, these posts are strictly my own opinion and not necessarily those of my employer or second cousin, twice removed.
Charlie - I may be technically all wet, but from what little I know, I do believe what you are looking at is an alias of the original signal. My layman's reading of Stuart's technical papers is that he was willing to accept a certain amount of aliasing, perhaps high by other's standards, in order to get the time domain performance he wanted, all based on his tested psychoustic experimentation, of course.
So, what you may be seeing is precisely that instead of an origami unfolding or other bug.
My limited listening so far, which is of course subjective and anecdotal, makes me believe there is ample reason to apply MQA to a 44 or 48k Master, even if no higher sampling rates are available. In fact, I find that MQA has MORE to offer there vs. the non-MQA original than it does when comparing MQA vs. non- at higher resolutions.
Stuart's published graphs in AES papers and elsewhere attest to the fact that 44k has much greater "temporal blur" than at higher sampling rates, and that impulse response is cleaned up much more by MQA at 44k than it is from, say, 192k masters. My subjective impression is that MQA/44K sounds much more like native hi rez.
I have not done much native hi rez vs. hi rez MQA comparison, but I feel there might possibly be considerably less to be gained from MQA by a hi rez listener like me.
> > I do believe what you are looking at is an alias of the original signal < <
I'm pretty sure that it is not aliasing. When you read a book on digital audio (such as Ken Pohlmann's popular one, "Principles of Digital Audio") they almost always show the aliased signals as mirror images of the original signal, extending upwards in frequency to infinity.
Mathematically this is how it works, but in real life there is no such thing as "infinite frequency". It's shown this way because the digital audio theory presented is based on an imaginary abstract concept called a "Dirac delta", which is defined as an infinitely narrow impulse that still contains a finite (quantized) amount of energy. There have been a handful of DAC chips made with pulse outputs (including the one in the original Sony SACD players, the SCD-1 and SCD-777). While not *infinitely* narrow, they would still create a good mirror image of at least the *first* aliased spectral reflection. With a chip like this, the unfiltered spectrum would have a "V" shape - the actual audio energy decreasing with frequency up to the 20 kHz cutoff of the anti-aliasing filter (in the A/D converter) and then mirror-imaged upwards to the sampling frequency.
However pulse-output DAC chips are extremely rare compared to a "zero-order hold" DAC chip that holds the value of a sample until the next sample is entered. These chips output a waveform that looks like a "stair-step" representation (see Figure 3 in link below), and finally the reconstruction filter in the D/A converter filters out the high frequencies (artifacts of the "steps"), leaving a smooth analog waveform without steps.
The "zero order hold" found in nearly all DAC chips performs an unavoidable combination of low-pass filtering and comb filtering (search for images of the "sinc function"). Specifically the audio will be about -4dB at Fs/2 and gradually falling to zero at Fs. However this curve has a known, specific signature, as does the natural spectral content of musical instruments, as do both "leaky" or "brickwall" or "apodizing" digital filters. One skilled in the art can examine the spectrum of the analog output and identify each of the "fingerprints" left behind by each. In my opinion, the waveform shown does not look like what would be caused by a "leaky" filter, but instead exactly what would be expected if the original file were recorded at 88kHz - except for the notch centered at Fs/2.
As far as "improving" the sound quality of an existing single-rate recording, the picture is far from clear regarding MQA. If you are comparing a 44kHz file to the MQA version of the same recording, the MQA file will use different digital filters during playback - even if listening through the exact same D/A converter. If you prefer the sound of the MQA version, it would seem to indicate that you prefer the sound of the MQA digital filter to the standard one in that same D/A converter. To me that simply confirms something that has been known for decades - that different digital filters sound different. In that case the question becomes "Does the MQA digital filter sound better than *all* other digital filters, or just the other one built into this particular D/A converter?" I believe this last question may explain some of the mixed opinions currently existing.
As always, these posts only reflect my personal opinions and not necessarily those of my employer or the local chief of police.
In my previous post I wrote "Clearly it is not just an alias of the original baseband audio, or it would be a mirror-image.". This would *only* be true if the DAC chip output a series of (ideally infinitely) narrow pulses. However virtually all DAC chips output a given level that is "held" until the next sample (called a "first-order hold"). This alters the frequency response of the output by convolving (basically superimposing) the frequency response of the "sinc" function, where sinc (x) = [sin(x)/x]. Regardless the original chart does not look like that either, so my confusion persists... :-(
OK, I'll answer my own question: It does make sense--at least I think it does. Is it correct? No idea, but it makes sense.
I don't remember what the MQA patent application says--I found that hard going and, like most patent applications, absurdly repetitive: sentence after sentence saying the same thing, or almost the same thing with minor (if important) variations. However, that JAES conference proceeding that outlines the MQA approach (without ever mentioning those three letters) is completely clear about the two families of MQA files, 44.1 and 48. One is not converted into the other; they remain separate. "We thus have recipes for downward and upward conversion within a hierarchy of rates such as 44.1, 88.2, 176.4 and 352.8 kHz, however these methods do not provide satisfactory conversion from, for example, 96 kHz to 88.2 kHz." They go on to give a recipe for converting, eg, 192 to 176.4, but it's clear they think such a procedure would not be satisfactory at lower sampling rates.
When fully decoded, the Beyonce track plays back at 44.1/24.
What, then, is the content above 22kHz? Again, I don't know, but this makes sense: There's a mirror image of any signal above Nyquist, unless it's filtered. MQA has a slow roll-off filter. (I should check back and remind myself HOW slow it is, but I'm too lazy to do that.) So, the image above Nyquist is attenuated but not suppressed. You'r seeing the attenuated image.
Given all that--and considering that, as the CD spectrum shows, the Beyonce track is brick-walled at about 21kHz--this is just what I'd expect to see.
What "bug" are you referring to?
> that JAES conference proceeding that outlines the MQA approach (without ever mentioning
> those three letters) is completely clear about the two families of MQA files, 44.1 and 48.
> One is not converted into the other; they remain separate. . . They go on to give a recipe for
> converting, eg, 192 to 176.4, but it's clear they think such a procedure would not be
> satisfactory at lower sampling rates.
The exception is when a 48kHz-family master is MQA-encoded as a CD, which, of course,
mandates a 44.1kHz rate. For example, the Piazzola CD started out as a 192kHz master
that had to be sample-rate converted to 176.4kHz before being encoded as an MQA stream.
Post a Message!
This post is made possible by the generous support of people like you and our sponsors: