The following graphic has been posted regarding MQA's audible effect on the noise floor of a sample 2L recording:
It would seem that the format introduces significant in-band noise whether it has been decoded or not.
It would seem that the format
does not introduce significant in-band noise whether it has been decoded or not
Extract from Computer Audiophile MQA Q&A article:-
PLEASE COMMENT ABOUT THESE INVESTIGATIONS/BLOGS
Q82. Please comment on these posts.
a. http://archimago.blogspot.ca/2016/01...s-and-big.html (Blog post has been removed by author - Editor)
In this blog you will see, that from the technical point of view, MQA have around 13 Bit of “lossless” information and everything below 14 Bit is “lossy”. Doesn't mean that is will not sound good, it just means, that this is not a lossless codec, it is lossy (from the technical point of view).
b. http://www.computeraudiophile.com/bl...ires-flac-674/
A82. We have paraphrased the assertions: 13
i) “MQA have around 13 Bit of “lossless” information and everything below 14 Bit is “lossy”
This is incorrect. In general, the MQA system can reach in excess of either 23-bit dynamic range capability or 3–6 bits below the content noise in the audio band.
ii) “Without a decoder we hear 13 bits, that isn’t CD quality”.
Here is a classic case of comparing apples to oranges. When we talk about CD quality sound we don’t expect an answer that says ‘it can’t sound like CD, because I can see only 13 bits’. Do we listen with our instruments? Even after years of working in this area we can’t look at an FFT plot and tell you how something will sound. We can maybe tell you the information capacity of the signal or the channel. One clue why it doesn’t help is in the second ‘F’ (for Fourier).
In any case the 13-bit number is wrong. Try as we might there is no way to tell the information capacity of a channel from a spectrogram (as in one of the cited posts) – the graphs look pretty but are basically meaningless.
As described earlier, if you don’t have a decoder, the channel capacity appears to be typically > 15 bits for the files on the 2L Testbench and this is limited by considerations of compatibility, not coding space. The noise is frequency shaped to minimise audibility, as it is for many well-produced CDs. If you have a decoder then, depending on the authoring parameters, the noisefloor in the recording should not be increased anywhere there is music signal.
iii) Paraphrase: ‘The Nielsen recording shows that MQA are cheating. They take a 16-bit recording and give us back a 24-bit file with lots of noise in it’.
Wrong. All one had to do was read Morten’s notes to guess it might have been remastered to 24 bits, See
A40.
Nielsen: 2L-120 Track 1
As can be seen in the following, the inherent noisefloor of MQA in this recording is actually:
Without Decoder: MQA channel noise is lowest around 4kHz @ 17.5 bits with a channel capacity of
15.8 bits which has been shaped. The MQA noise is always below that of the CD release.
With Decoder: MQA channel noise is lowest around 4kHz @ 24.3 bits with a channel capacity of
over 23 bits which has been shaped.
We have added to the graph (from our earlier note on the 2L website) to make this clear.
(High Resolution Image)
Note: the 24-bit Master and MQA (decoded) peak noise curves overlay and are not separately visible.
These graphs confirm that 2L’s Original, CD and MQA versions of the files are consistent in level and response. Of course spectral plots using FFT have no time-domain information, but we can use them to compare the peak spectrum of the Original, CD, and MQA with and without a decoder.
Also shown is a comparison of the background noise throughout each version and the reference level for 16-bit TPDF dither in a channel sampled at 44.1 kHz. 14 15
In the graphs the peak and noise-floor curves overlay for both MQA decoded and Original master. We can also see that the shaped noise introduced by the MQA encoder and ‘heard’ without a decoder is removed by the decoder and is also below that of the CD release, even without decoding.
Additional curves explained:
With a Decoder: Brown (with open circles): This shows the underlying end-to-end MQA channel noisefloor (with a decoder) in this recording, which
clearly shows that here the inherent noise of the MQA process is at least 10 bits (i.e. 60dB) below the noisefloor in the recording at all frequencies up to 22.05 kHz and close to 24 bits between 4kHz and 20 kHz.
Navy: shows the level of 24-bit TPDF dither for reference.
No Decoder: Magenta (with open stars): This shows the underlying MQA noisefloor for the listener with no decoder. It is lowest around 4 kHz and 12 kHz to minimise impact; is essentially below the 16-bit level up to 14 kHz and is always below the noise of the CD version. The inherent noise in the recording dominates below 15 kHz.
16
Note: The noise seen by a Legacy (no-decoder) listener is the sparse signalling channel, not lossy noise in the file.
iv) Paraphrase: ‘MQA increases the noise in some recordings’ (an experiment using Explorer 2).
The underlying thesis in this blog has been to demonstrate that, because MQA uses burying techniques in the lossless folds, that somehow the dynamic range is restricted to 16 bits or fewer. We showed this to be incorrect regarding the Nielsen recording. We also disagree with the blog’s findings in the case of 2L ‘Blågutten’ from Quiet Winter Night. The graph below shows analysis of:
(High Resolution Image)
· Files: Background noise levels in original DXD source and MQA file.17
· Explorer2 analogue output when receiving: MQA (decoded in Explorer2) and the 192 kHz PCM version on the 2L testbench (background noise).18
· References: showing 16- and 20-bit noisefloor @ 352 kHz (note, 9dB lower than at 44 kHz).
· Analysis: The underlying MQA channel noisefloor in this file.
· Hearing thresholds (steady-state) referenced to a playback acoustic gain of 105dB SPL.
The end-to-end core MQA noise floor in these encodings is always at least: 5 bits below the noise- floor of the recording up to 11 kHz, 4 bits below up to 22 kHz and is 3 bits below at 44.1 kHz (audio). However, no common DAC chip will reveal this due to internal noise. Even in these great 2L recordings we don't often see hall/microphone/ADC noise below the 16-bit noise spectral level -- not surprising given the fundamental thermal limit for microphones. See [2] and brown curve above.
In our experiment we don’t see the Explorer2 output deviate from the DXD or 192 kHz versions below 33 kHz. Above that there is rising dither from the DAC, but its origin is not lack of dynamic range.
The mastering engineer can set encoding or playback parameters where noise level can be increased or decreased in some frequency regions, but is not due to lack of dynamic range in the MQA system.
We should point out some key points for those less skilled in reading such plots:
I. FFT analysis like this does not give any clear indication of how it is going to sound because temporal information is excluded.
II. The dynamic range is huge; the silence in the recording is 1/3 the way up the graph. For steady noise we hear nothing in the shaded areas.
III. Even at high listening levels (e.g. acoustic gain of 112 dB), the noisefloor of the un-decoded MQA should be inaudible if the playback system is linear and has a flat response. With a decoder the noise is more than 20dB lower. [2][
5][
6][
7][
8]
IV. Very few headphones or loudspeakers can reproduce above 40 kHz (shaded blue area).
V. Very few microphones pick up above 40 kHz, including in this recording.
VI. Noisefloor above 44.1 or 48 kHz (especially at these levels) is more artefact than audio.[3]
13 There is an issue of bias: we take exception to blogs that block us from posting corrections!
14 The analysis uses 21.53Hz bins (=44100/2048 and 351800/16384) giving an offset of +13.33 dB wrt 1Hz.
15 2L sensibly use shaped quantisation for their CD releases.
16 Of course not all DACs can reach this low level of in-band noise.
17 Graph displayed up to 88.2 kHz for best comparison with blog.
18 The analogue output of Explorer2 was captured at 352.8 kHz/24bit in a Pyramix workstation using the Horus converter. The analogue noisefloor of the ADC is around 20 bits. Files were sent to the Explorer2 using Foobar.