Non Shannon-Nyquist sampling is applied at 24-48k bandwidth where only timing and amplitude are roughly capture. There’s very little frequency components exist in the ultrasonic range, thus it doesn’t require conventional sampling methods. This helps to dramatically reduce bandwidth.
Add: In a 24/192k PCM, one will get up to 24 bit resolution across the 96k bandwidth. MQA take advantage as there’s no much information above 24k, so it uses the least bit to code and combine with Non Shannon-Nyquist sampling to ultimately save a hell of bandwidth.
No sorry, that’s not what I meant. The first unfold IS lossy in the sense that the information to 24KHz audio freq (48KHz sample rate) is basically lossless, but the information from 24KHz to 48KHz audio freq (48KHz to 96KHz sample rate) is not lossless, it loses a great deal of resolution as it is encoded in about 8 bits. That is Non-Shannon sampling. The information past 48KHz audio freq (96KHz sample rate) is complete guesswork from an upsampler. So from 96KHz to 192KHz or higher (sample rate) it is all fake news.
Another point I wish to make is regarding on the ‘original sampling’ or simply ‘ORIGINALSAMPLERATE’, a tag ID found in a MQA file. MQA detection tool found in Roon can report the original sample of the recorded master either by reading the tag ID or simply play the file.
In the case of MQA, the original sampling does not necessary correspond to what is reflected on DAC display. For instance, if the original sampling is 48k, then MQA will decode and up-sample to 96k, this is usually referred to as MQA core, it will always output either 88.2/96k. FYI, there are many tracks in Tidal, especially pop genre are mostly mastered from 24/44.1/48k masters.
I believe why MQA core is always outputting 88.2/96k is that the reconstruction MQA filter (minimum phase slow roll-off) is designed cut off at 44.1/48k bandwidth so that impulse response or so called ‘de-blurring’ will look good.
Are you sure about this? My (admittedly weak) understanding of the case of 44/48 target rate was no unfolding is done at all.
Actually pretty easy to check by using a non-MQA DAC and checking the PCM rate out of a TIDAL or Audirvana first-unfold of such a file. I’m at work but might try and check when home.
Sorry for my bad, shouldn’t use the word ‘decode’, yes you are right 44.1/48k target rate does not do any unfolding but it still up-sample to MQA core at 88.2/96k, this one I’ve already verified on the Mytek Brooklyn MQA DAC display.
For Tidal app, I’ve yet to verify but I think it should be similar.
The upsample component is actually possibly the most ill-documented part of the process. Essentially this “rendering” part is just a choice of upsampling filter. The concept of “target rate” is essentially incorrect because most DACs will always upsample to the max rate their DAC chips can deliver that is in the “power of 2” family. For example, the ESS DAC in the Dragonfly always upsamples to 352/384 and what the “target sample rate” in MQA does is just a selection of the upsampling filter. That’s it.
In the case of the Mytek it would almost surely be the same: the ESS DAC will always upsample to it’s max operating rate (within the appropriate rate family, eg 352/384 or 704/768 whichever the max’s are). For each target sample rate there is a family of filters that can be chosen to upsample - that is the only difference. I believe that 88/96 has one such filter whereas 176/192 has 16 or 32 different ones, selected by the MQA embedded data.
I think you mis-understood me, I’m talking about the original sample rate from the recording masters. Say you buy a 24/48k FLAC download and playback, your DAC will display 48kHz, which is exactly intended.
However, if MQA encoded from 48k master and feed to a MQA DAC, it shows 96k, not 48k. However, MQA detecting tool inside Roon can shows the original sample rate. For example, Roon will show this, ‘Tidal 24 bit 48k, MQA 48k’ but not the MQA DAC. For me this gives an impression that is ‘up-sampled’ but for others, they may think they are listening to an 96k master! Cheat!
When MQA say lossless they do not exclude quantization and dithering. That is why I said almost lossless. I believe the MQA process analyzes the file and calculates the maximum bit depth required to encode the music. I didn’t think that this had anything to do with non Shannon sampling.
Because people are passionate about the subject so they speak out. My view is speaking out against MQA is only marginally less beneficial to the MQA cause than speaking in its favour because few are swayed by on line argument. It tends to entrench opinion rather than help shape opinion. I am pretty confident MQA HQ are more than happy their pet product continues to be discussed hotly all around the music and hifi communities.
To be clear, non Shannon sampling is sampling under the bandwidth required to reproduce the signal exactly without aliasing. So if you take a 192kHz audio file, and resample it at 96kHz (undersampling), that is non Shannon sampling. This is what MQA does in its encapsulation process, which takes files at 8x or 4x rate and undersamples to 2x rate. They use something like a B-spline kernel (see AES paper 9178 – its sitting on my desk as I type this). I believe that the exact kernel used depends on user input and/or analysis of the audio file.
The reason they do this is to obtain the temporal resolution of the 192kHz or higher sound file.
This process has nothing to do with the encoding stage, which takes a 96kHz file and uses an almost lossless process (with quantization though) to create the 48kHz end result. Naturally if the input is 48kHz to begin with none of the above has to happen at all.
Fair enough, agree with what you’re saying. My understanding was that non-Shannon sampling also referred to the 96KHz --> 48KHz process as well. Clearly all of these steps (downsample to 96 then to 48) are lossy.
I still don’t understand what you’re saying, sorry. If you have a target rate of 48, my understanding is there is no unfold so the rate out of the TIDAL app is still 48. The MQA data embedded in the PCM stream, which the Mytek reads, might be saying “use the very basic MQA upsampling filter” which the MQA DAC might interpret as 96 rate, I don’t know. But what is happening (my understanding) is: there is no unfold, but there is upsampling with an MQA-specified filter, and the upsampling is going all the way to the max rate of the DAC - 384 in the Mytek.
Edit: I think I misread your post. Is that plot out of the DAC or the PCM stream out of say TIDAL app?
Even if the MQA sample rate is 44.1/48kHz, Tidal desktop app and other certified MQA Core decoders will output 88.2/96kHz to a non-MQA DAC. This can only be verified by a non-MQA DAC. MQA DAC always reports the MQA sample rate.
Interesting… I wonder if this is an oversight of the code that does the first unfold, or if it is by design - meaning that “target rate of 44” really means “I will still unfold even if there’s no real info past 44 and choose the 44 filter thereafter.” Ultimately all DACs (except the rare NOS ones) upsample to their max rates.
This has been discussed on computer audiophile last year. As far as can be discerned, the up-sampling of MQA Core with Tidal to 96kHz (or 88.2kHz) is intentional. MQA Core output will always be at this rate unless the device does not support it. Like some have said, the DAC is most likely going to upsample anyway (MQA or not).
The conversion of 96kHz (2x) to 1x is not a downsampling, whether Shannon or non-Shannon. Its an encode. The encoder does throw a way a little bit of information (so its not technically mathematically lossless), but allowing for quantization etc you could say its essentially lossless. Its a very different process compared to the encapsulation/rendering.
I wonder what Audirvana does.
