MQA at CES

[Miska]

Most Meridian Active DSP speakers have a proprietary encrypted signal transfer technology called MHR (think of it like an audio HDCP). This ought to allow them to receive a deconvoluted MQA stream digitally without any potential software copyright infringement being possible from the source output. Given the AURALiC debacle I wonder how they will get comfortable with a straight MQA decoder outputting unencryted digital decoded MQA.

 

Why does it need to be encrypted? IOW, why does it need to have DRM?

 

I can buy plain normal unencrypted hires FLACs from NativeDSD, HDtracks, highresaudio.com and many other places. Why would we need to have this last step encryption? Which is ineffective anyway, because it is very simple to rip off the digital data straight from DAC chip's I2S pins.

[Miska]

Michael Lavorgna follows up on MQA:

MQA Continued | AudioStream

 

Funny part No.1:

By creating what I'll call DAC profiles, the MQA decoding process sends a DAC the data it can handle, whether that be a max sample rate of 48kHz in an smartphone, 96Khz in a tablet, and so on, while also correcting for unwanted digital artifacts and, in Bob Stuart's words, "make the DAC chip work at its optimum performance level."

 

But i's all PCM, so it doesn't bypass or implement a delta-sigma modulator which is half of the picture in digital domain. And they are very hard trying to prevent us from measuring the performance.

 

Funny part No.2:

If you do not own an MQA enabled DAC, yet you play a high-res MQA encoded file on it, the file will play at 24/44.1 regardless of the original resolution.

 

While the end result is actually less than 16/44.1 would be able to produce. Rest of the 8+ bits is useless waste of bandwidth.

 

Funny part No.3:

The MQA encoding process also reduces the associated file size of high-res files to roughly the size of a 24/44.1kHz file for transport. This makes the download process less painful for people with less-than-fast Internet connections, which is great but frankly not so sexy.

 

Since most of the 24-bits is pure noise from FLAC encoder point of view, it actually ends up being larger than equivalent normal hires FLAC file would be. But instead implements lot of constraints.

 

 

And sure enough the main reason why I have zero interest on such:

While the MQA decoder can incorporate software-based DSP for room correction or EQ (e.g. in AV Receivers), any such processing (including sample-rate conversion or software-based upsampling), can’t be introduced between the decoder and the DAC chip without breaking the chain.

 

IMO, DACs just don't have enough DSP processing power to do things properly and I wouldn't want to spend money buying another computer inside a DAC. Now I can use one computer with all my DACs, no need to embed a separate computer in each one of those.

[Miska]

I don't know if Miska has seen these or agrees with the analysis but I also think that SPZZZZKT (the one creating these filters & doing the above plots & analysis) changes his analysis at a later stage - I'll try to find it

 

This is current version of the new filter I created for playing MQA encoded files to fix the problems introduced by the MQA encoding:

[Miska]

You don't show an x-axis calibration so it's difficult to compare with above plots but in general you seem to favour a wide shoulder on these plots - in other words in the audible range above, let's say above -70dB, a smeared impulse response? Can you say something about this, please?

 

This is much narrower than any of the ones you posted. X-axis is not important, what is important is number of side lobes and the total attenuation. Notice that this plot has 200 dB Y-axis while the ones you posted have 100 dB Y-axis, which is 100000x difference. In my plot, the level difference between main lobe and first side lobe is 20 dB while in the plots you posted it is 6 dB less at 14 dB.

 

This one reaches -96 dB attenuation in 9 cycles which is as designed.

 

So frequency response looks like this, which I find more telling:

 

This filter doesn't have more attenuation because MQA is only 17-bit content, so the filter has been tailored accordingly.

 

For 32-bit content filter responses look more like this:

[Miska]

Does your filter also unpack the MQA hi-res fils ?

 

No, it just cleans up the noise added by undecoded MQA.

 

Or is it only for 16/44.1?

 

Since MQA FLAC is 24-bit of which 11 bits are more or less noise without decoder, it is intended for cleaning up those 11 bits, and also to give optimal time domain performance given the source content properties.

 

Output is 64-bit floating point which is then converted either up to 32-bit PCM or DSD at what ever rate the DAC can support.

 

Will your filter be part of HQP ?

 

It is already in the beta, but subject to be further tuning.

[Miska]

OK, so why not post the x-axis detail - I don't see why you make it a mystery?

 

Because it is normalized.

 

Your plots, without X-axis detail, makes them ambiguous & unreadable/uninformative. When plotting an impulse response, time is hugely important - I don't see why you say it's not?

 

No it is not, I didn't pay any attention to X-axis numbers of the plots you posted because it is irrelevant. What is relevant is how many zero-crossings the filter has and what attenuation it reaches.

 

You can run the same filter at 1 Hz or 1 GHz sampling rate, the impulse response looks the same, just X axis numbers change.

 

Yes, but over what timeframe does it drop 20dB - this is what I'm asking you to provide - the x-axis grid

 

Time depends on the sampling rate. What is important is how many cycles of ringing you have.

[Miska]

And being able to broadcast a 48/24 stream instead of a 192/24 stream allows Tidal to offer more music to more people over the same "pipes" and thus make more revenue, and if you have one of the MQA-certified "best with Tidal" DACs it sounds just like hi res!

 

But why would you send 48/24 stream that has 17-bit worth of data and requires special decoder, while consuming more bandwidth than ordinary FLAC compressed 192/17 stream would?

 

Note that the bandwidth consumed by FLAC is not simply related to sampling rate and bit depth of the uncompressed material, but also depends on compressibility of the source material. MQA encoded material has huge amount of noise that FLAC encoder cannot compress and thus it consumes relatively much more bandwidth. So yes, indeed 48/24 encoded FLAC can be larger and thus consume more bandwidth than even 192/24.

 

Results of my compression tests of the 2L content was that I can consume less bandwidth than MQA using ordinary FLAC while preserving more (all!) relevant information of the source!

[Miska]

Thanks Jud but I'm not even sure I understand any of this - I simply posted this impulse response plot that was in Harley's article (I presume it came from MQA?)

[ATTACH=CONFIG]23806[/ATTACH]

 

That filter of "new system" just has horribly bad filtering performance. That's why it would be much more interesting to see frequency response plot of that filter.

 

If I put my MQA-fixing filter to 192 kHz sampling rate scale it is twice longer than the "New system" but half of the "Linear-phase 192 kHz".

 

I prefer to have good performance in both time and frequency domains.

[Miska]

They are not interested in or even aware of your 192/17 compressed FLAC stream because you did not come to them with a reputation and a sales pitch (and a way to make more revenue if the commercial idea works out), Meridian did.

 

It is not rocket science, so it is hard to believe someone running a streaming service wouldn't have competence to run such comparative analysis.

 

OTOH, I'm not interested in solving Tidal's problems. If their offering doesn't fulfill my requirements I will just discontinue my subscription and not worry about it more.

 

In optimal case, one would run automatic software analysis of the library and select automatically sampling rate and bit depth that fits the source content but doesn't leave any unnecessary extra. Doing so for couple of 2L recordings I concluded 120 kHz sampling rate and 18-bit being optimal which resulted in significantly smaller standard FLAC file (less bandwidth consumption) than the MQA put into 44.1/24 package.

[Miska]

You're kidding, right? Or is it customary in Finland for people who run businesses to be familiar with what they consider the "technical aspects"? Here the executives don't need to know about such things.

 

Maybe executives don't, but they usually have people on payroll who have the competency. It is actually more customary here that executives are great engineers but really bad running a business...

[Miska]

If there is any basis in demanding lumped impulse responses be < 10 us for monaurally applied sound then we should expect them to point to this evidence, instead of hiding it in a waste dump.

 

Given that RedBook sample is 22.7 µs long and sample at 96 kHz is 10.4 µs long it is pretty steep requirement if proper anti-aliasing is to be applied and some extended bandwidth preserved. 10 µs is not problem for DSD though, so there's a solution for such without needing to resort to closed approach.

 

But overall, I understand that their idea is that anti-alias filter begins to roll-off already at 20 kHz and reaches maximum attenuation at 48 kHz (Nyquist frequency of 96 kHz sampled audio). Then they rely on this roll-off characteristics to minimize the amount they need to fold down into LSBs of the 48/24 container.

 

However, feeding data through such filter and cutting the word length to 17-bit as they do, the resulting standard 96/17 FLAC file is much smaller than the MQA-encoded one because the FLAC encoder can do much better job since it doesn't have to deal with data that is pure white noise (and thus uncompressible) from it's perspective.

[Miska]

All they can do for CD-rate recordings is assess if their anti-aliasing was done with half-band filters (**), and then apply the Meridian minimum phase apodiser (*).

 

I'm running my own apodizing upsampling filters at playback time. And I select between linear- and minimum-phase apodizing filters based on source content.

 

While I prefer things that way, someone else may prefer differently. Since it can be listener's choice, I think the choice is best left to be done by the listener based on his own preferences.

 

For hires sources it is also possible to run similar slow roll-off apodizing upsampling filters to reach desired behavior.

[Miska]

There is some new marketing material at Audiostream:

http://www.audiostream.com/content/supplementary-listeners’-notes-2l-test-bench

 

I wonder what they used as 44.1/16 source to make it look worse than undecoded-MQA variant? Why didn't they use plain TPDF dither for 16-bit because it performed best?

 

But anyway good demonstration to stay away from MQA and stick with the DXD master download instead. Or bake your own RedBook version from the DXD if you have to.

 

Of course they didn't include comparison to a standard 96/18 FLAC for example especially taking into account size of the produced file vs MQA encoded one.

 

3 2L sensibly use shaped quantisation for their CD releases.

 

Sensibly?

 

To make it perform worse than TPDF would have (minimal improvement in mid-range with massive damage in top octave).

[Miska]

Oh so I need to buy a new DSD capable DAC. I see.

 

No, why would you?

 

You can play DXD on a DAC that supports 352.8/24, or through player's conversion on any PCM or DSD capable DAC with your choice of conversion options.

 

Why would you buy anything else than the original master, especially if it's in some proprietary protected container that has limitations on how it can be used?

[Miska]

I thought that was HQPlayer? (Also sarcasm).

 

What I'd like to note here as being difference is that if you take a standard FLAC, the information how to decode it is public and well documented knowledge. You can choose HQPlayer or any other audio player to play it back, and you can change the player at any time at your willing, or implement your own. And every one of those have equivalent capability of decoding all the included information, based on public and free knowledge of the content format. They also have equal opportunity for employing any processing they like on the content, and what ever is the result is temporary and not preserved since it can be reproduced any time in same or different form the same source content.

 

Thus, I have no problem if:

- MQA is source-only process

or

- MQA is DAC-only process

and

- There is a non-degraded standard compliant alternative format

 

 

For streaming services the damage is not persistent, so in a way it is similar to choosing one player over another. However, the damage is persistent if file downloads are offered only in MQA but not in any standard compliant form. (like iTunes Store before they removed DRM)

[Miska]

I don't think this is related to correct use of dither as that's a ADC issue & not a DAC issue? I believe the phenomena being spoken of is how a digital processor's noise floor shifts (modulates) with changing input level. Suggested in an AES paper given by Dr. Richard Cabot, Principal Engineer at Audio Precision. Stereophile followed up on this & did measurements of DACs "measurement reveals shifts in the noise floor's spectral balance with changes in signal level. Ideally, the noise-floor spectrum should remain constant with level, producing curves that exactly overlay each other. Psychoacoustic research by Louis Fielder at Dolby Labs (those guys know something about noise-floor modulation!) indicates that noise-floor shifts of 2dB are audible. Further, Dr. Cabot's paper asserts that the ear is very sensitive to shifts in the noise floor's spectral balance; changes on the order of 1dB are reportedly audible." (Read more at Noise, Modulation, & Digital/Analog Conversion | Stereophile.com)

 

MQA cannot fix the DACs noise floor modulation by itself because it operates only in PCM domain and the noise floor modulation happens in the modulator section. However, it can be fixed by upsampling to DSD using a modulator that doesn't have noise floor modulation, since it bypasses the DAC's modulator.

 

However, the noise floor modulation is not an issue as long as the source material's dithered noise floor is higher than the modulator's noise floor which is the case in most practical situations. Since MQA reduces word length to 17-bit before encoding, this should be the case there too.

 

 

...of course the entire problem with digital filters and time domain performance is avoided by using DSD and that is the original purpose of DSD... So the problem MQA is trying to solve has been solved already long ago with similar end-to-end solution that is already more widely supported and adopted. (and luckily doesn't impose DRM schemes anymore since it's been freed from SACD jail)

[Miska]

However, are you saying DSD modulators don't have issues of noise modulation?

 

Some do, some don't. It depends on the modulator implementation. Better choose a good one...

 

How can you separate the dithered noise floor from the modulator's noise floor?

 

Higher, dithered noise floor buries the lower modulator noise floor. If you TPDF dither PCM for example to 16-bit, you have a constant signal restricting the dynamic range. If the modulator noise floor has noise modulation of +-2 dB at let's say -160 dB level, it is completely buried by the noise floor at -96 dB level.

 

Does DSD not suffer from the same issues of noise modulation & the added problem of dithering on 1-bit systems - see "Why 1-Bit Sigma-Delta Conversion is Unsuitable for High-Quality Applications"

 

As I said in the first part, it depends on the modulator used to produce the DSD. There is no problem dithering DSD, my modulators are dithered. You can verify this by feeding the same data to the modulator multiple times - the produced DSD stream is different every time. That article is ancient and the authors just demonstrate either complete incompetence in modulator design, or just want to make a point that bad modulator produces not so great results. But even those results are much better fidelity than what you get from the PCM -> SDM conversion that happens inside DAC chips...

 

 

As an example, here are MD5 checksums of the same source file converted three times using same parameters:

271702e70f23a2fc5693116fb5b05dd8 tmp1.dsf

74ad55c5469330ccb176692268874889 tmp2.dsf

02cbc4e82237a7d411c7b552f9d00323 tmp3.dsf

 

 

Also for comparison, here's a 1 kHz level sweep, one has been converted to 352.8/32 PCM and then to RedBook format, the other one has been converted to DSD64 and then to RedBook format (TPDF dither). Can you tell a difference?

 

[Miska]

Do you (all) have opinions why many high end DAC manufacturers are still choosing not to make their products DSD compatible given its apparent superiority?

 

Like dCS, Playback Designs, EMM Labs / Meitner, Esoteric, T+A, Chord, etc?

 

Where PBD and EMM/Meitner are explicitly DSD-only architectures and convert to DSD internally. While T+A has a completely separate discrete DAC section for handling DSD:

 

What high end manufacturer you have in mind who doesn't support DSD?

[Miska]

Related to the this thread and following Miska's comment that a 18bit/96KHz FLAC would actually be smaller than an MQA file plus would include more information (at least this is what I understood)... Is there a variable bit rate / variable bit depth format out there that could do this dynamically and efficiently? Yes, it would be lossy, but I'm not so concerned about that.

 

FLAC can support anything from 1 Hz to at least 384 kHz as sampling rate and 1 - 32 bits. You cannot change rate or word length during the track though, but you can vary it from track to track freely.

 

Maybe the answer is actually as simple as streaming in compressed DSD64? Miska: What's the bandwidth requirement of that?

 

DSD is simple, required bitrate is the sampling rate multiplied by channels. So stereo DSD64 stereo is 5.6 megabits. You can shave off at least about 30% of that with lossless compression.

[Miska]

In a sense Apple created the need for Tidal/MQA by never delivering a high res streaming capability. Nevertheless they still have time to do so, and have the technology chops to make this happen. If I owned the Tidal or MQA company, a high resolution format from Apple would be the seen as the biggest risk to my business. Given Apple's recent focus on managed services revenues, it would not be surprising to hear an announcement of a competitive response from them in the near future.

 

I'm pretty sure they'll have some 96/24 AAC compressed hires any time soon. Would be great if they'd do 96/24 ALAC though...

 

If I would have to choose between 96/24 AAC or MQA, I would take the AAC because I know exactly how it works...

[Miska]

Was thinking also of Octave and Aqua. Audio-Gd makes both but his 1704 DACs seem to get the most love. Not making a case against DSD, just curious.

 

I have also Metrum Musette which is discrete R2R (no DAC chips), it can take in 352.8/384k PCM, so a good target for software upsampling... Naturally it cannot deal with DSD because it is one of the rare PCM DACs.

[Miska]

Right. Rega would be another. Why the resistance? Of course Rega doesn't want MQA either.

 

Rega uses standard SDM DAC chip, they just don't support DSD inputs. But the DAC chip itself converts to DSD-like format before the actual conversion.

[Miska]

R2R PCM DACs are rare, but R2R/flash PCM ADCs are rarer still. Which means that even on the R2R PCM DACs what people listen to most of the time are delta sigma recordings (downsampled and decimated to PCM with on-board ADC filters). But that part of the story is rarely mentioned in R2R DACs' marketing materials

 

I'm not aware of any non-SDM audio ADC on the market at the moment. Is there any?

[Miska]

1-bit DSD into most significant bit and analogue low-pass filter would do it.

 

Or maybe they run it over a moving average filter of N-size window, which is almost "bit-perfect"...

 

That would give some use for more bits.

[Miska]

We could also ask why certain manufacturers are holding out on class D amplifiers, while others are embracing the technology. Or why certain manufacturers are holding out on tube amplifiers, while others firmly believe they deliver the best sound, etc. etc.

 

If you make a really sophisticated class-D amplifier it looks like a high-voltage DSD-DAC...