MQA is Vaporware

[FredericV]

1 hour ago, abrxx said:

 

Coming from my background as a software developer, and from reading the MQA patents, I would imagine a process whereby an internal 32-bit coding space is used to perform the "fold". Then, using subtractive dither (their words), the 32-bit coding space is transformed into a 24-bit one. If the dither can be reversed (which MQA claim), then the process can be reversed to construct a 32-bit/48Khz dataset, from which something resembling the input signal can be constructed.

This is probably never going to happen. They cannot reconstruct the original audio above 44.1 or 48 Khz.
All there is, is a first unfold + upsampling with a leaky filter causing aliasing.

Otherwise if there was a third unfold, and this data was buried and recovered, the plot would not be like this:

[mansr]

1 hour ago, abrxx said:

Coming from my background as a software developer, and from reading the MQA patents, I would imagine a process whereby an internal 32-bit coding space is used to perform the "fold". Then, using subtractive dither (their words), the 32-bit coding space is transformed into a 24-bit one. If the dither can be reversed (which MQA claim), then the process can be reversed to construct a 32-bit/48Khz dataset, from which something resembling the input signal can be constructed.

You can't turn 32 bits into 24 without losing something.

[FredericV]

They clearly do not read this topic, since when is DXD a multiple of 48 Khz?

[abrxx]

5 hours ago, mansr said:

You can't turn 32 bits into 24 without losing something.

Yeah, I thought about this a little more and realized my theory doesn't really work.

 

The first part, converting 32 bits (or perhaps 30 bits) to 24 bits, is the easy bit, and we know MQA does this, and how: subtractive dither and noise shaping. This adds the extra 6 bits of perceptual resolution. But this is perceptual resolution, not coding space.

 

The engineer within me is trying to crack this!

[Thuaveta]

3 hours ago, abrxx said:

The engineer within me is trying to crack this!

 

The engineer in you is possibly using the wrong tools: what's behind MQA likely has little to do with hard science, and much to do with marketing (let's not insult psychologists by calling it anything else).

 

Trying to make BS's and followers' science claims make sense in reality is assuming the solution to changing a lightbulb necessarily includes the use of a hammer.

[mansr]

9 hours ago, abrxx said:

Yeah, I thought about this a little more and realized my theory doesn't really work.

 

The first part, converting 32 bits (or perhaps 30 bits) to 24 bits, is the easy bit, and we know MQA does this, and how: subtractive dither and noise shaping. This adds the extra 6 bits of perceptual resolution. But this is perceptual resolution, not coding space.

 

The engineer within me is trying to crack this!

Reverse engineering the decoder, I have found that the bits in an MQA file are used like this (with LSB as bit 0):

• 9­­–15: Uncompressed audio data, 0-24 kHz with shaped pseudo-random dither.
• 8: Various low-rate data, including:
  • MQA identifier
  • Original sample rate
  • Desired "rendering" filter
  • Authentication signatures
  • Touch-up data for audio baseband
  • (Optional) Metadata (artist, title, etc)
• 0-7: Compressed audio data, 24-48 kHz

The decoding process works roughly like this:

1. Locate MQA identifier in bit 8 and initialise dither generator.
2. Subtract generated dither. This has two effects:
   1. Unmask compressed data in bits 0-7
   2. Reduce noise level in baseband
3. Extract compressed data (bits 0-7) into separate buffer.
4. Apply touch-up data (bit 8) to audio samples.
5. Decode compressed high band.
6. Join baseband and high band.
7. Encode render instructions (filter, etc) into LSB of 24-bit samples.

The rendering works like this:

1. Locate identifier in LSB.
2. Extract parameters:
   • Original sample rate (mainly used to lie on display)
   • Upsampling filter number
   • Dither level and shape
   • Gain
3. Apply gain if requested.
4. Upsample to device-specific rate using indicated filter. This typically the highest rate supported by the DAC chip and is not related to the original rate of the file.
5. Apply dither using parameters from step 2:
   • Bit depth is one of 15, 16, 18, or 20. Only 16 and 18 have been seen in the wild.
   • Shape is flat or one of two options with noised pushed towards high frequencies.

In devices like the Dragonfly that rely on the DAC chip for upsampling (with programmable coefficients), steps 4 and 5 are reversed.

 

Most of the things BS talks about are nowhere to be found in the actual product. You're wasting your time looking for them.

[mansr]

7 hours ago, Thuaveta said:

The engineer in you is possibly using the wrong tools:

Here's a very useful tool: https://ghidra-sre.org/

 

7 hours ago, Thuaveta said:

what's behind MQA likely has little to do with hard science, and much to do with marketing (let's not insult psychologists by calling it anything else).

BS calls it neuroscience.

[The Computer Audiophile]

5 minutes ago, mansr said:

BS calls it neuroscience.

Many others call is pseudoscience. 

[rickca]

1 hour ago, mansr said:

8: Various low-rate data, including:

• MQA identifier
• Original sample rate
• Desired "rendering" filter
• Authentication signatures
• Touch-up data for audio baseband
• (Optional) Metadata (artist, title, etc)

 

How can MQA do all that with just one bit?  

[The Computer Audiophile]

2 minutes ago, rickca said:

How can MQA do all that with just one bit?  

When it comes to MQA, be careful asking question to which you don't already know the answer :~)

[abrxx]

2 minutes ago, rickca said:

How can MQA do all that with just one bit?  

It's two dimension remember: one bit over time. Its just a block of data encoded. It could be binary data, although perhaps there is another way of doing it.

[rickca]

Just now, The Computer Audiophile said:

When it comes to MQA, be careful asking question to which you don't already know the answer :~)

I don't get your point, sorry.  Am I being dense?

[Iving]

5 minutes ago, The Computer Audiophile said:

When it comes to MQA, be careful asking question to which you don't already know the answer :~)

 

3 minutes ago, rickca said:

I don't get your point, sorry.  Am I being dense?

 

Old adage to trainee lawyers ... 😉

[abrxx]

1 hour ago, mansr said:

Most of the things BS talks about are nowhere to be found in the actual product. You're wasting your time looking for them.

 

Curious, what are you thinking of?

[rickca]

Just now, Iving said:

Old adage to trainee lawyers ... 😉

Sure I'm familiar with the risk to lawyers, but I'm sure that Mans can answer my question.  

[mansr]

1 minute ago, abrxx said:

It's two dimension remember: one bit over time. Its just a block of data encoded. It could be binary data, although perhaps there is another way of doing it.

Right, it's a sequence of data blocks serialised into one bit per sample. The format, to the extent I have deduced it, is documented here: https://code.videolan.org/mansr/mqa/blob/master/mqa.rst

[The Computer Audiophile]

9 minutes ago, rickca said:

I don't get your point, sorry.  Am I being dense?

I meant it more with respect to asking questions to MQA ltd. Ask a simple question and the answer will be a roundabout containing half-truths and fodder for the believers. 

[rickca]

Just now, The Computer Audiophile said:

I meant it more with respect to asking questions to MQA ltd. Ask a simple question and the answer will be a roundabout containing half-truths and fodder for the believers. 

OK that's what I figured you must have meant.  I'm done with MQA technobabble.  I wouldn't even bother asking them questions.

[Allan F]

19 minutes ago, Iving said:

Old adage to trainee lawyers ... 😉

With the exception that there are rare occasions when you can ask a question that, regardless of the answer works for you, a "Have you stopped beating your wife?" type of question.

[FredericV]

2 hours ago, mansr said:

Here's a very useful tool: https://ghidra-sre.org/

- strace
- running code in a chroot jail
- transprant proxy intercepting files from streaming CDN's and in flight firmware images for vendors who do not offer download links
- readelf
- objdump
- binwalk
....

[Iving]

5 minutes ago, Allan F said:

With the exception that there are rare occasions when you can ask a question that, regardless of the answer works for you, a "Have you stopped beating your wife?" type of question.

 

Oh yeah I would never have made a good lawyer 😉

[Cebolla]

3 hours ago, mansr said:

Reverse engineering the decoder, I have found that the bits in an MQA file are used like this (with LSB as bit 0):

• 9­­–15: Uncompressed audio data, 0-24 kHz with shaped pseudo-random dither.

This is a typo, presumably, so the range should be the high15 bits, ie, bits 9-23, rather than ending at bit 15.

 

Mind you, may be the high 7 bits 9-15 are actually the uncompressed 0-22kHz audio data in the MQA-CD file - yuck!

 

 

[mansr]

38 minutes ago, Cebolla said:

This is a typo, presumably, so the range should be the high15 bits, ie, bits 9-23, rather than ending at bit 15.

Yes, obviously. Thanks for noticing.

 

38 minutes ago, Cebolla said:

Mind you, may be the high 7 bits 9-15 are actually the uncompressed 0-22kHz audio data in the MQA-CD file - yuck!

MQA-CD simply lacks the low 8 bits of compressed high-frequency information. The remaining 16 bits are used the same way as in regular MQA.

[FredericV]

38 minutes ago, mansr said:

MQA-CD simply lacks the low 8 bits of compressed high-frequency information. The remaining 16 bits are used the same way as in regular MQA.

Bob claims MQA CD can produce content up to 36 kHz while also claiming the following:

 

Quote

When the transmission channel is limited to 16 bits, the decoder maintains good temporal performance, as the comparison between the olive and black curves show.  Audio below 20 kHz is maintained at full resolution, while some precision is lost in the next octave.

Notice the part in bold, with 8 bits gone, there is no data to do the first unfold. All MQA can do is upsample with their leaky filter, to generate a fake octave above redbook:
 

We can generate the same fake content with SOX. And Bob describes this as "some precision is lost in the next octave".

What precision????

[abrxx]

I think you misunderstand MQA, either that or it has changed since I last read about it.

 

With only 16 bits of coding space to play with, I would imagine 13 bits dedicated (on the MSB side) dedicated to 0-24Khz, and the remaining 2-3 bits holding the approximation stream for the hi-rez section (24-48Hz). Also, the MSB side will not be identical to the original master, but will be noise-shaped and dithered, so as to get say 15.9 bits of perceptual resolution from 13 bits.

 

The MQA decoder can then take all of this and render the high resolution section, not bit perfect, but as an approximation (since the touch-up stream, located beyond bit 15, is missing).

 

This has been described over and over again in various literature from MQA and Bob, as well as appearing in their patents. @mansr do you see something quite different in the code you reverse-engineered?

 

17 minutes ago, FredericV said:

Notice the part in bold, with 8 bits gone, there is no data to do the first unfold. All MQA can do is upsample with their leaky filter, to generate a fake octave above redbook:
 

We can generate the same fake content with SOX. And Bob describes this as "some precision is lost in the next octave".

What precision????