Dual-mono / multi-mono dilemma

[Pim]

I received the answer from the Devialet Support Team. Bad news...
"There is a 32bits to 24bits conversion occurring in the product, on the master's amplifier."

This means that using 32-bit USB audio input is somewhat compromised.
From what I saw, this also happens when using DSD files: they are converted to PCM, and between master and companion is sent PCM (not DSD).

So, how bad does it sound?

[daniel.avasilichioaei]

Basically speaking, the master plays a file (32-bit or DSD or MP3 etc.), and all companions another file (24-bit).
Since there is no practical way to make all amplifiers play the same file (32-bit or DSD or MP3 etc.), I have no way of knowing what the impact is and if it is noticeable...

In addition, the 32-bit -> 24-bit (in master) conversion, followed by the 24-bit -> 40-bit internal native format (in companion) conversion, can't be a good thing. Whether the effect is felt or not is another matter.

[Redwingnine]

So, here's the next question: Within the master, is the file resolved as 32 bit, OR, is it converted to 24 bit during playback?

[ada]

While not being expert in the domain (and hard to know because it is a Devialet property) I think there are a few things here mixed. (I want to believe in that way).

First of all, I am not sure if there is a 32bit recorded audio. Golden standard for digital recording is 24bit 48kHz but there could be some recorded at 24bit 96kHz and I am sure there are some super/uber obsessive ones recorded at 24bit 192kHz. Saying all these, you should take these into account:

1. Finding the hardware that records in 32bit
2. Finding a DAW supports and performs at 32bit
3. Finding all the plugins used for mastering and mixing supports 32bit

Let's assume that song exists (still hard to believe it didn't pass through any 24bit path until now but it is fine), as it stands 24bit audio is capable of 144db dynamic range and 32bit audio is capable of 1528db dynamic range but on the other side, the human ear's dynamic range is around 120 dB at the peak! (assuming you hear 20Hz to 20kHz with no lose!)

Without a matter of 32bits or 24bits and how they are presented, our ears will be limited to way less than 24bits in the best case scenario. Then you can ask, why Devialet is using 40bit internal system and I think that was already discussed in this forum, that 40bit data doesn't need to contain only the song information but more about configurations, real time flags etc to be taken into account.

All in all, I think what should be happening in a dual mono setup is: Source (32bit) -> Devialet USB -> initial internal default conversion to 24bits (enough for our ears) -> Master and Slaves and then both units process the same audio internally. But again, that's what I want to believe and if that's not the case it sounds like faulty design and both channels process different data!

[thumb5]

Good post @ada. To pick up one point you mentioned: I believe the 40-bit format Devialet uses is floating-point, and it is probably used to avoid overflow or underflow in DSP operations (in other words, it is not meant to provide 40-bit audio resolution throughout the audio path). In any case the 40-bit floating-point format would "only" have 31 or 32 bits of precision.

Also, the ADH stage is based on a PCM1792 DAC which has 24-bit resolution, so whatever goes into the Devialet will presumably be at 24-bit resolution when it hits the ADH in both master and companion. And as you said, that should be enough resolution for all practical purposes when human hearing is involved.

In addition, the 32-bit -> 24-bit (in master) conversion, followed by the 24-bit -> 40-bit internal native format (in companion) conversion, can't be a good thing.

I don't see why it should be a problem. If the ADH stage operates at 24-bit resolution, you have to have the truncation from 32 to 24 bits at some point. Extending the data from 24 to 40 bits won't lose any further precision, and it will presumably be followed by truncation back to 24 bits for the ADH stage in the companion(s). The point is, I think, that the Devialet is inherently limited to 24-bit resolution by the ADH DAC so the fact that it will notionally accept and process 32-bit input doesn't mean that it is "giving you" 32-bit resolution at the analog output whether you're talking about a single stereo unit, a master or a companion.

[Jean-Marie]

Good post @ada.  To pick up one point you mentioned: I believe the 40-bit format Devialet uses is floating-point, and it is probably used to avoid overflow or underflow in DSP operations (in other words, it is not meant to provide 40-bit audio resolution throughout the audio path).  In any case the 40-bit floating-point format would "only" have 31 or 32 bits of precision.

Also, the ADH stage is based on a PCM1792 DAC which has 24-bit resolution, so whatever goes into the Devialet will presumably be at 24-bit resolution when it hits the ADH in both master and companion.  And as you said, that should be enough resolution for all practical purposes when human hearing is involved.

In addition, the 32-bit -> 24-bit (in master) conversion, followed by the 24-bit -> 40-bit internal native format (in companion) conversion, can't be a good thing.

I don't see why it should be a problem.  If the ADH stage operates at 24-bit resolution, you have to have the truncation from 32 to 24 bits at some point.  Extending the data from 24 to 40 bits won't lose any further precision, and it will presumably be followed by truncation back to 24 bits for the ADH stage in the companion(s).  The point is, I think, that the Devialet is inherently limited to 24-bit resolution by the ADH DAC so the fact that it will notionally accept and process 32-bit input doesn't mean that it is "giving you" 32-bit resolution at the analog output whether you're talking about a single stereo unit, a master or a companion.

Correct, the 40 bits is floating point. It is the same as IEEE Standard 754/854 (which is 32 bits float) but with 8 more bits for the mantissa. This allows to have 31 bits of mantissa instead of 23, therefore fitting a full 24bits without any loss and benefiting from the exponent part for functions like volume in a way that do not loose any definition whatever is your volume (+- 762 dB   )

[daniel.avasilichioaei]

Using 32-bit can be useful for those who do software upsampling before sending the audio signal to the amplifier.
I don't see the point of having a 32-bit audio input, but it doesn't use the signal as such (but converted to 24-bit).

[Delija]

Basically, 32-bit is only useful in music production software (DAW) - for digital audio data processing. E.g. Merging Technologies Pyramix.

[ada]

Using 32-bit can be useful for those who do software upsampling before sending the audio signal to the amplifier.
I don't see the point of having a 32-bit audio input, but it doesn't use the signal as such (but converted to 24-bit).

That's the thing, 32bit audio is different then 32bit float point. When you send 32bit audio, you are still sending 21-22bits of audio there and you can't use 32bit audio in practice because that will cause damage to your amp, speaker or if they are able to handle 1000dB you won't hear anything or you will damage your hearing permanently because of the clipping that supposed to happen and block the audio to reach to higher dB won't be clipped and will reach to your ears. (A good reference: http://www.cochlea.org/en/hear/human-auditory-range) That's why the headroom that we are talking in the music should still be in the limits of the nature.

All in all, I wouldn't stress myself about this at all, Devialet is having a great dynamic range, THD+N numbers, probably one of the best you can find in the world. Just keep enjoying it, I generally criticize Devialet a lot but this domain is something they have nailed down perfectly! Any 32bit DAC that says they process 32bit audio signal, it can't be true.

One last note about up-scaling: When you upscale 24bit audio to 32bit audio, probably you end up a lot of empty bits there, there is no way to increase the dynamic range from narrower to wider. That's the limitation of the universe both for sound and visuals. You can use some machine learning models to fill that part and scale it up with some prediction algorithms but that is reproduction of the music and probably won't same as the original if it sounds any good and again be careful about what you are doing because if there is a DAC process 32bits audio I don't think there is any speaker capable of playing that at all.