XXhighend digital filters question

[gsbrva]

Hi there, I'm new to the forum.  I finally took the plunge and decided on XXHE as my player.  Thanks, Peter! 

Great player, but I have a few questions :-)

My question is:  What bit depth are the upsampling modes running.  Is it fair to assume that it is interpolating 16 bit data to 24 bits and that any dac less than 24 bits will be truncating the output?

Do I understand arc prediction correctly?  Is it keeping all the original data points and adding new samples in between, but not post filtering the final product? 

Has anyone used a hardware upsampler/digital filter after arc prediction and liked the sound?  I can only output x4 from the Hiface Evo and I can definitely hear the image in NOS.  I tried the 3 filters in XX and the sound differences are major.

Thanks,
Greg

[PeterSt]

Hi there Greg - thanks for the nice words.

Very good questions ...

Is it fair to assume that it is interpolating 16 bit data to 24 bits and that any dac less than 24 bits will be truncating the output?

Almost correct;
With a DAC that is 16 bit capable, no "upsampling" to a higher number of bits take place (this is taken care of). Only when the DAC has more than 16 bits but less than 24, truncation happens.
In addition, the output really is 32 bits for DACs denoted as "DAC Needs" = 32 bits (see Setting). In this case truncation to 24 bits happen when the "DAC Is" (see Setting) is 24. All nothing much different from how it happens in-DAC where most often the interface is 32 bits, but the DAC can process 24 only.
Notice that with the HiFace (Evo ?) you can set the DAC Needs at either 24 or 32. 24 should sound better (don't ask). Be careful at trying it (analogue volume low).

Do I understand arc prediction correctly?  Is it keeping all the original data points and adding new samples in between, but not post filtering the final product?

100%. 

I tried the 3 filters in XX and the sound differences are major.

... with the notice that the other two really are no good. Ok, to my standards (and Linear Interpolation just *is* no good ).

Regards,
Peter

[gsbrva]

I did notice XX sounds better on the 24 bit setting with the EVO.  How should one set a 16 bit dac with an EV0 to avoid data truncation?   Is it possible to hold the bit depth to 16 and still output upsampled to the EVO driver?  The settings don't seem to permit that.   

I know I need to upgrade to 24bit :-)  The chips are sitting on the workbench calling me...... 
   

Greg

[PeterSt]

That could be a nice strange kind of problem; I don't think the HiFace (Evo) allows for 16 bit input.

On the other hand, nobody tells you to upsample. So, leave that out and nothing strange happens (and just set DAC Needs to 24). But now you also won't upsample of course ...

So no, there is no provision for upsampling and leave that at 16 bits - this would be an illegal thing btw. Which ... makes setting DAC Needs to 24 *plus* DAC Is to 24 as illegal. Maybe you won't notice it as "bad", but it would be. Maybe still better than do nothing ?

Peter


PS: If you have a spare minute, please put your relevant stuff in your signature. This is always useful ad juding problems.

[stefanobilliani]

Hi Peter and all ,

the situation in the attacched screenshot , does denotes if all will be going alright? I mean not knowing if the dac really processes 24 or 32 bit ?
Anyway I seem to hear some difference in sound when the DAC is 32 bit 96K , rather tha 24 bit 96k while the dac need is still 32 bit .

Thank you

sb

[PeterSt]

Hi Stefano,

DAC Is really shouldn't make a difference, unless your DAC really *is* 32 bits. But if you don't know, then sure it is not ...
(what about putting your relevant data in you sig ?).

Peter

[stefanobilliani]

In addition, the output really is 32 bits for DACs denoted as "DAC Needs" = 32 bits (see Setting). In this case truncation to 24 bits happen when the "DAC Is" (see Setting) is 24. All nothing much different from how it happens in-DAC where most often the interface is 32 bits, but the DAC can process 24 only.
Notice that with the HiFace (Evo ?) you can set the DAC Needs at either 24 or 32. 24 should sound better (don't ask). Be careful at trying it (analogue volume low).


Peter

I guess in my case , the interface would process 32 bit data , then the dac "only" 24 .

No idea really , if there is truncation or not or ... editing post , yes ...there is . Anyway I will set 24 bit needs and listen .

sb

[gsbrva]

Thanks much for the explanations. 

So when playing redbook and upsampling via arc prediction, is there any advantage to having more than 16bits of dac resolution? 

Greg

     

[PeterSt]

Greg,

So when playing redbook and upsampling via arc prediction, is there any advantage to having more than 16bits of dac resolution?

But of course. I mean, when I said something like "upsampling without adding additional bits to the output data is illegal", something must be beneficial in just doing that (adding bits). Think like this :

Each upsampling step needs one additional bit. So, upsampling from 44.1 to 88.2 now needs 17 bits in the DAC to have it right. When this is not done, two samples are in the place of the earlier one, but there may not be space to add the interpolated volume. Thus, the two samples are the same while in (faked) reality they would not be. The one additional bit now allows for one additional volume step.
This is how the NOS1 allows for 16x upsampling (2x2x2x2), which needs 4 additional bits (thus 20 bits) while it is a 24 bit (NOS) DAC.

In your situation, the 1543 with a HiFace in front of it, will allow you to output 4x upsampling (or 8x with the Evo ?) because XXHE sees that it can do it (once you also denoted the DAC Is as such). But next all is out of (XXHE) control, because somewhere along the lines those 2 additional bits for 4x get chopped off. And again the additional samples will be the same or not in proper balance for their volume.

Clear a little ?

What you could do in your case is denote 4x upsampling, and *not* engaging any filtering (like Arc Prediction). What will happen now is that the samples stay the same on purpose, which is legal again. Why ? because that improper balance is not implied. But ... it is and stays illegal for the reason of your 1543 being NOS and now the THD is huge (think in terms for 30% for certain frequencies).
But this is what you were used to in the first place ...
(and liked).
Btw, I assume your 1543 runs at 176.4 when I talk abou 4x upsampling. Otherwise it won't go at all of course.

Peter

[gsbrva]

Yes, my limit of upsampling is 176.4 right now.  With the 32 bit data frames out of the evo, this is all the 1545A should be able to handle.  I haven't tried higher.  Chip sheet says 18mhz.  32x2x176.4k=11.2mhz, so I probably cannot double it further right now without changing some hardware.

Can you clarify?

The problem of not having an intermediate bit to interpolate to should not occur very often, right?  I should think at high levels near 0dbm, that the precision of the 16bit sample is adequate for any amount of interpolation needed.  The issue to my mind is at low signal levels and high frequencies, say below -75db or so.  I can see how there might be no available "bits" to interpolate to.  However, would the resulting signal be more distorted  than redbook is in the first place?  It's not obvious, since the recordings are defined by the limit of the 16 bit resolution.  After all, near the noise threshold, the sample of any redbook recording is the gross approximation of a 2 or 4 bit sample.  So yes, nothing to interpolate to, but nothing much to start with either.   

Perhaps the issue is that upsampling on 16bit dac would yield a less linear filter, since the distortion/noise image due to the sample rate would be spread down to a lower frequency at lower levels than at high levels when compared to 24bit?  Would this be correct, since there are some repeating bits due to truncation at low levels/high frequency?

Just a thought
Greg 

[gsbrva]

After all, near the noise threshold, the sample of any redbook recording is the gross approximation of a 2 or 4 bit sample.

That didn't make sense.  I should have said a few bits sample instead of 2 or 4 - don't post after the bar is the rule

[PeterSt]

The problem of not having an intermediate bit to interpolate to should not occur very often, right?

This happens all over;
The 16 bit resolution of course is just "utilized". So, with a max level of decimal 32767 (for plus, for minus it's 32768) steps, all those steps are used by guarantee. Think transients. So, the one transient may skip a few of those levels (like 12, 27, 50 could be utilized), but the other transien twill use just all those and in a fashion like 10,11,12,13,14. So, between 12 and 27 there's something to interplolate (1x), but between 12 and 13 there's nothing. So or you'd have two samples at 12 and 12, or at 13 and 13.
Besides that, to have it accurately interpolated between 12 and 27 with just one values in between (at 2x upsampling) additionally is also a bad thing. With 8 bits more (24 instead of 16) there's 256 values in between to choose from ...

Peter

[gsbrva]

Thanks, that explanation made sense.  I was not thinking of it the right way.  It's not the same problem as resolution near the zero crossing while recording.  The way I'm understanding it now, my 16 bit dac will have noise artifacts to a lower frequency because the interpolated values are truncated at all signal levels.  So there is not a way to get the benefits of filterless NOS w/ arc prediction at 16 bits.  Sigh.......

I surrender :-)  I'll concentrate on getting 24 bit working instead of tinkering with this 16 bit chip.