ASR Audio Science Review forum YouTube Channel

[Andyman]

18 minutes ago, The Computer Audiophile said:

 

HI Andy, it will probably be easiest to digest this little pieces at a time, at least for me. 

 

I'm listening to a DAC that enables external DSP processing. When it's used, the DAC measures much better and is more linear. This provides a more accurate analog output through DSP. Removing nonlinearities is a good thing. It isn't a matter of mine is more perfect because the built-in filters can be bettered. No filter is perfect. 

 

 

 

 

 

Bringing HQP back into the conversation (it is the paradigm for such discussions and presumably what you are using?)

I occasionally glance at his software thread and all anyone ever asks is "which filter should I use?"

I'l bet you've measured more than one filter. Would be interesting to know if you have noticed a correlation purely between linearity and perceived SQ? And what else, if anything, have you measured that you feel might be notable...

 

[Andyman]

 

 

46 minutes ago, The Computer Audiophile said:

Exactly what my wife told me at breakfast this morning. 

 

Me too...

(...not your wife 😳!)

[opus101]

1 hour ago, pkane2001 said:

 

A reconstruction filter is a mathematical necessity for proper reproduction of frequencies below Nyquist/2.

 

The '/2' here I believe is superfluous - Nyquist is already fs/2.

[danadam]

54 minutes ago, opus101 said:

 

The '/2' here I believe is superfluous - Nyquist is already fs/2.

Maybe he meant Nyquist rate, not Nyquist frequency? :-)

Although there are others who define it as one and the same:

Quote

Fortunately, most authors are careful to define how they are using the terms. In this book, they are both used to mean one-half the sampling rate.

 

[acg]

6 hours ago, PeterSt said:

Somehow this worst thread turned into a best thread ?
I am serious.

 

I was thinking the same thing.

[firedog]

3 hours ago, The Computer Audiophile said:

Exactly what my wife told me at breakfast this morning. 

Sounds like you found the perfect woman: "audiophile wife". Congrats!

[PeterSt]

4 hours ago, pkane2001 said:

Some are "designer" filters meant to add something to the sound

 

Paul, you should start making DACs. This is really the strangest BS 😉 of all times. Really. Undoubtedly unintentional, but it is (again) getting more and more crazy in this thread.

 

Ask Miska why he implemented a filter A, B, C etc. Try to drag out of him that he did that to create a certain sound. I'll bet you he won't say that. He merely would say something along the lines of:

- Distortion shifts to an other location, less harmful to you if you are sensitive to it.

- Filter B is an improvement of filter A. A remains for those who like it.

- Filter C is a general improvement for XYZ.

And no filter is perfect because that can't exist in reconstruction.

 

Happy to lose the bet. I don't care. I have my own filters. 🤗

[manisandher]

7 hours ago, Racerxnet said:

Does HqPlayer help reconstruct the wave form more accurately??

 

7 hours ago, Samuel T Cogley said:

How would you make the comparison?  What would you use as your "accurate" sample?

 

That's exactly what I attempted to do in this thread:

 

Mani.

[PeterSt]

4 hours ago, pkane2001 said:

I suspect most poorly filtered DACs are done so because it was cheaper and easier to implement.

 

There is nothing like "poorer". There's just an enormous difference in approach between e.g. genuine NOS (no filter at all) and Rob Watts with 1M filter taps. Let's say that DSD is somewhere in that spectrum too, working with different properties (like upsampling way more, right into the DSD domain).

 

What you guys seem to have fortotten (in the midst of what should be 10K posts about it, if not way more) is that NOS is accurate towards the time domain and 1M taps is accurate towards the frequency domain. The more "accurate" in either of these ends of the spectrum, the more inaccurate the other end will be.

So it is the skill to be something in the middle (anywhere but not at one of the ends) and make the one the best for your (designer's liking) and the other the least harmful technically.

 

Miska will tell you the same. It is only that my liking is towards the time domain and his liking is towards the frequency domain.

Nice eh ? two blokes ever back starting the same thing, but with different technical aims. Both worked out to its best. Both overruling what happens in-DAC. that preferably being a NOS/filterless DAC because it does nothing in the first place.

[PeterSt]

5 hours ago, Andyman said:

Probably mad Peter also but not entirely sure.

 

How did I do ? ... Not quite "accordingly" I'm afraid.

DSD has a purpose in itself because of the high upsampling rate, and with that additional filtering possibilities. Cookie Marenco claims to record in pure DSD. And yes, she can do that. Albums with that type of recording exist. But if anything needs to be done for "mastering" it happens in PCM. Is this important ? well, sort of. You partly transfer to another domain and it wouldn't be a good thing (at fade-ins and -outs for example, or blending in another voice (track)).

 

99% of DAC's transfer to a (kind of) DSD because they are SDM based. So they already molest (in my definite book). They would be good at playing genuine DSD though, although even that is doubtful (because they will still touch the sound).

 

The only DAC/Chip that would not touch the sound is multibit PCM. That is, if no filter is used in front of it (NOS/Filterless). This has become oldfashioned, and maybe I am to blame. ... A NOS/Filterless DAC which is fed with unfiltered data (meaning : not reconstructed from 16/44.1 which is too few) will be as inaccurate as hell (0.04% THD at best). It will sound like sh*t although ever back a large crowd liked it because, well, it is (more) accurate in the time domain. However, give such a DAC a nice 24/192 file and it will sound as good as can be (24/96 would be quite all right just the same).

 

... So what we do is trying to re-construct that 24/192 file from 16/44.1 as good as we can, and next let the DAC do nothing. Only genuine PCM can do that. Only PCM1704 can do that. Only a discrete (e.g. Total)DAC can do that. The Yggy would be able to do it at 20 bits.

They can NOT do DSD, unless Holo indeed managed to play DSD through PCM discrete ladders (my NOS2 design could).

 

Done.

[opus101]

43 minutes ago, PeterSt said:

What you guys seem to have fortotten (in the midst of what should be 10K posts about it, if not way more) is that NOS is accurate towards the time domain and 1M taps is accurate towards the frequency domain.

 

Bruno Putzeys showed in one of his 'masterclass' presentations that filterless NOS isn't invariant with respect to sample clock phase. Which makes it about as inaccurate in the time domain as its possible to be.

[PeterSt]

1 minute ago, opus101 said:

Which makes it about as inaccurate in the time domain as its possible to be.

 

That is why I was careful not to state "100%" (in this case represented by "towards" and in my English). Besides, I assume "after upsampling" which in my case always has been to 705.6/768 (except fo the first 6 months of the DAC's life, when it was 352.8/384.

 

Also, don't forget that Bruno is a sheer "DSD" man. He will do everything to prize his own findings (also found back in everything Hypex).

 

I recall that you too are a DSD adherer. So please continue. 🙂

(maybe I'm wrong)

[opus101]

1 minute ago, PeterSt said:

I recall that you too are a DSD adherer. So please continue. 🙂

(maybe I'm wrong)

 

Yeah, you recall wrong. A search on here will confirm. I even got posts deleted from an Australian audio forum because I was perceived as bad for DSD business.

[PeterSt]

32 minutes ago, opus101 said:

Yeah, you recall wrong.

 

Sincere apologies for the accusation. It wasn't meant bad anyway.

 

[opus101]

I didn't take it as an accusation. Its a bit off-topic though don't you think?

 

<afterthought> If tribal loyalties are important then I'd like to disclose I'm a NOS DAC designer. I don't think NOS is bad just I have a preference to employ some kind of analog filtering in conjunction with it. I do hear an improvement when moving from a relatively shallow filter to a steeper one.

[Miska]

2 hours ago, PeterSt said:

There is nothing like "poorer". There's just an enormous difference in approach between e.g. genuine NOS (no filter at all) and Rob Watts with 1M filter taps. Let's say that DSD is somewhere in that spectrum too, working with different properties (like upsampling way more, right into the DSD domain).

 

All Chord DACs are SDM DACs...

 

2 hours ago, PeterSt said:

What you guys seem to have fortotten (in the midst of what should be 10K posts about it, if not way more) is that NOS is accurate towards the time domain and 1M taps is accurate towards the frequency domain. The more "accurate" in either of these ends of the spectrum, the more inaccurate the other end will be.

 

When you guys talk about NOS DACs, you forget that there are many NOS DSD/SDM DACs out there.

 

For perfect time domain accuracy you need infinite sampling rate. If you run DAC at 705.6k, your timing error is +-709ns. If you run DAC at 44.1k your timing error is +-11.3µs. Because your timing event will never coincide with the sampling event. If you have steep enough analog filter to reach -96 dB by first image frequency, you have 16-bit accuracy. If you have steep enough analog filter to reach -144 dB by first image frequency, you have 24-bit accuracy.

 

The digital filters I use most of the time, reach -240 dB (48-bit accuracy) by first image frequency and produce 11.2 MHz sampling rate. From this rate it is simple to have steep enough analog filter, that doesn't have notable phase error at 20 kHz, to reconstruct the signal accurately.

 

[Miska]

2 hours ago, PeterSt said:

And no filter is perfect because that can't exist in reconstruction.

 

By definition of Nyquist sampling theory, you need filter, because you need to remove all the image frequencies. That is why it is called "reconstruction filter". The original analog waveform exists only after it has been band-limited to fs/2 at least to the precision provided by the sampling word length. So you need 96 dB attenuation for 16-bit and 144 dB attenuation for 24-bit.

 

[Miska]

5 hours ago, opus101 said:

The '/2' here I believe is superfluous - Nyquist is already fs/2.

 

No it is not. Since 1 kHz tone sampled at 44.1 kHz rate will have image at 43.1 kHz (and also at 45.1 and 87.1 kHz and so on and so on). And 20 kHz tone sampled at 44.1 kHz rate will have image at 24.1 kHz (and also at 64.1 kHz and 68.2 kHz and so on and so on). To reconstruct 20 kHz tone from 44.1 kHz 16-bit sampling, you need to get the 24.1 kHz image down to at least -96 dB.

 

[opus101]

6 minutes ago, Miska said:

 

No it is not. Since 1 kHz tone sampled at 44.1 kHz rate will have image at 43.1 kHz (and also at 45.1 and 87.1 kHz and so on and so on). And 20 kHz tone sampled at 44.1 kHz rate will have image at 24.1 kHz (and also at 64.1 kHz and 68.2 kHz and so on and so on). To reconstruct 20 kHz tone from 44.1 kHz 16-bit sampling, you need to get the 24.1 kHz image down to at least -96 dB.

 

Yes I agree with all of that, but the first sentence. I can't see how its relevant to my original contention.

[Miska]

7 hours ago, lucretius said:

No matter what format (PCM vs DSD) you digitize in, ultrasonic artifacts are created. The more bits you have, the lower the noise floor. Noise floor is lowered by roughly 6db for each bit.  DSD has significantly higher quantization noise than PCM, and the noise is much closer to audible frequencies, requiring significantly more sophisticated digital filters, as well as noise-shaping and upsampling algorithms.  When converting from PCM to DSD, more quantization noise and/or quantization errors are added to the recording.

 

Wrong... You cannot compare it this way, because you are not using equal bandwidths. If you look at how 44.1 kHz samples 16-bit PCM looks like at 2.8 MHz bandwidth it is clear that it has massive amount of ultrasonic images throughout the specrum.

 

One of these sampling systems is "Nyquist" sampling where you typically would look only at the frequencies between 0 and fs/2. But you can look also for example at frequencies from -fs x 64 to fs x 64 as well.

 

7 hours ago, lucretius said:

DSD128 has the majority of its quantization noise around 50KHz, which is fairly close to the same frequency as the majority of the quantization noise in a 44.1KHz PCM recording, which is centered around 44.1KHz.

 

No, DSD128 has noise floor beginning to increase at about 50 kHz. Below that it could be for example at -160 dB or lower, exceeding resolution of 24-bit PCM. For practical purposes, DSD128 gives you flat noise floor to 50 kHz. This is also what you get from most modern ADCs with PCM output. PCM output from a modern ADC typically looks just like DSD128 converted to PCM.

 

DSD256 gives you flat noise floor to 100 kHz, DSD512 to 200 kHz, DSD1024 to 400 kHz and DSD2048 to 800 kHz. This flat noise floor means that the quantization noise within this band is below analog noise floor of even best analog electronics. So the noise floor of analog world defines the noise floor.

 

Majority of DSD quantization noise is close to the sampling rate, 5.6 MHz for DSD128 for example.

 

[Miska]

11 minutes ago, opus101 said:

Yes I agree with all of that, but the first sentence. I can't see how its relevant to my original contention.

 

Just that for proper reconstruction you need to be band-limited to fs/2. There's no way around it. And that makes RedBook very tricky because the Nyquist (fs/2) is right at the edge of audio band (in traditional sense of audio band). Meaning that images begin right above fs/2.

 

[opus101]

1 minute ago, Miska said:

 

Just that for proper reconstruction you need to be band-limited to fs/2. There's no way around it. And that makes RedBook very tricky because the Nyquist (fs/2) is right at the edge of audio band (in traditional sense of audio band). Meaning that images begin right above fs/2.

 

Yes I agree with all of that too. But note you wrote 'Nyquist' not 'Nyquist/2' which was what I have been taking issue with.

[Miska]

Just now, opus101 said:

Yes I agree with all of that too. But note you wrote 'Nyquist' not 'Nyquist/2' which was what I have been taking issue with.

 

Ahh, ok. That's probably a typo...

 

[Miska]

10 hours ago, Rexp said:

If the consumer uses the exact PCM playback chain the Artist heard in the mastering studio they will hear what Artist heard. What do you think the consumer will hear if they then change to DSD upsampling playback? So yes I agree with @Andyman

 

Except that in practice nobody will ever have the exact PCM playback chain the Artist had in mastering studio. Already for the reason that there are way too many studios that people could possess the same gear and playback chain. Even less so the same acoustics.

 

[pkane2001]

6 hours ago, opus101 said:

 

The '/2' here I believe is superfluous - Nyquist is already fs/2.

Yes, I actually typed Fs/2 first, then changed it but not not the /2 :)