Mastering in Pure DSD - Mixing and EQ with no PCM

For years I’ve been told it was not possible to Mix or EQ in DSD. We’ve discussed that in the PS Audio Forums multiple times. But it is possible! Eudora Records and Gonzalo Noqué are doing it. I wanted to find out how and I wanted to hear the difference it made in the sound quality. Gonzalo was willing to explain his process (in some detail) and also to provide music file samples from three of his albums. So, in collaboration with Gonzalo, we pulled together this article for Positive Feedback:

Mixing in DSD - No PCM Allowed

Gonzalo agreed to make available for free download three pairs of sample files from three different Eudora albums, including a string quartet, a violin concerto, and a violin and harpsichord duo. In each file, the data comes from the same DSD256 microphone tracking channels mixed either a) via a DXD project in Pyramix or b) via a Pure DSD modulation in HQPlayer Pro. The link to download these sample files is in the article.

Three Eudora albums in which multiple microphone tracking channels are captured in DSD256 and then mixed and EQ’d in post-production completely in the DSD domain, no PCM:

@Paul, @cookie, @tedsmith, I’d love to get your thoughts and input on the sound quality of the sample files, on Gonzalo’s process, and whether other perfectionist labels might be able to effectively use this same process with certain of their releases. Granted, not all recordings will be amenable to this processing. But certainly many single performer and small ensemble recordings can benefit from this same processing when multiple microphone channels are needed.

I’ve been stating for quite a long time that you can edit in DSD, the problem is that people all have a different idea what DSD is and what PCM is.

To edit DSD well you DON’T downsample it. You edit it just like you would PCM but at it’s original rate and you let the samples widen so you don’t lose precision (which is what you should do in PCM as well.). Then just where you’d dither PCM to narrow the samples to the width appropriate for delivery to the consumer, you should “dither” DSD to one bit for final delivery. But for DSD narrowing to one bit uses a sigma delta modulator instead of simply adding a small amount of noise which is what you do in PCM.

The confusion comes when people seem to get the idea that you should keep things in one bit during editing. As a PCM example, you don’t keep material intended for a CD in 16 bits during editing… Letting the sample width grow doesn’t lose any accuracy or “goodness” of DSD. Wider is fine. It’s when you are done editing that you go back to one bit (or 16 bits in CD PCM.) The “DSD goodness” comes from the high sample rate and skipping the steep antialiasing filters needed in PCM, etc.

10 Likes

@tedsmith, all good. And, yes, we’re really talking about PDM which can be many bits wide. The single-bit variant of PDM is what Sony chose to call DSD for marketing purposes. So, for simplicity of communication, I’ve stuck with that. The main point from my listening experiments, and what I hear in the downloadable sample files for which a link is provided in the article, is that keeping the data in PDM and not making a PCM conversion (as done when converting in Pyramix to DXD for editing) results in a cleaner, nicer sounding final product.

Does Octave Records offer any recordings in which the PDM tracking channels have been mixed in the PDM domain and not gone through an interim PCM stage for mixing? I’d love to hear their results from that effort.

And Sony called the editable multi-bit version DSD-Wide;

Yep, absolutely correct. And…?

The point I was trying to make was don’t lower the sample rate and you don’t have to worry about anything. DSD can widen to multibit PDM or multibit PCM if needed, it doesn’t matter. When done editing then go back to one bit at that high sample rate. The point is stay away from “low rate” PCM. When going to multibit, always staying in the original sample rate is what matters.

I’m not the one to ask about anything Octave…

2 Likes

Thanks, Ted.

Can’t DSD be used like an analog multitrack where it’s just converted to analog for mixing? I imagine punch ins would just occupy another track.

Yes, and that’s what I used to prefer. Just treat your DSD like tape, it’s the archive format.

However, when people don’t lower sample rates to edit and use quality software SDMs (instead of the SDM in their A/D) I think we’ll be better off. We’re getting closer.

4 Likes

(Replying to your comment but most words here are written for others to read – I know you know all this, and you may well have corrections or clarifications to add :slight_smile: )

100% this. There’s a fascinating conversation to be had around this and very few people seem to be willing to have it. I only had my eyes opened when the first descriptions of how the DS DAC performs upsampling came out and it nearly broke my head.

All digital audio plots amplitude versus time. Not just PCM, but DSD and every other variant of PDM as well. Any audio transformation you can perform on PCM you can perform on the others, and if you provide sufficient precision you can do it losslessly, or near enough to not matter. DSD purists think that anything with more than a single bit per sample must by definition be inferior – but in the same way that jitter only matters at the point of A/D/A conversion, likewise the 1-bit-ness of DSD is only an advantage for physics reasons related to the construction of audio electronics.

As you point out, Ted, in the pure information domain the key is the super high sample rate, and I would add to that the noise shaping which mitigates the 6dB SNR of direct 1-bit sampling. Compact Disc and its “PCM” brethren use lots of bits to provide high SNR across the entire sampled frequency spectrum, which is a narrow but deep approach. DSD uses a single bit but at least 64x the sample rate of CD – wide but shallow – with sigma delta modulation to provide a high SNR in the audible range and pure noise in the upper frequencies.

But they are fundamentally both plots of a waveform. Waveforms can be scaled, added/subtracted and transformed in a myriad of other ways regardless of how many bits per sample they have. You do need to have more bits in your answers than what you started with to avoid losing your signal in quantisation noise, but whatever it is that allows DSD to sound so good is not lost in this process, especially if you’re merely adjusting levels and mixing multiple tracks together.

And why does DSD seem to be able to capture and hold better sound than low-rate PCM? I don’t think you mentioned it in this thread but you’ve made the point elsewhere: it’s the filters. Because the Nyquist cut-off (half the sampling frequency) is so far above the audible band we can use far gentler filters which avoid roll-off and group delay in the audible band; filters which are needed during recording, in the digital realm prior to SDM processing back to 1-bit, and in playback equipment to let the audio out without all the ultrasonic noise.

DXD is both wide and deep (32-bits). You can bring your DSD recordings into that space, transform them without losing any of the benefit of those shallow filters, mix them directly with PCM content if you need to, and then run an SDM process over the whole lot (which includes another low-pass filter to eliminate the ultrasonic component of the DSD waveforms you mixed in) to produce a final DSD output.

(Edit: the paragraph above is wrong because I forgot that DXD downsamples to 384kHz. So it’s good but not as good as what Ted does at full sample rate inside the DS DAC.)

DSD doesn’t lose its essential quality as a result of being processed through what is indistinguishable from a PCM system, so long as the high sample rate is maintained the whole way through.

Hmmm, okay… I’ll admit to being the non-technical listener who is getting confused. So let me just ask two simple questions: 1) is DXD a high sample rate PCM variant? 2) When Pyramix runs a DXD project using DSD256 tracking channels as the input, should the output to DSD from that DXD process be indistinguishable, that DSD has not lose “its essential quality” as a result of this processing?

If so, have you listened to the comparative sample files from this process to which I’ve given a download link in the article? (This is really an opportunity to listen and compare.)

I’ve stated in the article what I hear, and the DXD-processed file out of Pyramix is consistently quite distinguishable from the non-PCM processed file out of HQPlayer Pro (which keeps the file in the DSD domain throughout the mixing process). To my ear something is being changed by the DXD conversion.

1 Like

Oh I’m sorry – I got my definition of DXD wrong because I was careless. DXD is only 384kHz. So it violates the rule of keeping the DSD sample rate of at least 2.8224MHz.

2 Likes

DXD is “low rate” in that it’s 8 times lower than DSD64 (352.8kHz). I’m talking about editing DSD64 at it’s nominal rate of 2.8224MHz, DSD128 at 5.6448MHz and DSD256 at 11.2896MHz. That’s an entierly different kettle of fish than DXD.

6 Likes

Just on this, I can believe that HQPlayer keeps the sample rate at full DSD rate, but processing cannot be done in the 1-bit space. So it’s going to go through a step that involves multi-bit sample values, which is indistinguishable from PCM. If you didn’t run into materials science limitations you could build an R2R PCM ladder DAC and play that multi-bit variant through it without modification. So is it still DSD at that point? Or is it PCM?

I would say it’s a meaningless distinction to attempt to make, because the original DSD waveform is still perfectly intact inside that higher-resolution space.

Technically noise shaped data isn’t PCM, simply because PCM is completely specified by the sample rate and the sample width - pulse code modulation.

Also technically very high-rate wide samples aren’t DSD because DSD is pretty specific about one bit.

Still the high-rate wide sample have the benefits of DSD’s high sample rate and PCM’s precision and is a superset of each. You lose no information widening DSD or upsampling PCM.

4 Likes

Ah, okay. But are there any tools currently available that mastering engineers have access to use as @tedsmith and @dvorak have suggested for processing in 11.2896MHz PCM?

Look, I’m just a music listener who has a passion for the best sound quality reproduction I can get an opportunity to hear. The point I’m trying to get to in my article is: WHY are folks mixing in DXD when there is an opportunity to mix in the DSD domain at the full 11.2896MHz rate of their DSD256 tracking channels? The notion that “you cannot mix in DSD” in now a myth. People are currently doing so. They’ve explained how they are doing it. They are using currently available “off the shelf” tools (e.g., HQPlayer Pro). And the resulting sound quality is better than when DXD processed, in my opinion as a listener.

Sure, if you need to add reverb or any variety of other additional things that a DAW like Pyramix allows to be done in PCM, you’ll need to go to PCM. But, if all you need to do is mix and EQ DSD256 tracking channels, you can do that entirely in the DSD domain today.

If you’re thinking in terms of product standards and interoperability specs… sure. But at a more abstract level they are both just regular coordinate systems for amplitude vs time of some bandwidth-limited signal. We happen to be talking about human-perceptible audio as the application so we’re really concerned about high SNR, flat frequency response and consistent phase up to about 20kHz.

Your own product (or at least the marketing for it) claims to keep DSD as DSD right throughout its internal processing. But in reality you go to at least 50 bits deep even for DSD. I don’t see an inconsistency in that, because the waveform of the original DSD signal maintains its shape until you do a new SDM pass. But it was a brain-exploding moment to learn how you map PCM and DSD inputs into that exact same sample space using the exact same upsampling and filtering. Why is that even possible? Because the sonic characteristics of DSD do not depend on it being represented by single-bit samples at all times. It’s the frequency content of DSD which carries the sound in the abstract, and the single-bit detail helps with making a great DAC from physical components.

DSD is a special case of the general form of PCM, if you think of PCM in an information theory sense rather than an engineering spec. The waveform we represent in DSD is artificially created via SDM, true, but the outcome is that in the sub-20kHz portion of it we have encoded a high SNR audio signal.

Literally any transformation you could do on “PCM” you can do on “DSD” if only you have the computing power to do it at DSD sample rates and are willing to accept that mapping your waveform into a multi-bit sample space doesn’t violate the rules of what you want to consider “DSD”. Reverb is no exception. There are no exceptions. The algorithms are identical. It’s just a dramatically higher sample rate and at the end you’ll need to run another SDM pass to get back to a 1-bit version.

The first part of what Dvorak says is correct, as far as I know. HQPlayer Pro is using a wide variant of PDM. And to Ted’s point, this is why Jussi prefers to use the terms “Pulse Duration Modulation or Pulse Width Modulation (PWM)” rather than PDM. But, I think he would also say that this is not PCM within our typical understanding of what we mean by PCM.

So, I come back to my questions from my last post here: 1) What tool is out there to do what you are describing with ultra high rate PCM? 2) Have you listened to the sample files comparing DXD processed files to Pure DSD?

Just trying to get back to real world options available to mastering engineers today.

I can’t speak to the tools widely available for editing at these rates. But the article mentions that the HQ Player related stull is good for SDM, etc. but not (at this time) for editing like people are used to. Conversely the standard tools aren’t quite up to snuff yet.

I was talking about where they are headed (where many are headed in a little time.)

2 Likes