Audio Myth - "DSD Provides a direct stream from A/D to D/A."

[Hiro]

Yes, multi-bit SDM. The only way to get DSD out of them is by noise-inducing quantisation. The only way to get PCM is by downsampling. Pick your poison.

 

Recorders such as the new TASCAM DA-3000 use 1-bit, 5.6MHz D-S Modulators per channel so when they record to DSD no digital processing occurs. Ditto for KORG recorders. New AKM ADC chips will also be able to record to DSD (up to 11.2MHz, i.e. their native sample rate), not sure if it will be derived from 1-bit or 5-bit D-S modulator though...

[mansr]

Recorders such as the new TASCAM DA-3000 use 1-bit, 5.6MHz D-S Modulators per channel so when they record to DSD no digital processing occurs. Ditto for KORG recorders.

 

Then they cannot offer the same noise levels as state of the art multi-bit converters.

[Hiro]

Then they cannot offer the same noise levels as state of the art multi-bit converters.

 

Grimm AD1, for example, offers -129dB of THD. Do you know of any multi-bit SDM ADC that can match this performance? BTW, with DSD256 modulators you can reportedly get better than 24bit performance.

 

And I've seen the output of 5-bit SDM converters, and their noise performance looked very similar to DSD converters.

[Miska]

Then they cannot offer the same noise levels as state of the art multi-bit converters.

 

Not really. Better than you get from any real world analog source apart from state of the art signal generator used for test signals. Those recorders use TI's PCM4202 chip, I have pile of those chips and have been using it for a quite a while together with Cirrus CS4398 DAC chip in Direct DSD mode. You need to consider and measure the entire chain, so don't get stuck to digital domain but instead measure the performance at DAC output. That's what I've been doing. It's not all about plain theory, but it's also about the implementation in practice, in the real world.

 

Main problem with multi-bit SDM ADCs that convert to PCM are the bad decimation filters with at best about -120 dB stop band attenuation and the overall filter design. Plus bunch of issues you get with multi-bit design. If you use the full multi-bit space for noise shaping you don't win anything in terms of noise levels. If you don't, the converter becomes 1-bit at any sample values that are at LSB level or below (for 6-bit that means sample values below -36 dBFS).

 

I'm all fine with introducing multi-bit end-to-end SDM file format, but until such becomes reality DSD offers the way that has leas amount of damaging processing involved when considering end-to-end chain.

 

Problem with PCM are the correlated noise levels (images) above fs/2, at every multiple of fs (the DAC runs at) and intermodulation of the negative and positive images around each multiple of the fs.

[Miska]

Grimm AD1, for example, offers -129dB of THD. Do you know of any multi-bit SDM ADC that can match this performance? BTW, with DSD256 modulators you can reportedly get better than 24bit performance.

 

And it's no COTS chip-crap, it's a proper discrete converter implementation.

[Hiro]

And it's no COTS chip-crap, it's a proper discrete converter implementation.

 

One can only wonder how a converter of similar build quality, but using 5.6 million samples per second would perform. I don't even want to think about an 11.2 MHz one

[The Computer Audiophile]

One can only wonder how a converter of similar build quality, but using 5.6 million samples per second would perform. I don't even want to think about an 11.2 MHz one

OK, I'm wondering why I would need 5.6 million or 11.2 million samples per second.

[Hiro]

OK, I'm wondering why I would need 5.6 million or 11.2 million samples per second.

 

Depending on Delta-Sigma Modulator design, of course, using the higher rates can give better noise performance, and makes filtering of out-of-band noise easier.

[mansr]

OK, I'm wondering why I would need 5.6 million or 11.2 million samples per second.

 

To make up for the lack of precision in each sample.

[徐中銳]

From Q&A with Andreas Koch :

Double rate DSD pushes the noise shaper up in the frequency domain as shown in Fig. 2 below:

 

That is most interesting for recording and post production when the intent is to release the product in DSD, because DSD2x gives the extra headroom that recording engineers need in order to record and edit without causing any degradation when releasing their final product in single rate DSD.

 

It may also be interesting for hobbyists who for instance want to archive their analog music library to a digital format. In such applications you may not care about the extra storage space that is required and you certainly wouldn’t be bothered with bandwidth bottlenecks when sending DSD2x files through the internet.

 

However, as a delivery format from studio to end user single rate DSD seems to offer an optimal combination of sound quality, bandwidth and storage space.

[tailspn]

To make up for the lack of precision in each sample.

 

There is no precision in each DSD sample, there are no samples. There are no digital arithmetic values represented in a DSD bit stream as there are in PCM words. There's simply the relative level of the modulating analog level, expressed in the bit density. And there's lots of precision in the DSD format, since there's not the overhead of carrying forward the redundant constant data from word to word as there is in PCM.

[徐中銳]

Last week :

Extracted from Koch's DSD - the new Addiction :

The term Direct Stream Digital (DSD) was coined by Sony and Philips when they jointly launched the SACD format. It is nothing else than processed Delta-Sigma modulation first developed by Philips in the 1970’s. Its first wide market entry was not until later in the 1980’s when it was used as an intermediate format inside A/D and D/A converter chips.

 

Fig. 1:

 

Figure 1 shows how an analog source is converted to digital PCM through the A/D converter and then back again to analog via the D/A converter. The A/D internally contains 2 distinct processes:

 

1. Delta-Sigma modulation: the analog signal is converted directly to DSD with a very high sampling rate. Various algorithms are in use depending on the application and required fidelity. They can generate 1-bit DSD or multibit DSD oversampled at 64x or 128x compared to regular CD rate.
   
2. Decimation filter: the DSD signal from the previous step is downsampled and converted to PCM. Word length is increased (for instance 16 or 24 bits) and sample rate reduced to CD rate or a low multiple of it for high resolution PCM formats.
   

The D/A process is very similar where:

 

1. the PCM signal is upconverted to a much higher sample rate.
   
2. then converted to DSD via the Delta-Sigma modulator (to reduce word length)
   
3. then converted to analog.
   

This technology was chosen because of its improved linearity and consistent quality behavior across physical components' date=' as most of the heavy duty signal processing was shifted to the digital domain where it was not susceptible to variability of electronic components. It was quickly adopted in most converter systems and we can say that since about the late 1980’s we have been listening to some form of DSD without even knowing it.

 

As science progressed as well as our experience with digital audio, we started to realize that the algorithms for the DSD-to-PCM and PCM-to-DSD conversions can have a profound impact on the sonic performance when they are developed according to classic formulas. These are relatively complicated algorithms and they introduced a new phenomenon that we describe as “digital sound” or ringing effects. Hence the motivation by the engineering teams of Sony and Philips to remove these steps altogether from the conversions between analog and digital. This simplified DSD path that bypasses the PCM path is shown in Fig. 1 above. As is usually the case most simplifications in the signal path lead to sonic improvements and so it didn’t come as a surprise when first listening tests were so astonishing that this format was considered as an archiving format for recording studios. That alone says something about its sonic fidelity. At the time no recording studio was even considering using any PCM format to archive its analog recordings.[/quote']

[rredline]

The main advantage of DSD is not the so called AD-DA direct philosophy, it is the elimination of decimation filter (which is kind of "lossy" process) in the ADC stage.

[Paul R]

DSD is mathematically difficult to work with. Even a seemingly trivial operation like splicing two streams has issues. Consider the simplest possible case of splicing two streams of silence. In DSD, silence is encoded as alternating ones and zeros, 101010. If we splice two such streams without due care, we may end up with two consecutive ones or zeros, 1010110101, which amounts to a slight blip in the continuous waveform. In general, switching from one DSD stream to another easily results in pops or chirps.

 

Similar effects can be seen with PCM where a sudden level change is also equivalent to a wideband burst. With PCM it is, however, trivial to pick a splice point where the levels match or to smooth the signal around the splice point.

 

More "advanced" operations like volume adjustment or mixing are impossible to perform without a requantisation step. Although it is possible without going all the way to a PCM representation, the effect is the same as if one had done so: another round of quantisation noise is added.

 

As for things like EQ or reverb, don't even think about it.

 

Going back to the article, the main point John Siau makes is that the most accurate ADCs and DACs today are all multi-bit designs. From there, conversion to 1-bit DSD is a step back no matter how you look at it.

 

Mathematically, DSD is much simpler than PCM. Not sure why you would think otherwise, except that PCM has become a standard because there just wasn't enough computer storage and processing power to do DSD manipulations cleanly - and at a reasonable cost - 30 years ago.

 

Manipulating DSD data is very efficient for a computer too, by the way. There is just a whole whale of a lot of it to manipulate.

 

Despite media propaganda, there really isn't anything that I can see inherently *better* between DSD and PCM. Both have their own unique kinds of problems, which are solved differently.

 

But given how much easier DSD playback equipment is to build *well* - I tend to prefer DSD for playback. Economic reasons are not an insignificant factor there.

 

Also, "a step backwards" is misleading. Of course it is technically stepping back to a one bit pulse density density format, but that format can offer advantages at playback time that are enormous expensive to create with a PCM only DAC. I do get that the opposite is true in the engineer's studio. At least, today.

 

-Paul

[Paul R]

OK, I'm wondering why I would need 5.6 million or 11.2 million samples per second.

 

You wouldn't - but DSD is not made up of "samples". There is no instant value for any discrete time period in a DSD data stream.

 

The higher the data rate, the closer to a continuously varying analog signal you get, with DSD. PCM operates on a complete different principle, though in a roughly analogous way, the same is true with high sample rate PCM. Not exactly, but some of the same operators come into effect.

 

-Paul

[Synfreak]

Just a a side note:

 

Many of the Plug-Ins which are used to "shape" the sound *) of a given PCM recording use internal oversampling.

 

So one always ends up with multiples of added (anti-aliasing) filters and lots of decimation processes ...

 

 

*) i.e.: compressors, limiters, EQs, ...

[Maldur]

The main advantage of DSD is not the so called AD-DA direct philosophy, it is the elimination of decimation filter (which is kind of "lossy" process) in the ADC stage.

 

...And interpolation filter in DA stage .

[mansr]

Mathematically, DSD is much simpler than PCM. Not sure why you would think otherwise, except that PCM has become a standard because there just wasn't enough computer storage and processing power to do DSD manipulations cleanly - and at a reasonable cost - 30 years ago.

 

Being mathematically simple doesn't make it easy to work with. As you said yourself right there, it takes tremendous computing power.

 

Despite media propaganda, there really isn't anything that I can see inherently *better* between DSD and PCM. Both have their own unique kinds of problems, which are solved differently.

 

This I agree with.

 

But given how much easier DSD playback equipment is to build *well* - I tend to prefer DSD for playback. Economic reasons are not an insignificant factor there.

 

Which is why modern PCM DACs all convert to a high-rate low-bit format internally.

 

At least for home use, one should also consider that PCM offers the ability to apply digital room correction filters. Unless your listening room has studio-quality acoustic treatments, this alone will make more of a difference than you'll ever hear between PCM and DSD.

[Miska]

Which is why modern PCM DACs all convert to a high-rate low-bit format internally.

 

Unfortunately all the in-chip implementations do it poorly. Doing it well just requires hefty amount of computation power and putting that kind of computing power into $10 mixed-signal chip that doesn't have cooling has number of challenges. Not least because you'd also want to avoid polluting the analog side.

 

That's one reason why I took all the processing out and designed the DSC1 that is just a discrete SDM-only DAC.

 

At least for home use, one should also consider that PCM offers the ability to apply digital room correction filters. Unless your listening room has studio-quality acoustic treatments, this alone will make more of a difference than you'll ever hear between PCM and DSD.

 

I'm doing digital room correction in real time for DSD and while doing so also upsampling DSD to higher rate for increased dynamic range.

[mansr]

Unfortunately all the in-chip implementations do it poorly. Doing it well just requires hefty amount of computation power and putting that kind of computing power into $10 mixed-signal chip that doesn't have cooling has number of challenges. Not least because you'd also want to avoid polluting the analog side.

 

Doing multi-bit SDM well is much easier than single-bit.

 

I'm doing digital room correction in real time for DSD and while doing so also upsampling DSD to higher rate for increased dynamic range.

 

That means you're at a minimum requantising the signal, even if not doing a full PCM roundtrip. This adds noise. Of course, any operation on PCM also has a level of inaccuracy, so there's noise/distortion there as well.

[Hiro]

Just a a side note:

 

Many of the Plug-Ins which are used to "shape" the sound *) of a given PCM recording use internal oversampling.

 

So one always ends up with multiples of added (anti-aliasing) filters and lots of decimation processes ...

 

 

*) i.e.: compressors, limiters, EQs, ...

 

So much for NOS PCM.

[Miska]

Doing multi-bit SDM well is much easier than single-bit.

 

It doesn't really make any difference. For example with my modulators I can generate any number of target levels. But it is just a single parameter for the modulator.

 

But for on-chip implementations you don't need to even look at the modulator to see the shortcomings. The first shortcomings are already in the oversampling area.

 

Many properties of the DACs are clearly apparent from the measured output performance.

 

That means you're at a minimum requantising the signal, even if not doing a full PCM roundtrip. This adds noise. Of course, any operation on PCM also has a level of inaccuracy, so there's noise/distortion there as well.

 

If the target container has lower noise than the source, there's practically no change in the noise level. Upsampling as part of the process (for both PCM and DSD) allows target container to have significantly lower noise than the source, ensuring that there's no SNR tradeoff from the processing.

 

In any case you can choose what you want to do. For quite a while I've had two alternative routes, either doing Eq in player software, or doing external digital room correction only for the subwoofer using Anti-Mode 8033S-II.

[audiventory]

Into computer no 1 bit internal format. I.e. DSD processed as multibit value always.

 

During converting 1 bit to multibit information do not lose.

 

From DSP point of view no difference in quality PCM vs. DSD.

 

 

 

"DSD provides a simple and direct digital path between the A/D and D/A."

 

Only if don't process it.

 

 

"DSD is simpler than PCM."

 

Need give definition: what is «simpler»?

 

DAC simpler. Yes.

 

 

"DSD is not PCM."

 

Possibly consider DSD as noise shaped 1-bit PCM.

 

 

 

 

Mathematically, DSD is much simpler than PCM.

 

Math of DSD processing is more complex and demand more computing resources. Due higher sample rates.

 

 

 

Just a a side note:

 

Many of the Plug-Ins which are used to "shape" the sound *) of a given PCM recording use internal oversampling.

 

So one always ends up with multiples of added (anti-aliasing) filters and lots of decimation processes ...

 

 

*) i.e.: compressors, limiters, EQs, ...

 

 

 

Decimation and filters is not so scary as may seem at first look.

 

Mixing process more harmful than intermediate decimating with qualitative filters.

[Hiro]

Main problem with multi-bit SDM ADCs that convert to PCM are the bad decimation filters with at best about -120 dB stop band attenuation and the overall filter design. Plus bunch of issues you get with multi-bit design. If you use the full multi-bit space for noise shaping you don't win anything in terms of noise levels. If you don't, the converter becomes 1-bit at any sample values that are at LSB level or below (for 6-bit that means sample values below -36 dBFS).

 

There you have it, multi-bit SDM demystified.

 

I'm all fine with introducing multi-bit end-to-end SDM file format, but until such becomes reality DSD offers the way that has leas amount of damaging processing involved when considering end-to-end chain.

 

+1

[Hiro]

Need give definition: what is «simpler»?

 

DAC simpler. Yes.

 

ADC too.

 

In general a Direct SDM path cuts out the PCM middleman and the necessary downsampling, upsampling, oversampling processes that PCM has to go through in delta sigma AD/DA's.