DSD vs PCM: can you hear the difference?

Are there any (relatively cheap) external devices that convert DSD (HDMI) to high rez PCM (SPDIF)?

 

I just read the article. I am far from an expert in this topic. Could you explain what was so wrong about the article or give a link for an article that you think is right?

 

DSD vs PCM: can you hear the difference?

Between a native recording done in 24/96 and DSD64: Yes.

Between a native recording done in 24/192 and DSD128: Can't say. I do not, presently, have the capability to record in DSD128. (24/192 sounds pretty dang good, though.)

I currently have no preference at 24/96 and DSD64. Neither are high enough resolution for me. The verdict, for me, is not out on 24/192 and DSD128. If DSD128 is just that last little bit less granular at the top when compared to 24/192, DSD128 would then be my preference... ...sonically. For the sake of comprehensive editing, I would be sticking with PCM.

 

The mistake the author makes is to assume that PCM sampling-rate relates to resolution. This is the main misconception people have about PCM sampling. The sampling rate only impacts bandwidth, setting the upper limit to the half-sampling frequency (also, timing accuracy of the samples is a source of error - jitter).

The word length is what sets the resolution in PCM encoding - it dictates the maximum signal to noise ratio you can encode by setting the number of levels it can determine between 0 volts and full-scale. As you turn up the volume during playback you hear down into the recording further - how far that goes is determined by word-length in PCM encoding. 24-bit gives plenty of depth for any recordings you can listen to. 16-bit, not so much.

The author is also confused about the noise up above the audio band on DSD, stating that it is dealt with by noise-shaping when in fact it is there as a result of the noise-shaping.

These fundamental misunderstandings alone render the piece worthless.

 

The Vanity HD Card will do exactly what you want. The DoP version will do DSD too. But apparently this does not matter to you.

 

You have asked this before. And the answer is the same again this time: Nothing will do this directly. But you can get 24/88.2 PCM out of an HDMI de-embedder which is connected to an Oppo 103 which has already down-sampled the DSD to something which is theoretically high-res PCM.

Unfortunately the output of this down sampled PCM sounds inferior to the redbook layer of the same SACDs to my ears. So it largely defeats the purpose in my book.

But you can always just buy yourself a de-embedder and find-out for yourself. These are not expensive.

 

Yeah, they didn't even put all their own releases out on SACD, or even put SACD in all their disc players. No guts, no demand from the top for all divisions to support, no glory. I still have a chip on my shoulder due to their "pure 1 bit stream" baloney lies marketing, though I heard some very nice stuff at a Hi Fi show in Newport.

 

Actually, it's worse, I hadn't read the full article (my bad). It's a sales piece for.... Mojo Audio's - guess what? - non-DSD, non-upsampling new DAC!

Apologies guys, my mistake!

 

Can't tell from the site, but does the DoP trade off the LPCM encoding capabilities?

 

Assuming one is stuck with an Oppo. Higher end (yet cheaper than Oppo) Sony players output 176.4khz from the HDMI, not limited to 88.2. Which theoretically is better than 88.2 (but Oppo claims otherwise).

 

What recordings do you know that need more than 96 dB, or apparently up to 120 dB perceived dynamic range with noise-shaped dither?

Tim

 

Nope. DoP vs high-res PCM transcoding (including multiple filtering options when transcoding) is a switchable option on the card. This switch is repurposed on the DoP version of board from switching some surround option on the regular board. Sorry, but I don't recall exactly which surround option is dropped in the process.

I also think that it converts DSD at 24/176, rather than 24/88 like the Oppos can do internally. But its been a while since I switched it to PCM mode, so I am not certain about this.

 

I know of at least one (a 24/88 LPCM master)... Its downsampled/dithered/downrezzed to 16/44 version was used as an example in an earlier epic thread (now closed, sadly).

 

As I said earlier in this thread, this option clearly will get you the best bang for the buck. Although how this actually sounds is unknown to me. So it might sound great, or it might sound no better than Redbook (like the 24/88 transcoding on the Oppo does to my ears). I can't opine since I haven't heard this setup.

Although with this approach, you don't get a two-box universal player in the process, nor HDCD decoding with this option like you do with the Oppo. The Oppo and the Vanity HD card (DoP Version) is the only way I know of to have a two-box genuine universal disk player which plays virtually every type of media in its fully-decoded native form, including genuine DSD. But if you don't care about native DSD, nor about having HDCDs decoded, nor about DVD-A, then the Sony combo sounds-like a killer deal.

 

Human hearing is a mysterious thing - you don't want to focus on the total dynamic range alone - it's the micro-dynamics that come in to play here.

What is the smallest level change you can hear? I don't mean, adjusting the volume of the whole program - I mean, the fluctuations of the level of each individual element in the recording. How small are the changes before they are not contributing to the subjective listening experience?

When we increase the volume on playback we are effectively exposing the resolution of the recording. Of course, for very quiet signals you can actually hear the difference between 16-bit and 24-bit encoded audio in the noise floor at highly elevated levels. The question is, does that difference translate into an audible difference at normal playback levels.

 

I wasn't kidding when I said virtually everything. I really don't have time to correct all the errors, but picking up on what Mal has already written (small correction, Mal, signals are bipolar -V to +V, not 0V to +V, and are encoded in 2's complement binary values from -32768 to +32767), here are a few more to give you a flavor:

- In the early 1980s, when digital recording became readily available, studios converted from analog to digital to save money. Far from it! DASH machines were hugely expensive, as were digital editors, and digital consoles did not come until later. Many studios/labels were angry with Sony because they thought that they were being ripped off (Sony had a virtual monopoly on much of the digital infrastructure in the early days).

- This new 1-bit technology was achieved by outputting from the monitoring pin on Crystal’s new 1-bit 2.8Mhz Bit Stream DAC chip. The output required is from a 1-bit ADC, not DAC!

- Of course, from the time the SACD was conceived until the time it came to market, DAC chip manufacturers had advanced from 64fs to a higher 128fs sampling rate and from 1-bit to a higher-resolution 5-bit format. Oops. Philips' very first BitStream DAC chip worked at 256 Fs. The choice for 64 Fs was dictated by the disc capacity of DVD media. As in the case of PCM/CD, the release format was state-of-the-art at the time, but it was never the intention for the recording format to be limited in the same way. So, just as we saw the introduction of 20-bit then 24-bit converters, and then higher sampling frequencies (96 kHz, etc.) in PCM production chains for CD, the plan was always to have a higher than 64 Fs sampling frequency on the studio side, but recording and processing 64 Fs in 1998/99 was already a tough challenge!

1-bit converters are a special case of sigma-delta modulators. Matsushita had already used a few-bit design in their MASH converters, and dCS, for one, was using a 5-bit converter, but there has never been a proposal to use anything other than PCM (in some form) or DSD.

- Long before the DVD, SACD, or DSD formats were developed, the Bit Stream DAC chip was introduced to the consumer market as a lower-cost alternative to the significantly more expensive R-2R multi-bit DAC chip. Bit Stream DAC chips have built-in algorithms to convert PCM input to DSD, which is then converted to analog. Once again, the result was a huge cost saving at the expense of fidelity. The main driving force behind the switch to sigma-delta modulation was linearity. 1-bit systems are inherently linear, and do not require complicated precision-matched components. As a consequence, a 1-bit chip is cheaper/easier to make, however, high order sigma-delta systems are more complicated to design, and have to be computer-modeled. From the moment the cost factor was announced, all kinds of 'golden ears' have declared sigma-delta solutions inferior to ladder designs, despite the fact that sigma-delta systems are more linear, and remove the zero-crossing distortion inherent in ladder DACs. Virtually every ADC and DAC design available today uses a sigma-delta modulator (whether 1-bit or few-bit).

- DSD encodes music using pulse-density modulation, a sequence of single-bit values at a sampling rate of 2.8224MHz. This translates to 64 times the Red Book CD sampling rate of 44.1KHz, but at only 1/32,768th of its 16-bit resolution. You cannot directly compare DSD and PCM in this way. As Mal points out, noise-shaping is an inherent part of sigma-delta modulation, and as a result of the sampling frequency and modulator-order, PRACTICAL DSD systems can achieve slightly better than 120 dB SNR over the audio band. This equates to 20-21 bit performance in PCM terms. That is state-of-the-art as a result of practical circuit thermal noise limits. No 24-bit PCM converter is going to measure better!

In data terms, DSD64 uses slightly more data than 24/96 PCM, but a lot less than 24/192. Of course, due to the characteristics of the noise-shaper, DSD cannot achieve 20-21 bit performance over the entire frequency range, so cannot be considered 'rectangular' like a PCM system. However, the rising shaped noise is still very low in level, and is outside the audio band, so can be considered benign, and it is unclear whether anyone designing a 24/192 system would opt for a brickwall filter characteristic at around 90 kHz to be able to ensure a 120 dB SNR for 85 kHz tones! The proof of the pudding is in direct comparison listening tests, and I'll comment on that (the topic of this thread) in another post.

- You can’t quantize to a fraction of a bit, and you can’t quantize to a fraction of a sampling rate. I've already written reams on sampling theorem in various threads here on the Forum (not sure how many got ultimately deleted, though! ), so let me keep this very brief. Firstly, I have no idea what he means by quantizing to a fraction of a sampling rate. The sampling rate is the sampling rate. If you needed to sample faster, you should have! As Mal mentions, the sampling rate (for PCM) determines the highest frequency which can be unambiguously resolved. The key word is unambiguously. To avoid the problem of aliasing, the input to an ADC is filtered, and it is the filtered signal which is quantized. That is, the rate of change of the signal being quantized is limited so as not to conflict with the sampling theorem. Many people seem to forget this fact. From a purely mathematical stand-point, sampling faster will not add any meaningful data. However, there are limitations in practical anti-alias/reconstruction filters, and using a higher sampling frequency can have benefits with regard to the phase/amplitude response of the converter (less ringing, less phase distortion, etc.).

While it is true that you cannot quantize to a fraction of a bit, provided that you use sufficient bits, the quantization error can be made very small. Moreover, the reconstruction is not a simple join the dots process. If you had perfect anti-alias/reconstruction filters, the reconstruction process would equate to the weighted sum of (sin x/x) functions placed at the sample points. It is like fitting weighted-curves to the data. This not only gives us a continuous time signal again, but also gives essentially the correct amplitude values again. Remember, we have constrained the input signal via the anti-alias filter, so you cannot have some fast little squiggle between sample points!

- For example, when Sony decided to archive their analog master libraries to DSD64 back in the mid 1990s, they were wrong to believe that these masters would be future-proof and able to reproduce any consumer format. The fact is, these masters could only properly reproduce a format that was divisible by 44.1KHz. This assumes that the only operation you can do on the data is division. However, in their very first White Paper on DSD, Sony showed how you arrive at 48 kHz-based PCM formats:

[DSD64 x 5] / 294 = 48 kHz

[DSD64 x 5] / 147 = 96 kHz

If you want to get 192 kHz, you would multiply by 10, and so on.


I hope this gives you an idea of how much is wrong in that article. My advice is to completely ignore it.

 

Right, I'm talking about ± full-scale

 

We're just talking about the 103, right? I thought the 105/105D output DSD direct, be it via HDMI or the analog outputs?

 

Well, that's where the Nyquist-Shannon theorem comes in. Intuitively you might expect that a more precise sample would result in a smoother waveform closer to the original, but the theorem tells us that the waveform can be exactly reconstructed *subject* to the constraints of dynamic range (or noise floor) and frequency response that are well known. Of course that doesn't prove that DACs succeed in doing this but it is possible in theory.

Tim

 

The sampling theorem is an ideal mathematical proof in which you can fully reconstruct the input waveform from the encoded waveform. This is true if you have the exact sample amplitudes at the exact sample intervals and ideal filtering at input and output. In reality, we have imprecision in the recorded sample values (quantisation error), timing error (jitter) and non-ideal filtering, along with other noise.

What we are discussing here relates to how far we must minimise these unwanted effects before we can no longer hear any degradation from these errors. I think it is unwise to assume that 16-bit level resolution is fully beyond the ability of the human ear/brain's capabilities as far as micro-dynamics are concerned.

 

A good test to run would be using a DoP capable DAC with an Oppo 103 + Vanity HD card. Send the SACD 24/88.1 first, and then DSD. See if the difference is audible. Too bad you can't audition the Vanity HD card in advance.

 

It still 'downsamples' to 88.2 when converting DSD to LPCM via HDMI. It also sends straight DSD if that's your choice and you have a receiver capable of decoding.

 

My understanding is that the Vanity card allows for full resolution from the optical out (176.4 vs 88.2)?

 

Then even better. But for me the question is: is the Vanity103 HD card worth almost twice the cost of the original unit? If I can get 99% of the sound quality outputting 23/88.1 to the same DAC, then I would be happy. If the difference is not noticeable, then forget it.