World’s First Valid Comparison of PCM versus DSD?

[Maldur]

How does DSD sound if done in PCM?

...Sounds like PCM, nothing more and nothing less.

[Daudio]

Can't DSD can be converted to analog for editing rather than PCM?

 

I can understand why DSD can't have EQ, compression, and other tricks applied to it, but there should not be any problem doing simple editing: splitting, adding and/or removing snippets, and such. All that needs to be known for that are word boundaries, not frequency interpolation.

[Miska]

I can understand why DSD can't have EQ, compression, and other tricks applied to it, but there should not be any problem doing simple editing: splitting, adding and/or removing snippets, and such. All that needs to be known for that are word boundaries, not frequency interpolation.

 

Word boundary on DSD is one bit. But just as with PCM you cannot just take two random pieces and put these together, because it would create discontinuation point in the signal and you hear snap/pop. If ends of the two waveforms don't meet, you need cross-fade or similar to avoid discontinuous signal.

 

With DSD one sample doesn't encode any absolute state of signal unlike PCM, but instead the signal is continuous function of bit stream. There is no clear number on how many DSD samples are needed to encode a full state because it is continuous function.

 

It is still entirely possible to process EQ etc for DSD just as for PCM too. The process is just more complex and thus computationally more expensive. But now modern processors have enough capacity to perform such processing.

[Boris75]

Not the original poster, and don't know soccer well enough.

 

You probably know in baseball a pitcher stands at 18 meters from a batter with a stick. He will use the stick to hit a small palm sized ball. The pitcher throws the ball, and the batter tries to hit it with his wooden bat or stick.

 

Nolan Ryan was known for how fast his fastball was. He simply threw it so fast it was difficult for a human batter to react and hit. Some pitches were around 170 kph. He had a better record in some respects than any pitcher in history.

 

Greg Maddux was almost as successful. He used techniques that made the ball curve and move. Also making it hard to hit with a stick, but more due to movement than speed. However his fastest fast ball was only 140 kph. But combined with movement of even slower pitches he still was highly effective at preventing hits with a stick or bat.

 

So one guy is effective due to speed. Another guy is effective due to movement. So you set up a test to pick the best pitcher. You only test a fastball which gives an edge to Ryan. And you evaluate the result by how the catcher does catching the ball. None of which really are anywhere near the things that matter about which pitcher will be effective. In fact both were among the most effective in history. You then declare a winner with great fanfare. But you really haven't done anything other than make a spectacle of yourself.

 

Many thanks for the very clear explanation.

[elcorso]

It's more like DSD is a one trick pony. It's said to be sublime, but only under very limited circumstances. (In rooms so large EQ won't help or subs, with a single resistor DAC, when the recording was made with a single bit DSD recording, and on and on) When Nolan's fastball was on, he was literally unhitable. But that wasn't very consistent. OTOH, Maddux never needed to rely on one pitch to win. Maddux could even hit and field. How many gold gloves did he win? Maddux could work the plate ump's strike zone like a master as well. Nolan Ryan is best known for his rare no-hitters and highlight reel. Sorry it's the truth. PCM does it all: EQ, delay, cross-over, volume control, wide DR and super extended FR.

 

Hi,

 

Single resistor DAC...? But what a wonder resistor, to the MGHZ range...!

 

We could do EQ, delay, cross-over, volume control, et al under DSD too. But what about those like me that doesn't like 'adjustment' like EQ to the music, because feel it destroys the music?

 

Regarding wide DR, dynamic contrast and super extended FR, maybe you are talking about DSD...

 

And maybe some people needs all those adjustment when listening to PCM, trying to hide the embedded noise inherent to PCM recordings, no so under DSD. The proof is very easy: Try to record an empty (no music) PCM track and then do the same with a DSD one, you can listen to the noise in the PCM one, but not in the DSD track.

 

I'm sorry but I don't know nothing about Nolan & Maddux, but yes about Messi & Ronaldo. I don't to want to cite'm in order not to complicate the thread.

 

At the end everything (again) could be a "matter" of taste".

 

Best,

 

Roch

[Daudio]

Word boundary on DSD is one bit... With DSD one sample doesn't encode any absolute state of signal unlike PCM, but instead the signal is continuous function of bit stream. There is no clear number on how many DSD samples are needed to encode a full state because it is continuous function.

 

Miska,

 

Poor choice of words on my part. But I thought there was zero crossing, or other kind of time sync somewhere in a DSD stream.

Are you saying that is just an high-low, up-or-down, 1-or-0, sequence for the entire selection/track file, no matter how many Mb or Gb it is ?

[Miska]

Poor choice of words on my part. But I thought there was zero crossing, or other kind of time sync somewhere in a DSD stream.

 

Signal of course has zero-crossing, but you have to take undefined N number of preceding samples into account to figure out where the signal is going. Then the cut point needs to not create a discontinuation in this window. So it's doable, but it's not simple as cut-and-paste.

 

Are you saying that is just an high-low, up-or-down, 1-or-0, sequence for the entire selection/track file, no matter how many Mb or Gb it is ?

 

Yes, it is just continuous stream of bits.

[Daudio]

Signal of course has zero-crossing, but you have to take undefined N number of preceding samples into account to figure out where the signal is going. Then the cut point needs to not create a discontinuation in this window. So it's doable, but it's not simple as cut-and-paste.

 

Ah, doable ! Is anyone 'doing' it, that you know of ?

[Le Concombre Masqué]

Hi,

 

 

 

We could do EQ, delay, cross-over, volume control, et al under DSD too. But what about those like me that doesn't like 'adjustment' like EQ to the music, because feel it destroys the music?

 

 

ok, i found out that using A+ to convert dsd to pcm to be able to EQ was killing the magic and with a clear negative balance for pure dsd acoustic music .

 

But I don't dig how you can use the plain bad src version embedded in A+ (that takes flac and what have you, converts up to 384 when RX3 advanced does not and does not), which sounds nasty compared to RX3, and applies iZotope designed algorithms affecting multiple weird parameters, and accuse( iZotope designed as of my system) EQing algorithms affecting frequencies' levels of destroying the music !

 

I once showed my room's response ; is yours perfect ?

[dallasjustice]

Roch,

It's time you bought a calibrated mic and learned the truth about your room. :-)

 

Hi,

 

Single resistor DAC...? But what a wonder resistor, to the MGHZ range...!

 

We could do EQ, delay, cross-over, volume control, et al under DSD too. But what about those like me that doesn't like 'adjustment' like EQ to the music, because feel it destroys the music?

 

Regarding wide DR, dynamic contrast and super extended FR, maybe you are talking about DSD...

 

And maybe some people needs all those adjustment when listening to PCM, trying to hide the embedded noise inherent to PCM recordings, no so under DSD. The proof is very easy: Try to record an empty (no music) PCM track and then do the same with a DSD one, you can listen to the noise in the PCM one, but not in the DSD track.

 

I'm sorry but I don't know nothing about Nolan & Maddux, but yes about Messi & Ronaldo. I don't to want to cite'm in order not to complicate the thread.

 

At the end everything (again) could be a "matter" of taste".

 

Best,

 

Roch

[Miska]

Ah, doable ! Is anyone 'doing' it, that you know of ?

 

Someone more studio oriented may tell more, but as far as I know, Pyramix is able to keep the bitstream untouched elsewhere than just the splice crossfades.

[Teresa]

Are you saying that is just an high-low, up-or-down, 1-or-0, sequence for the entire selection/track file, no matter how many Mb or Gb it is ?

 

As Miska said it's a continuous stream of bits. However I will try to explain why many editing tasks can't be performed in 1 bit DSD, this is oversimplifying and may have parts that are incorrect as I barely understand the very basics of DSD.

 

So here goes:

 

Computer language uses the binary numeral system and has two states: 1s and 0s. However, DSD has three states, and here is how it accomplishes that feat: DSD compares the previous slice of the musical waveform in the pattern buffer to the current one to determine if it goes up, down, or stays the same. "1" corresponds to a pulse of positive polarity, "0" corresponds to a pulse of negative polarity. A run consisting of all 1s would correspond to the maximum positive amplitude value, all 0s would correspond to the minimum negative amplitude value, and alternating 1s and 0s would correspond to a zero amplitude value. Thus, there must be two 1s in a row for the waveform to go up, or two 0s in a row for it to go down. If the DAC sees 10 or 01 the output stays the same. The continuous amplitude waveform is recovered by low-pass filtering the bipolar PDM bitstream. It does this 2,822,400 times a second.

 

On the other hand PCM uses a word of defined length to describe a single slice of the waveform for example with 24 bit 96kHz PCM a 24 bit word every 1/96th of a second. Thus in PCM a sample can be moved and/or replaced with another sample, as each 24 bit sample completely describes that 1/96th of a second of the waveform.

 

In short, PCM is a digital translation of music for a given time frame and DSD follows the musical waveform up and down much like analog recording does.

 

DSD experts please feel free to correct what I have gotten wrong. Also I have a question about the first sample of a live DSD recording, there will be audience noise before the music starts, thus there is no previous sample in the pattern buffer to compare the first sample to at beginning of the recording, how does it know the level of the first note?

[PopPop]

Diogenes,

 

I didn't learn much from that other than you think Theresa is under educated. Could you describe better why editing tasks can't be performed in 1 bit DSD?

[Teresa]

All digital data, dsd, pcm, whatever is binary. Take one or two undergraduate computer science courses and get back to us.

 

Duh that is exactly what I said "Computer language uses the binary numeral system and has two states: 1s and 0s." Doesn't that mean that all digital data is binary?

 

What I was saying is that DSD requires a third state as it measures the movement of the musical waveform up and down as opposed to completely describing each sample as PCM does.

 

• "1" corresponds to a pulse of positive polarity
  
• "0" corresponds to a pulse of negative polarity
  
• alternating 1s and 0s would correspond to a zero amplitude value.

 

Thus that third state is alternating 1s and 0s indicating no movement of the waveform either up or down.

 

You must have misread my post as your response has nothing to do with what I stated. Also I don't want to take an undergraduate computer science course and you can't make me!!!

 

I was just trying to be helpful in explaining the difference between PCM and DSD and why DSD is hard to edit. You didn't reply to that at all!!!

[JohnSwenson]

Hi Teresa,

you almost have it. What you are describing is called delta modulation, its a cousin to sigma delta modulation. (it can be called sigma-delta or delta-sigma, both are used interchangeably)

 

In delta modulation the output pulses are the difference from the previous state to the current state. With this a string of ones would represent a rising voltage ramp. In sigma delta the stream of ones represents a constant highest value.

 

Sigma delta is little bit harder to understand, its the quantizing of the error signal computed from subtracting the quantization from the input. (don't worry about understanding what that means if you don't want to)

 

Some differences are: in delta modulation the single bit represents the smallest CHANGE in a signal, whereas in sigma delta the bit is the largest VALUE of the waveform.

 

Thus in delta modulation there is a limitation on how fast a signal can change, but with sigma delta it can go from full negative to full positive in one clock cycle. (stream of 0s changes to stream of 1s)

 

In delta modulation the decoder has to maintain the current value and add or subtract small values each cycle, THEN low pass filter that. With sigma delta the output just has to be low pass filtered.

 

Your description of the states are correct. Alternating 1s and 0s represents 0 in the waveform. A positive value will have a larger number of 1s than 0s over time. A negative value will be represented by more 0s than 1s. Thus to figure out what the current value is you have to look at a large number of pulses. This is the reason editing is difficult, it cannot just be done with data from one time, it has to be done with data over a long time frame.

 

The sigma delta modulator does contain state, but it is the error signal not the input signal value. At turn on this error state is assumed to be zero. It never causes a "pop" or other type of discontinuity because it is the error, not the actual value. BTW that question was a very astute question proving you really do have some insight into this stuff.

 

John S.

[magnum innominandum]

Teresa,

 

However, DSD has three states, and here is how it accomplishes that feat: DSD compares the previous slice of the musical waveform in the pattern buffer to the current one to determine if it goes up, down, or stays the same. "1" corresponds to a pulse of positive polarity, "0" corresponds to a pulse of negative polarity.

 

You mean it applies averaging (or lowpass filtering)?

 

So DSD is standard PDM? Then 2.8MHz seems a little low.

 

On the other hand PCM uses a word of defined length to describe a single slice of the waveform for example with 24 bit 96kHz PCM a 24 bit word every 1/96th of a second.

 

You seemed to have missed the K in the sample rate. In your case it does 1/96th Milliseconds or every appx. 10uS. I wonder what slew rate this equates to.

 

In short, PCM is a digital translation of music for a given time frame and DSD follows the musical waveform up and down much like analog recording does.

 

Isn't it more precise to state both PCM and analogue describe an absolute value of the input comparedto an absolute reference at any given point in time with an accuracy determined by a variety of details, while Delta Sigma describes...

 

Well, actually what does it describe? It does not describe an absolute value, unless the number of bits observed is approaching infinity...

 

So it describes a DIFFERENCE (DELTA) over TIME (SIGMA).

 

And the steps are 1/2.8 Millionth of the Peak-Peak analogue signal (theory) for 1/2.8 Millionth of a second. The step size is very small (though 24 Bit PCM in theory allows us to describe any of 16.7 Million absolute values for each and every sample) as is the time domain step size.

 

However there seems to be a very clearly defined and limited slewrate here (unlike PCM) if an old fogey amplifier designer looks at this. So, may I ask the DSD Guru's, what is the theoretical slewrate of DSD (the time taken for the signal to go from -2.8V to +2.8V to wit)?

 

Also I have a question about the first sample of a live DSD recording, there will be audience noise before the music starts, thus there is no previous sample in the pattern buffer to compare the first sample to at beginning of the recording, how does it know the level of the first note?

 

Well, there is no "first sample", only a first bit. If there was along enough silence before the first bit, the modulator would have approached midpoint, as it would if the system just came out of reset.

 

If the system was not centred when the first bit arrived, it would continue from whatever previous value and might eventually clip the system. Oh, is that the reason why DSD is notionally limited at 50% modulation? But what if the next bit sequence again starts at the maximum modulation of the prior sequence, which in turn was at the max? We clip again.

 

As a strictly analogue/RF guy I get a more and more worried feeling about this DSD thing. If it can work at all,they must be using Voodoo and Magic(or maybe just smoke and mirrors?).

 

Maybe someone has some 'scope traces of a DSD output so we can see how it compares to analogue (and not those that Miska declared where from a very poor DSD DAC)?

[Boris75]

Take one or two undergraduate computer science courses and get back to us.

 

Am I alone in finding this remark unpleasant to read?

[magnum innominandum]

John,

 

Thus in delta modulation there is a limitation on how fast a signal can change, but with sigma delta it can go from full negative to full positive in one clock cycle. (stream of 0s changes to stream of 1s)

 

With sigma delta the output just has to be low pass filtered.

 

These two statements seem at odds with each other. If a lowpass filter is involved there is no instant change of state within 1Bit, the filters settling time must be considered.

 

Further, if as you say, the bits reflect the Error then a continous string of 1's in effect represents the system slewing if a dynamic signal is involved (read for example a fullscale step). At least if the system involved was an Amplifier we would call it slewing. So one would have to either accept this slewing limitation and the resultant distortion (which we call TIM in amplifiers) or filter the input to the system same as for feedback amplifiers.

 

So I still wonder, what is the slew rate of DSD (that is fastest change of the input from fully negative to fully positive that does not cause the modulator to saturate)?

[Miska]

However there seems to be a very clearly defined and limited slewrate here (unlike PCM) if an old fogey amplifier designer looks at this. So, may I ask the DSD Guru's, what is the theoretical slewrate of DSD (the time taken for the signal to go from -2.8V to +2.8V to wit)?

 

PCM has much more clearly defined limited slew rate since it is heavily band-limited system by definition. Nyquist frequency clearly defines the maximum slew rate. PCM has same speed limitation regardless of signal level.

 

With DSD the amount of time required to slew a signal depends on signal level (just like analog), lower the swing, higher the available bandwidth is (and also better the linearity becomes). At 0 dB level DSD64 slew rate is equivalent of 117.6 kHz PCM. At -6 dB level it is equivalent of 235.2 kHz PCM and so on...

 

So pushing DSD to 0 dB doesn't gain much, while PCM needs to be pushed a lot because otherwise you are constantly leaving lot of value space unused. DSD always uses the full value space regardless of signal level.

 

If the system was not centred when the first bit arrived, it would continue from whatever previous value and might eventually clip the system. Oh, is that the reason why DSD is notionally limited at 50% modulation? But what if the next bit sequence again starts at the maximum modulation of the prior sequence, which in turn was at the max? We clip again.

 

DSD modulator design is one kind of art form in itself. There are some tricks to make it unconditionally stable. However what you described doesn't happen in proper implementation, because delta from the original signal would grow. The system tries to follow the source signal. So for sake of simplicity think it as similar to negative feedback in amplifier circuit. Without the feedback, high gain amplifier could quickly latch-up at one rail.

 

Maybe someone has some 'scope traces of a DSD output so we can see how it compares to analogue (and not those that Miska declared where from a very poor DSD DAC)?

 

I can try to make some, I didn't find any good ones from my stored measurements because I don't consider scope traces very useful.

[Jud]

Am I alone in finding this remark unpleasant to read?

 

It's OK, John Swenson, who actually knows something about these matters, provided helpful explanation a few comments later.

[Miska]

Now to the correct thread...

 

Here's 20 kHz sine at DSD128 with 60 kHz reconstruction filter:

 

Second harmonic is -93 dB and third harmonic at -92 dB level from analog stages.

[beetlemania]

Am I alone in finding this remark unpleasant to read?

 

You are not alone and I think there have been several posts in this thread that were uncivil. I've noticed that pretty much any DSD-related thread seems to bring out the crowd that is mostly interested in tearing someone down.

[mav52]

You are not alone and I think there have been several posts in this thread that were uncivil. I've noticed that pretty much any DSD-related thread seems to bring out the crowd that is mostly interested in tearing someone down.

Just about any thread where hearing, cables, jitter are involved moves towards the uncivil.

[alfe]

Just about any thread where hearing, cables, jitter are involved moves towards the uncivil.

 

It's called behind the keyboard effect ¨-)

[Teresa]

Teresa,

 

...You seemed to have missed the K in the sample rate. In your case it does 1/96th Milliseconds...

 

Thanks for catching my error, it should have read: “describes that 1/96,000th of a second of the waveform." Oops!

 

I don't know enough about the workings of DSD to answer your other questions, as I said I just repeated the simplified versions I've read and saved on my hard drive for later references such as this. The DSD experts hopefully can address your other concerns.