[LAD] Re: Direct Stream Digital / Pulse Density Modulation musing/questions

Phil Frost indigo at bitglue.com
Tue Sep 25 13:51:30 UTC 2007


On Tue, Sep 25, 2007 at 10:37:17AM +0200, Fons Adriaensen wrote:
> On Mon, Sep 24, 2007 at 09:41:39PM -0700, Maitland Vaughan-Turner wrote:
> 
> > > > Intuitively, one could also say that more sample points yield a
> > > > waveform that is closer to a continuous, analog waveform.  Thus it
> > > > sounds more analog.
> > >
> > > This is completely wrong. Sorry to be rude, but such a statement
> > > only shows your lack of understanding.
> > 
> > Why is it wrong?  If I drew some dots on a waveform and then connected
> > the dots, to try to reconstruct the waveform, wouldn't I get a better
> > result with more dots?
> 
> If you just connect the dots, yes. But that's not how an analog waveform
> is reconstructed from PCM samples. 'Connecting the dots' is not even part
> of that process.
> 
> > > > Thanks for the link.  My whole point of digging up this old thread
> > > > though, was to say that I've tried it, and my ears tell me that the
> > > > papers are incorrect.
> > >
> > > Then please point out the errors in the paper by Lipshitz and Vanderkooy.
> > 
> > my ears tell me that... that's all; it's just subjective.  haha, I see
> > subjective reports don't get you far around here.
> 
> You said the papers were incorrect. Then point out the errors.
> And indeed, a subjective evaluation is useless if not the result
> of a double blind test. 
> 
> If you have ever been involved in organising a controlled listening
> test you should know how easy it is to get completely invalid results
> and to fool yourself into believing things that are just an illusion.
> 
> > > I'm not saying that DSD is crap. It sounds well. But it doesn't meet
> > > the claims set for it (as shown by L&V - you need at least two bits
> > > to have a 'linear' channel) and  as a storage or transmission format
> > > it's inefficient compared to PCM. That means that if you use PCM with
> > > the same number of bits per second as used by DSD, you get a better
> > > result than what DSD delivers.
> > 
> > well, what do you mean by better?  It seems like 24 bit is already
> > better in terms of dynamic range at any sample rate, but if you mean
> > more detailed representation of a waveform (in time), it seems like
> > you necessarily need to have the highest possible sample rate.
> 
> There is a point where more detail becomes irrelevant because it's
> way below the noise. For a properly dithered PCM signal it can be
> shown that the error that remains is in a strict mathematical sense
> indistinguishable from noise. Hence if it's below the analog noise
> floor it doeasn't matter any more. BUt you need at least _two_ bits
> for this to work. 
> 
> For 'detail in time' the situation is even simpler. If the waveform
> is bandlimited, then *every* detail is captured by sampling it at a
> rate equal to twice the bandwidth. Each sample does not only represent
> the waveform at the time it was taken, but in fact contains information
> about the entire waveform. The samples completely describe the waveform
> in the same strict sense as three points are sufficient to define a
> circle. It's not an approximation.

Here is a good document on sampling theory, how it works perfectly in a
mathematical sense, and how differences from ideal (lack of infinitely
narrow impulses, ideal lowpass filters, quantization error) are made
insignificant in practical implementations:

http://www.lavryengineering.com/documents/Sampling_Theory.pdf



More information about the Linux-audio-dev mailing list