IPB

Welcome Guest ( Log In | Register )

5 Pages V  « < 3 4 5  
Reply to this topicStart new topic
xiphmont’s ‘There is no point to distributing music in 24 bit/192 kHz’, Article: “24/192 Music Downloads are Very Silly Indeed”
Patrunjica
post Oct 14 2012, 01:55
Post #101





Group: Members
Posts: 28
Joined: 25-April 10
Member No.: 80142



QUOTE (splice @ Oct 14 2012, 03:39) *
So if you take your "jagged" join-the-dots line and join the dots with a sine wave curve, you'll find that there's exactly one curve that will join all the dots: one with the same frequency as the original frequency. Drawing that curve is exactly what the D to A reconstruction filter does.


No disagreements there, I was just wondering how it can perfectly reproduce the wave when the sample points have to be approximated, the ideal position for them lies at the end of an infinite series of decimals so naturally those get truncated at one point or another and what I'm wondering is if this truncation implies any errors in the reproduction concerning amplitude or the resulting frequency itself.

The deviations can't be that large since obviously PCM works but I'm wondering how large can they be for those frequencies that creep up closer to the given sampling rate?
Go to the top of the page
+Quote Post
greynol
post Oct 14 2012, 01:59
Post #102





Group: Super Moderator
Posts: 10338
Joined: 1-April 04
From: San Francisco
Member No.: 13167



If you were to look at the waveform in an editor that actually assumes proper reconstruction you will in fact see a proper sine wave regardless of the frequency. This is all about the specific application you used to display the waveform.

Your application is displaying it incorrectly!

This post has been edited by greynol: Oct 14 2012, 02:01


--------------------
Your eyes cannot hear.
Go to the top of the page
+Quote Post
Wombat
post Oct 14 2012, 02:06
Post #103





Group: Members
Posts: 1118
Joined: 7-October 01
Member No.: 235



QUOTE (Patrunjica @ Oct 14 2012, 02:55) *
The deviations can't be that large since obviously PCM works but I'm wondering how large can they be for those frequencies that creep up closer to the given sampling rate?

And another link:

http://www.hydrogenaudio.org/forums/index....

Now read the post especialy from Woodinville. I don´t understand all this but since you asked... Btw. when you use search in here you´ll find several similar threads dealing with such things.

This post has been edited by Wombat: Oct 14 2012, 02:07
Go to the top of the page
+Quote Post
mjb2006
post Oct 14 2012, 02:15
Post #104





Group: Members
Posts: 860
Joined: 12-May 06
From: Colorado, USA
Member No.: 30694



Go to the top of the page
+Quote Post
greynol
post Oct 14 2012, 02:29
Post #105





Group: Super Moderator
Posts: 10338
Joined: 1-April 04
From: San Francisco
Member No.: 13167



Thanks for that!


--------------------
Your eyes cannot hear.
Go to the top of the page
+Quote Post
splice
post Oct 14 2012, 02:34
Post #106





Group: Members
Posts: 143
Joined: 23-July 03
Member No.: 7935



QUOTE (Patrunjica @ Oct 13 2012, 17:55) *
No disagreements there, I was just wondering how it can perfectly reproduce the wave when the sample points have to be approximated, the ideal position for them lies at the end of an infinite series of decimals so naturally those get truncated at one point or another and what I'm wondering is if this truncation implies any errors in the reproduction concerning amplitude or the resulting frequency itself.


The displayed sample points only appear to be chaotic. They are actually accurately positioned to the limits of the numeric resolution. There is one sine curve that exactly (to the limits of the resolution) passes through all of the sampling points. Any inaccuracy introduced by "truncation of an infinite series" appears as noise in the output from the D to A conversion. This noise will be equal to or less than that represented by the least significant bit of the digital signal.

When you digitise a signal to 16 bits, each sample has to be represented by one of 65535 different numbers. If it falls part way between two possible values, it is set to the nearest. This means the digitised value can be up to half a "bit's worth" different than the original input. When you reconvert it back to analog, the output can be up to half a bit plus or miinus different than the original. This is noise, because it is different than the original signal.

For a 16 bit signal, this noise is insignificant. Try it for yourself. Use Reaper or Audacity to generate a 1 KHz signal using all 16 bits (0dBfs). Do it again using only the least significant bit (-96 dBfs). Play the 16 bit signal at the loudest level you can stand. Without changing the volume setting, play the 1 bit signal. Can you hear it? (I doubt it).

Once you understand this, we'll move on to dither.

Edit: mjb2006's picture would have saved me a thousand words if I'd seen it in time...

This post has been edited by splice: Oct 14 2012, 02:35


--------------------
Regards,
Don Hills
Go to the top of the page
+Quote Post
saratoga
post Oct 14 2012, 03:59
Post #107





Group: Members
Posts: 5157
Joined: 2-September 02
Member No.: 3264



QUOTE (Patrunjica @ Oct 13 2012, 20:13) *
True, but each time they repeat they are different then before,


I didn't realize you had picked such an unround number as 13001, so the periods you showed don't actually repeat exactly. But the function is still periodic, just with a much longer period. They have to repeat every LCM(44100,13001) samples given that the underlying sin function is itself periodic.

QUOTE (Patrunjica @ Oct 13 2012, 20:13) *
dividing up the sampling rate to that particular frequency results in a number with an infinite number of decimal places, approximating the location of one of the sampling points will influence the position of the next sample point and so on


This isn't true because subsequent values of sin(x) do not depend on previous values. How could they? sin(x) has an infinite domain, so if it could only be computed recursively it would be impossible to compute it at all.

This post has been edited by saratoga: Oct 14 2012, 04:01
Go to the top of the page
+Quote Post
Kees de Visser
post Oct 14 2012, 11:50
Post #108





Group: Members
Posts: 737
Joined: 22-May 05
From: France
Member No.: 22220



QUOTE (greynol @ Oct 14 2012, 02:59) *
Your application is displaying it incorrectly!
No need to shout smile.gif
Why not propose him an application that does it right (or at least better) ? My favorite (and cross-platform) is iZotope RX, which has a free demo version available here.
You can vary the "Waveform interpolation order" in the Preferences/Display between 0 and 64.
Nothing beats empirical evidence IMO.
Go to the top of the page
+Quote Post
greynol
post Oct 14 2012, 12:13
Post #109





Group: Super Moderator
Posts: 10338
Joined: 1-April 04
From: San Francisco
Member No.: 13167



Probably because I am WindowsAdobe Audition-centric and don't know of other programs. I'm shouting because I provided an immediate (and correct!) explanation to the dilemma which was ignored and now the discussion has needlessly gone off-topic. I'm tempted to split it off now.


--------------------
Your eyes cannot hear.
Go to the top of the page
+Quote Post
Porcus
post Oct 14 2012, 23:41
Post #110





Group: Members
Posts: 1995
Joined: 30-November 06
Member No.: 38207



QUOTE (Patrunjica @ Oct 14 2012, 02:55) *
No disagreements there, I was just wondering how it can perfectly reproduce the wave when the sample points have to be approximated


By 'knowing', in the appropriate sense, that it is a sine wave. In principle, not very far from the way that I can perfectly reproduce a line using merely two distinct points: by knowing that it is a line.

Not that it matters. A 20 kHz sine wave and a 20 kHz triangle-shaped wave (or whatever shape!) sound the same, as the difference is above the hearing limit. Only the sine wave part of the triangle will contribute to the ≤20 kHz range.


--------------------
One day in the Year of the Fox came a time remembered well
Go to the top of the page
+Quote Post
yourlord
post Oct 15 2012, 01:48
Post #111





Group: Members
Posts: 252
Joined: 1-March 11
Member No.: 88621



QUOTE (Patrunjica @ Oct 14 2012, 02:55) *
No disagreements there, I was just wondering how it can perfectly reproduce the wave when the sample points have to be approximated


Look up the Nyquist–Shannon sampling theorem. Because it's known that the highest possible represented frequency is X, there is only 1 possible frequency below X that can match any 2 sample values. Based on that fact the DAC will "connect" the 2 sample points with a voltage curve that matches that single possilbe frequency below X...

The fact that the samples can only be quantized to 65536 levels introduces what's known as quantization noise. At 16 bit resolution you can basically ignore it for audio delivery.

QUOTE (Porcus @ Oct 14 2012, 18:41) *
Not that it matters. A 20 kHz sine wave and a 20 kHz triangle-shaped wave (or whatever shape!) sound the same, as the difference is above the hearing limit. Only the sine wave part of the triangle will contribute to the ≤20 kHz range.


It should be stated that a 20kHz triangle wave can't be represented with 44.1kHz sample rate. It would be low pass filtered to a sine wave before being sampled.
Go to the top of the page
+Quote Post
Porcus
post Oct 15 2012, 07:37
Post #112





Group: Members
Posts: 1995
Joined: 30-November 06
Member No.: 38207



QUOTE (yourlord @ Oct 15 2012, 02:48) *
there is only 1 possible frequency below X that can match any 2 sample values. Based on that fact the DAC will "connect" the 2 sample points with a voltage curve that matches that single possilbe frequency below X...


That is wrong. The number of samples required, approaches infinity as frequence approaches sampling frequency / 2. For any frequency below, you can do with a finite (but possibly large) number of samples.

Simplify down to 3 bits, i.e. values 0 through 7. Pick two successive samples, say, a 4 and a 3. Matched by a 7654321012345676543210123456 sequence, a 6564321234565432345 sequence, a 543234543234543234 sequence and a 4343434343434343434343. What's the lowest frequency in each of those signals?


--------------------
One day in the Year of the Fox came a time remembered well
Go to the top of the page
+Quote Post
Arnold B. Kruege...
post Oct 15 2012, 13:32
Post #113





Group: Members
Posts: 4475
Joined: 29-October 08
From: USA, 48236
Member No.: 61311



QUOTE (Patrunjica @ Oct 13 2012, 20:13) *
I know the waveform can be rebuild since obviously I can hear it, but I can't help but doubt that it's a more of an imperfect reconstruction then compared to pretty much anything lower then a 11025Hz sine.


I'll put your mind at rest. A digital reconstruction of any digitized or synthesized analog signal is always imperfect for theoretical and/or practical reasons.

I think you're asking the wrong question. The most important question is not whether the reconstruction is imperfect, but rather how those imperfections compare to what you get when you do the same thing by other means, for example by keeping the signal in the analog domain.

When it comes to storing music, digital from day one of consumer digital audio back in 1982 has always been vastly more perfect than analog media.

When it comes to playing music, the digital outputs of music players are always vastly more perfect than their analog outputs, at any price from high to low.

When it comes to passing music through even a single audio component such as an AVR the same thing is true. Digital (DSP-based) AVRs have always been vastly more perfect than analog AVRs at any price from high to low.

When it comes to loudspeakers, if we had implementations of speakers that were as sophisticated as our AVRs, the speakers that were more fully implemented in the digital domain would have a greater potential to be accurate. That day is coming!

Go to the top of the page
+Quote Post
krabapple
post Oct 15 2012, 21:17
Post #114





Group: Members
Posts: 2518
Joined: 18-December 03
Member No.: 10538



QUOTE (Arnold B. Krueger @ Oct 15 2012, 08:32) *
QUOTE (Patrunjica @ Oct 13 2012, 20:13) *
I know the waveform can be rebuild since obviously I can hear it, but I can't help but doubt that it's a more of an imperfect reconstruction then compared to pretty much anything lower then a 11025Hz sine.


I'll put your mind at rest. A digital reconstruction of any digitized or synthesized analog signal is always imperfect for theoretical and/or practical reasons.


I'd note here that an analog recording of an acoustic event -- say, a sound recorded on magnetic tape -- is always imperfect. And an analog reformatting of a recorded signal (e.g., an LP pressing sourced from a tape, or from digital ) -- is also always imperfect. Look at how the imperfections pile up! wink.gif

Go to the top of the page
+Quote Post
xiphmont
post Jul 18 2013, 18:40
Post #115


Xiph.org


Group: Developer
Posts: 176
Joined: 24-September 01
Member No.: 16



QUOTE (krabapple @ Oct 11 2012, 14:53) *
Just want to point out that there's a sustantial diff in terms of peer-review between AES convention presenations/publications, and JAES publications. Oohashi et al. never made it past convention, as far as I can tell. Their work ended up in a low-impact neurophysiology journal.


Ah, I suppose that would be an important distinction, and one I wasn't aware of. I've been an AES member on and off over the past 15 years, but mostly for the journal access. I find it hard enough to make the local BAS meetings :-)

My point was that the publication (presentation) of the paper did not automatically elevate it beyond a status of 'let's discuss', not that the AES should be faulted for presenting it.
Go to the top of the page
+Quote Post
xiphmont
post Jul 18 2013, 18:50
Post #116


Xiph.org


Group: Developer
Posts: 176
Joined: 24-September 01
Member No.: 16



QUOTE (bandpass @ Oct 11 2012, 23:51) *
QUOTE
Papers with interesting ideas and no data (eg, the J. Dunn 'equiripple filters cause preecho' paper, which presents a fascinating insight, even if it doesn't work out in practice)

Dunn refers to R. Lagadec and T. G. Stockham, ‘Dispersive Models for A-to-D and D-to-A Conversion Systems’ for data. Is there data elsewhere to support that it doesn't work out in practice?


Dunn says something a little different than Lagadec, although Lagadec was certainly hinting hard at the direction Dunn went.

In any case, Dunn presents simulations of an approximation that's a partial fit to the frequency responses of several DAC reconstruction filters in the field at the time. The partial-ness of the fit is not accounted for in the simulated figures, nor does Dunn show any measurements of the actual filters to show if and how well they match the approximation. Thus we don't know how good or bad the approximation is in practice.

But my primary thought is that it's a filter strategy and family that aren't used in DACs anymore. This work is from the bad old days of high-order analog equiripple IIR anti-imaging filters, which at the time were dominant. This is not a fault of the paper in any way, but we do not use these filters today and so it would seem the concerns simply do not apply anymore.

Of course, we also have test equipment today that should make verifying that assumption pretty easy. I've been wondering when I'd have a lazy weekend afternoon with nothing better to do than go explore the Dunn paper experimentally.
Go to the top of the page
+Quote Post

5 Pages V  « < 3 4 5
Reply to this topicStart new topic
1 User(s) are reading this topic (1 Guests and 0 Anonymous Users)
0 Members:

 



RSS Lo-Fi Version Time is now: 22nd December 2014 - 22:02