R/E/P Community

R/E/P => R/E/P Archives => Dan Lavry => Topic started by: tonymite on October 10, 2004, 12:14:42 PM

Title: High sample rate Plug-Ins
Post by: tonymite on October 10, 2004, 12:14:42 PM
Hi Dan -
before i get to ball breakin,thanks for putting up with us - hope to learn a shitload.

I've read that certain Plug-Ins perform better at high sampling rates - contrary to your report on your website re: Nyquest.
are these people just full of non-sense ??? and just imagining ??

or do plug-ins benefit from hi sample rates ???

tenx
Title: Re: High sample rate Plug-Ins
Post by: bobkatz on October 10, 2004, 04:17:07 PM
tonymite wrote on Sun, 10 October 2004 12:14

Hi Dan -
before i get to ball breakin,thanks for putting up with us - hope to learn a shitload.

I've read that certain Plug-Ins perform better at high sampling rates - contrary to your report on your website re: Nyquest.
are these people just full of non-sense ??? and just imagining ??

or do plug-ins benefit from hi sample rates ???

tenx



Actually, Dan has supported the need for non-linear processing at higher sample rates to minimize problems with aliasing. There has been an AES Journal article on the subject. The problem is well known.

BK
Title: Re: High sample rate Plug-Ins
Post by: danlavry on October 10, 2004, 05:54:53 PM
bobkatz wrote on Sun, 10 October 2004 21:17

tonymite wrote on Sun, 10 October 2004 12:14

Hi Dan -
before i get to ball breakin,thanks for putting up with us - hope to learn a shitload.

I've read that certain Plug-Ins perform better at high sampling rates - contrary to your report on your website re: Nyquest.
are these people just full of non-sense ??? and just imagining ??

or do plug-ins benefit from hi sample rates ???

tenx



Actually, Dan has supported the need for non-linear processing at higher sample rates to minimize problems with aliasing. There has been an AES Journal article on the subject. The problem is well known.

BK


It is NOT contrary to my report on my web site! The opposite is true!

If you read my paper “Sampling Theory” you will see that I made an clear distinction between the sample rates used for storage and transmission of digital audio, and LOCLIZED sample rates, for a specific task.

I stated some examples where very fast sampling is in order. The first example was the rate used in a front end of a modulator of modern sigma delta AD’s which is somewhere between 64 fs and 512 fs !!! That is between 2.8MHz and 22MHz !!! We should not confuse it with that 192KHz data rate. The very fast rates for the AD’s front end is way beyond what is needed to make the design of the anti aliasing filter an easy task. It is about a converter architecture that begins with one bit (or a few bits) at very high rate to be “exchanged” for many bits at the lower rate, which is the DATA rate – audio sample rate.  

The second example I used was DA’s. Almost all DA’s use some amount of up sampling. The main reason here is to allow for a use of a simpler anti imaging analog filter. Making a 44.1KHz DA without any up sampling is a very difficult and costly affair. Some PCM DA’s up sample to as high as 16 fs (700KHz or so), and sigma delta types go higher. Again, this is a localized sample rate.

In both cases, one benefits greatly from a LOCALIZED higher sampling, and this is not the issue about 192KHz sampling DATA rate.

I did not talk about plugins in my paper. It was long enough, and I felt the 2 examples are sufficient. But as Bob Katz stated I supported higher rates for plug ins as a localized rate. That is of course for certain types of plug ins that intentionally introduce non linear behavior, such as in “tube sound”. Any intentional non linearity in fact.

I would go beyond the word “supported”. I do the math and figure out what is required to do the job right. We all agree that aliased energy is non musical. When doing a non linear process, the energy content requires more bandwidth. An X^2 function doubles the bandwidth, X^3 triples it and so on. So a non linear curve such as (for example) y=a0+a1X+a2X^2+a3X^3+a4X^4+a5X^5 takes 5 times the original bandwidth. A 22KHz digital audio would require 110.25KHz signal bandwidth, thus 220.5KHz (or more) sample rate. A 10th order non linearity would require 441KHz rate or more…

Of course, once the non linear process produced the results under that fast rate, you are now able to filter the high frequency energy and down sample to the proper DATA sample rate, be it 44.1KHz or 96KHz… So the LOCALIZED fast rate is NEEDED to avoid aliasing. There is no other way to accomplish the task in digital form.

Plug-Ins perform better at high sampling rates - contrary to your report on your website re: Nyquest.  It is not contrary to the report on my website re: Sampling Theory.    

I hope I stated it clearly. Many people told me that my paper is difficult to follow. I am sorry if such is the case.

BR
Dan Lavry

Title: Re: High sample rate Plug-Ins
Post by: JackJohnston on October 11, 2004, 10:24:34 PM
I was originally confused by Nyquest theory because it's natural to assume that if the frequency of a sample can be represented, meaning that the actual frequency of the sound is maintained, then the sampled sound must sound the same as the original. Which, of course, is not necessarily the case.

My questions would be:

How many samples are used to represent a 20K Hz frequency sound at 44.1 Khz sample rate?

Does that number of samples accurately describe all of the characteristics of that 20K Hz sound in its original continues form?

Does it seem likely that in a good listening environment, most people could hear the difference?

Thanks,

Jack Johnston
jackjohnston99@hotmail.com



Title: Re: High sample rate Plug-Ins
Post by: Nika Aldrich on October 12, 2004, 10:56:37 AM
JackJohnston wrote on Tue, 12 October 2004 03:24

I was originally confused by Nyquest theory because it's natural to assume that if the frequency of a sample can be represented, meaning that the actual frequency of the sound is maintained, then the sampled sound must sound the same as the original. Which, of course, is not necessarily the case.


It is necessarily the case.  The exception is if there is distortion or other forms of error in the sampling process.

Quote:

My questions would be:

How many samples are used to represent a 20K Hz frequency sound at 44.1 Khz sample rate?


The real answer to this question is very complicated.  Each individual cycle has little more than two samples in it, but little more than two samples does not accurately represent that waveform.  It actually takes hundreds of samples prior to and after the cycle to accurately convey what goes on throughout those two sample points.  The technology here uses "look ahead" and "look behind" filters to determine what should be happening between the sample points, and the steeper the transition band the further it has to look ahead and look behind to reconstruct the waveform properly.

To answer your question it takes hundreds of samples to represent that 20KHz waveform properly - only 2 samples throughout the cycle itself - the rest conveying the data in front of and behind the cycle.

Quote:

Does that number of samples accurately describe all of the characteristics of that 20K Hz sound in its original continues form?


ABSOLUTELY!  That is the basis of Nyquist's work - was proving that, despite common sense, this actually does work.  It's very non-intuitive.

Quote:

Does it seem likely that in a good listening environment, most people could hear the difference?


Nope, by definition (or mathematical law).

Nika.

[/quote]
Title: Re: High sample rate Plug-Ins
Post by: PookyNMR on October 12, 2004, 12:43:19 PM
danlavry wrote on Sun, 10 October 2004 15:54

If you read my paper “Sampling Theory” you will see that I made an clear distinction between the sample rates used for storage and transmission of digital audio, and LOCLIZED sample rates, for a specific task.
BR
Dan Lavry




Dan, is there a link where I could find your papers?

Thanks,

Nathan
Title: Re: High sample rate Plug-Ins
Post by: danlavry on October 12, 2004, 02:02:40 PM
PookyNMR wrote on Tue, 12 October 2004 17:43

danlavry wrote on Sun, 10 October 2004 15:54

If you read my paper “Sampling Theory” you will see that I made an clear distinction between the sample rates used for storage and transmission of digital audio, and LOCLIZED sample rates, for a specific task.
BR
Dan Lavry




Dan, is there a link where I could find your papers?

Thanks,

Nathan


Hi Nathan,

www.lavryengineering.com under the support section.
The last paper is "Sampling Theory". There are other papers there as well.

BR
Dan Lavry
Title: Re: High sample rate Plug-Ins
Post by: danlavry on October 12, 2004, 04:11:48 PM
JackJohnston wrote on Tue, 12 October 2004 03:24

I was originally confused by Nyquest theory because it's natural to assume that if the frequency of a sample can be represented, meaning that the actual frequency of the sound is maintained, then the sampled sound must sound the same as the original. Which, of course, is not necessarily the case.

My questions would be:

How many samples are used to represent a 20K Hz frequency sound at 44.1 Khz sample rate?

Does that number of samples accurately describe all of the characteristics of that 20K Hz sound in its original continues form?

Does it seem likely that in a good listening environment, most people could hear the difference?

Thanks,

Jack Johnston
jackjohnston99@hotmail.com





Hello JJ

You said:
I was originally confused by Nyquest theory because it's natural to assume that if the frequency of a sample can be represented, meaning that the actual frequency of the sound is maintained, then the sampled sound must sound the same as the original. Which, of course, is not necessarily the case.

Well, it is not the same, but is half way to being there. The difference between the sampled (digitized) waveform and the original one (the analog) is very well defined, and it has one nice characteristic to it: The difference is in fact all made of high frequency energy. In other words, the part of the energy that makes up the signal is under Nyquist (0-22.05KHz for 44.1KHz sampling) and the error signal (the difference between the analog and it’s sampled representation) resides above Nyquist (22.05KHz to 44.1KHz).

So that sampled wave (just binary numbers) is converted to “voltage steps” at say 44.1KHz. We are now in hardware world. We run the voltage steps through an analog filter which removes the high frequencies which is the difference between the original analog and the sampled wave. Did we say “remove the difference”? I guess we did. So with no difference we end up with the original.

The confusion is often due to the difference between two statements:
1.   “The sampled wave contains all the information”
2.   “The sampled wave is all the information”.

1. Is the correct one.  
2. Is missing something (a filter)

How many samples are used to represent a 20K Hz frequency sound at 44.1 Khz sample rate?

Infinity. I see what you are getting at. Well, if we agree on some performance goal, we can figure it out. I did not analize it to be able to answer you with numbers (so and so many samples for such and such maximum deviation). It is a day’s work, and I am busy (maybe after the AES).
But I was wondering myself about the “startup” of a DA. Well if you have an anti imaging filter (analog low pass filter, after the DA, passing audio and blocking energy above Nyqusit), and you apply 1 sec of digital black, you will be doing fine. But that takes 44100 samples. If you are after, say .001% maximum deviation, I bet it takes fewer samples…

Your question might interest a lot of people, from data compression, to ear research, wavelets and more. It is not an issue as far as making converters. From a practical standpoint, for converters, the data is an on going steady flow – “it just goes for ever”. Technically speaking, 1 second is a long time, It is pretty close to “forever”.  
Does that number of samples accurately describe all of the characteristics of that 20K Hz sound in its original continues form?

Well, the number of samples is half the story. You also need the value of each sample (think of it as an XY plot). And with that information one can plot a “stair case” like plot” - a waveform which is a filtered  away from being a duplicate of the original.

Does it seem likely that in a good listening environment, most people could hear the difference?

I do not know all the answers. I know very little about the ear brain behavior. But once I get some answers from you we can proceed. For example:

120dB dynamic range is very good for the ear brain combination
0-20KHz audio frequency range (or 0-30KHz or to 40KHz…).

With such information, my job is well defined. I can now concentrate on “waveform in” is the same as “waveform out”.  Personally, my job is mostly around “electrical wave in” and “electrical wave out”. Microphone people are about “air motion in” "electrical wave out”. Speaker people do the opposite. Are there any obstacles between “picking the musical air vibrations at the performance space” and “regenerating the same into you ears”? Of course. There is the whole list (starting with room acoustics…)

If  I can restrict the answer to your question to from “vibration of a mic membrane” to “vibration of speaker cone”, I would say that the idea is to make the difference very small. We hardware people make the difference small until the ear research people, tell us it is small enough.

In theory, if we respect Nyquist rule, and go for enough bits, there is no difference within the specified boundaries of the agreed bandwidth and dynamic range.

BR
Dan Lavry


Title: Re: High sample rate Plug-Ins
Post by: JackJohnston on October 13, 2004, 12:38:39 AM


Thank you very much for your insightful answers. I didn't mean to ask such leading questions. I know the complexity of these questions require some thought and I really appreciate the depth of you answers.

It's interesting that you mention Wavelets. It seems that with Wavelets, plug-in manufacturers should be able to emulate the non-linear characteristics of analog gear very well. Do you think that this is happening? The popular opinion seems to be that plug-ins do not accurately emulate the non-linear charecteristics of analog gear. It also seems like much of modern commercial music is sterile sounding. Do you think this may be caused by the use of plug-ins that only emulate gear in a linear manner and/or may not be adaptive to the waveform like analog gear is adaptive to the signal?

Thanks very much.

Jack Johnston
jackjohnston99@hotmail.com


Title: Re: High sample rate Plug-Ins
Post by: Nika Aldrich on October 13, 2004, 10:15:42 AM
JackJohnston wrote on Wed, 13 October 2004 05:38

It seems that with Wavelets, plug-in manufacturers should be able to emulate the non-linear characteristics of analog gear very well.


They can do it without wavelets as well.  There is nothing in the analog world that can't be represented numerically, and anything that can be represented numerically can be done so with only traditional FFT analysis - given, as Dan said, performance goals.  

Quote:

Do you think that this is happening?


No.

Quote:

The popular opinion seems to be that plug-ins do not accurately emulate the non-linear charecteristics of analog gear.


Plugin manufacturers are still limited by performance goals that prevent them from accurately reproducing analog gear in some situations.  In other situations they try rather to design stuff that sounds as good as analog gear but without the limitations of analog - they lower the noise floor, reduce inharmonic distortion, provide more flexibility, et al.  In these situations they aren't trying to actually emulate analog gear - they are trying to improve upon it.

Quote:

It also seems like much of modern commercial music is sterile sounding.


First, this is a very broad, sweeping statement without very much specific corroboration or detail.  It will be difficult to answer to this accusation of "much of modern commercial music" without a basis for comparison and some reasonable information that can help substantiate this.  I agree that this notion is thrown about a lot, but that, unto itself, should not necessarily be accepted as an adequate basis for expounding on it.

Quote:

Do you think this may be caused by the use of plug-ins that only emulate gear in a linear manner and/or may not be adaptive to the waveform like analog gear is adaptive to the signal?


And even if we did accept the above statement, I do not believe it is for this reason.

Nika.
Title: Re: High sample rate Plug-Ins
Post by: JackJohnston on October 14, 2004, 10:57:35 PM

Thank you very much for all of your perspectives. I love to hear proponents of DSP as I intend to eventually release DSP products. But I always enjoy playing devils advocate, if for nothing more than finding out what industry experts believe.

Thanks again,

Jack Johnston
jackjohnston99@hotmail.com