Jump to content

Sampling rates


Tony Johnson

Recommended Posts

If I may ask a question that may steer us back to the original topic a bit: don't most modern AD converters oversample internally already? Meaning that the oversampling would be handled in the digital domain to have a more gradual anti-aliasing filter (thus the concrete benefits of extended bandwidth above Nyquist) with the storage/processing efficiency of the standard 44.1/48k sampling rates?

 

From Wikipedia:

 

"By increasing the bandwidth of the sampled signal, design constraints for the anti-aliasing filter may be relaxed.[1] Once sampled, the signal can be digitally filtered and downsampled to the desired sampling frequency."

Link to comment
Share on other sites

  • Replies 106
  • Created
  • Last Reply

Top Posters In This Topic

I think everyone can agree on that

 

Allow me to summarize the points of agreeance here:

- In order to properly sample audio, the audio's frequency content must lay entirely below the sampling rate's Nyquist frequency, otherwise aliasing artifacts would be present. To account for that fact, the input signal either goes through analog filters or gets oversampled and digitally filtered before conversion to the intended sample rate.

- Intermodulation is an effect by which two or more frequencies can be mixed to generate other frequencies at sum and difference multiples of the original frequencies.

- Higher sampling rates are beneficial to certain post processing

- Higher sampling containing ultrasonic frequency content is necessary for post processing which will transfer ultrasonic frequency content to audible range (eg time stretching)

 

And the points of contention here:

- Does sampling at 96khz add quality to audio whose only frequency content is within 20hz - 20khz?
- Can humans hear or be otherwise affected by ultrasonic frequencies?

- Where can and cannot intermodulation occur: in open air, in audio equipment, in your ear, or in your mind?
- If intermodulation does occur in open air, is it necessary to record ultrasonic frequencies when their intermodulation has already had its effect on auditory frequencies?
- Is intermodulation in audio equipment considered distortion of the audio signal?
- Can upconverting 48khz data have equal benefit to post processing as data originally sampled at 96khz when ultrasonic frequency content is not required?

Let me know if I forgot anything, but I think I've made my opinions on this subject pretty clear, and so shall excuse myself. As a final note, I'd like to reiterate that this article spells it all out pretty clearly with illustrations and samples: http://people.xiph.org/~xiphmont/demo/neil-young.html

Cheers,

James

Link to comment
Share on other sites

Big picture.

It's very difficult to discuss these topics in text. I think a interactive real time video discussion is much more appropriate. If people are interested maybe we should make that happen.

Then we can easily have pictures and audio comparisons and so on...

 

I'm sorry to bring this up again, but I didn't get to watch this video before now, and it vividly tests and illustrates every question related to sampling rate (not the frequency mixing in air questions): http://xiph.org/video/vid2.shtml

Link to comment
Share on other sites

 

1- Does sampling at 96khz add quality to audio whose only frequency content is within 20hz - 20khz?

 

2- Where can and cannot intermodulation occur: in open air, in audio equipment, in your ear, or in your mind?

 

3- Is intermodulation in audio equipment considered distortion of the audio signal?

4- Can upconverting 48khz data have equal benefit to post processing as data originally sampled at 96khz when ultrasonic frequency content is not required?

 

1 .  No

2. like you said, no intermodulation in linear systems.

    there is a psicoacustical efect of intermodulation, sum and difference tones, discovered by the violinist and acustician Giuseppe         Tartini, We had a demonstration with a violinist at  the university. The sound can be listened. but it doest not exist in the acoustic         realm, only in the brain.

3. yes, It adds frecuencies not  present in the original sound, it is a type of non linear distortion.    

4. yes , it is the same. If ultrasonics are not required, 

 

I repeat that it seems that in general, people dont understand the conversion process. The comon view of a converter as a device that fill  voltage values thruogh time as seen visualy in the waveforms of a computer editor make people to think that  with more samples the signal is better represented, and better sounding. That is incorrect, digital audio samples are not exactly the same than pixels in pictures. Into an editor, a 1 khz frecuency waveform looks uglyer sampled at 3 khzsf but when reproduced thruogh a properly designed converter for that sample frecuency , it sounds the same as it would sound if sampled at 96 khzsf.

 

If there are tiny details between samples in a signal,  that are not captured at 48 khz sample frecuency, is because that details are made of frecuencies higher than 24 khz 

 

the reason for CDs to use 44.1 khzsf has nothing to do with aliasing, is because of in the beginig of digital audio, the conversion of the masters to digital was made into U-matic analog videotapes with a special code of analog patterns. 

 

from wikipedia:

" with the famous compact disc 44.1 kHz sampling rate based on a best-fit calculation for the U-matic's video horizontal-sync rate. "

http://en.wikipedia.org/wiki/U-matic

 

Also, I have noted a  confusion in the terms "conversion, sampling and cuantizing".  Sampling is an analog precess, the first part of the conversion process, Cuantizing is the measurement of the result of sampling. Only when the result of cuantizing is buffered, then the signal has been converted.

Link to comment
Share on other sites

NewEndian: "  Does sampling at 96khz add quality to audio whose only frequency content is within 20hz - 20khz? "

VR sez "no", but

I would say: Maybe: "   If the higher rate sampling is done well, it can increase the accuracy of the potential reproduction of the analog waveform as their is less interpretation of what is the actual waveform between the individual samples. "

so maybe it adds a bit of quality, but virtually undetectable by human hearing...

Link to comment
Share on other sites

I'm sorry, Senator. I have to point to this video again. He very clearly demonstrates that the input and output of the 44.1k A-D-A conversion process are identical: http://xiph.org/video/vid2.shtml

96k does not in any way better represent <20khz audio than 48k. It is mathematically proven, and he demonstrates it in the video. Watch the video.

Link to comment
Share on other sites

It's not my integral calculus (though I did take that class). This is the calculus that is actually related to sampling: http://en.wikipedia.org/wiki/Nyquist%E2%80%93Shannon_sampling_theorem#Shannon.27s_original_proof

Yes, one might intuitively say that the more points, the better. However, it should not come as a surprise that not all curves require the same number of points to 100% accurately describe the curve: One half of a square wave for instance only requires 4 points. Nyquist and Shannon have mathematically proven that the least number of points required to 100% accurately describe a waveform is just more than 2x the highest frequency.

Watch the video: http://xiph.org/video/vid2.shtml

Link to comment
Share on other sites

Ok, aside from the question of the benefit of higher sampling rates to the audible spectrum, I still think that more resolution is beneficial to various post processes like time stretches and noise reduction. And while this has not been answered yet, here, by an authoritative voice, it seems very plausible that upsampling is not good enough, or not as good anyway

Link to comment
Share on other sites

we agree to disagree...

including your integral calculus vs my integral calculus...

http://online.math.uh.edu/HoustonACT/videocalculus/index.html

#21:

javascript:var%20x=window.open('SV3/21-area.mov','popup','width=828,height=644,%20menubar=no,scrollbars=no,resizable=yes,location=no');

 

we disagree because conversion is not an example of integral calculus. Sampling is about modulation, and cuantizing is about discrete value measurement.  Conversion is to sampling and then cuantizing. Conversion is not directly related to area under the curve calculus nor primitivisation of mathematical functions.

  All the asumptions on better representing is just a missunderstandig on how converters do really work.

Link to comment
Share on other sites

Ok, aside from the question of the benefit of higher sampling rates to the audible spectrum, I still think that more resolution is beneficial to various post processes like time stretches and noise reduction. And while this has not been answered yet, here, by an authoritative voice, it seems very plausible that upsampling is not good enough, or not as good anyway

I agree, I don't have concrete evidence for that notion like the Nyquist-Shannon Theorem really is for audible frequency content.

This is just a notion: I would argue that if the ultrasonic frequency content isn't there (has been filtered out by standard audio equipment) then higher sampling rates are not more beneficial than 48k.

Link to comment
Share on other sites

Ok, aside from the question of the benefit of higher sampling rates to the audible spectrum, I still think that more resolution is beneficial to various post processes like time stretches and noise reduction. And while this has not been answered yet, here, by an authoritative voice, it seems very plausible that upsampling is not good enough, or not as good anyway

If you want to record  ultrasound to investigate , or if you want to hear a bat`s sonar by reproducing it at a low sample rate , it is necesary to use high sample rate recording. If you only cares about sub 20 khz content, in order to facilitate interpolation for time stretch, upsample gives the same result for the audible bandwidt.

Link to comment
Share on other sites

VR: " if you want to record  ultrasound to investigate , or if you want to hear a bat`s sonar by reproducing it at a low sample rate , it is necesary to use... "

special gear, capable of capturing such sounds properly... eh?

yes, ultrasound capable gear for ultrasound recording.

20hz-20khz sound do not need 96khz recording, 48khz is perfect

 

apart from extending the audio bandwidt, the other benefit of higher sample rates that I can think of is that with proper design, it is posible to reduce the latency.

Link to comment
Share on other sites

... I still think that more resolution is beneficial to various post processes like time stretches and noise reduction.  And while this has not been answered yet, here, by an authoritative voice, it seems very plausible that upsampling is not good enough, or not as good anyway 

 

What would you consider an authoritative voice? Someone who designs these processors? Someone who uses them on a regular basis in production? Someone who looks at the math involved?

 

For what it's worth - and I'm not claiming to have enough authority to satisfy you - time stretching can benefit from a higher sample rate in the original recording... but only if the higher rate is actually supported by higher bandwidth analog and converters, and usable HF audio in the original source.

 

Noise reduction, on the other hand, doesn't seem to gain anything by running at higher than normal sample rates. It just means more work for multiband expanders, which will remain closed anyway... and unless you're actually trying to remove noise above the normal 20 kHz, would be ignored by "retouching" wavelet processors. (One would wonder why you're trying to remove noise that won't make it through the release process anyway, but that's a philosophical question.)

Link to comment
Share on other sites

If you accept the premise that 48k 100% accurately represents <20khz audio, it stands to reason that upsampling to 96k would also 100% accurately represent that audio (it wouldn't reduce the quality), and therefore would be of equal quality to the same audio (<20khz) sampled originally at 96k. Again this presumes that all audio is bandlimited at 20khz which is fair for the majority of audio recording equipment.

I submit the logical conclusion that for <20khz audio, original sampling at 96k provides no benefit over sampling at 48k and converting to 96k (or any other higher sample rate that post processing might require)

Link to comment
Share on other sites

What would you consider an authoritative voice? Someone who designs these processors? Someone who uses them on a regular basis in production? Someone who looks at the math involved?

For what it's worth - and I'm not claiming to have enough authority to satisfy you - time stretching can benefit from a higher sample rate in the original recording... but only if the higher rate is actually supported by higher bandwidth analog and converters, and usable HF audio in the original source.

Noise reduction, on the other hand, doesn't seem to gain anything by running at higher than normal sample rates. It just means more work for multiband expanders, which will remain closed anyway... and unless you're actually trying to remove noise above the normal 20 kHz, would be ignored by "retouching" wavelet processors. (One would wonder why you're trying to remove noise that won't make it through the release process anyway, but that's a philosophical question.)

Sorry, Jay, I didn't intend to imply that you or others here did not have an authoritative voice. However, you and James, as well as the Senator to some degree, seem to disagree on some of these issues and I suppose I was kind of hoping for a third voice of expertise to break the tie. And yes, I would love to hear from a manufacturer of one of these devices and hear their considerations.

One answer is given by StageTech, I suppose, who arguably build the cleanest and most neutral a-d converter (TrueMatch) which is popular with classical music recordists. Its highest sampling rate is 48kHz.

Link to comment
Share on other sites

If you accept the premise that 48k 100% accurately represents <20khz audio, it stands to reason that upsampling to 96k would also 100% accurately represent that audio (it wouldn't reduce the quality), and therefore would be of equal quality to the same audio (<20khz) sampled originally at 96k. Again this presumes that all audio is bandlimited at 20khz which is fair for the majority of audio recording equipment.

I submit the logical conclusion that for <20khz audio, original sampling at 96k provides no benefit over sampling at 48k and converting to 96k (or any other higher sample rate that post processing might require)

yes, you're logic seems flawless. Within the parameters you mentioned.

My idea was that 96kHz provides a finer resolution of the recorded audio and if a time-stretch is applied there is simply more information for the computer to stretch, rather than "filling in the blanks". I am completely wrong in this assumption?

Link to comment
Share on other sites

Just wanted to mention I appreciate everyone's level-headed responses.

As for "filling in the blanks": because the waveform is so accurately represented by 48k, the interpolation is not doing any guess work. It is merely restating the parts of the waveform that are otherwise still accurately represented by 48k.

 

Very interesting about the TrueMatch: they forgo the use of preamps in favor of 4 extra bits of sampling depth. I may be looking at the wrong thing though. This seems to suggest a 96k sampling rate (and only <20khz frequency range): http://usa.stagetec.com/index.php?option=com_content&view=article&id=205&Itemid=213

In spite of that, I still stand by the idea that 96k provides no benefit for <20khz audio.

I do think that Jay's and my thoughts align on most things (the exception being intermodulation in air). He has more practical experience of course, and I'm looking at it more theoretically.

Link to comment
Share on other sites

This is a great discussion. 

I listened to a talk by Rupert Neve a couple years ago where he stated his belief (again not proof, belief) that audio in the system above the audible frequencies does in fact have an effect on the quality of sound, and therefore he designes his products to have a flat frequency response well above the 20-20 spectrum.  He was going flat up to 100khz for a while and more recently has settled on flat to 50 or 60 khz.  His belief is based on years of listening tests and comparisons.

 

yes, you're logic seems flawless. Within the parameters you mentioned.
My idea was that 96kHz provides a finer resolution of the recorded audio and if a time-stretch is applied there is simply more information for the computer to stretch, rather than "filling in the blanks". I am completely wrong in this assumption?

If the <24khz audio is reproduced 100% accurately as stated by the Nyquist theorem (i believe this to be accurate and there is math to prove it), then additional sampling points are unnecessary, and do not further increase the accuracy of the reproduced waveform.

The exception? to this may be manipulation where higher sampling rates include frequencies above the audible spectrum that are brought down into the audible spectrum, but wouldn't exist sampled at 48k.

However, for dialog specifically, most delivery specs and playback specs of center channel content is band limited well within the audible spectrum, therefore I'm not convinced that the small amount of higher frequency content would survive through to reproduction.

Link to comment
Share on other sites

constantin: " to break the tie. "

sorry, this isn't an election.

but if I were voting I would vote that: 96k provides no discernable benefit for <20khz audio.

 

WE: " is reproduced 100% accurately as stated by the Nyquist theorem "

I believe the weakness may be in the "is", as there are technical issues, limitations, not accounted for in the theorem.

Edited by studiomprd
Link to comment
Share on other sites

constantin: " to break the tie. "

sorry, this isn't an election.

but if I were voting I would vote that: 96k provides no discernable benefit for <20khz audio.

WE: " is reproduced 100% accurately as stated by the Nyquist theorem "

I believe the weakness may be in the "is", as there are technical issues, limitations, not accounted for in the theorem.

Ok. Are you saying that recording at 96k necessarily accounts for those technical issues that processes like oversampling do not?

If so, what specifically?

Link to comment
Share on other sites

Sorry, Jay, I didn't intend to imply that you or others here did not have an authoritative voice. However, you and James, as well as the Senator to some degree, seem to disagree on some of these issues and I suppose I was kind of hoping for a third voice of expertise to break the tie. And yes, I would love to hear from a manufacturer of one of these devices and hear their considerations.

One answer is given by StageTech, I suppose, who arguably build the cleanest and most neutral a-d converter (TrueMatch) which is popular with classical music recordists. Its highest sampling rate is 48kHz.

 

 

You should read the white papers here. Dan Lavry designs some very highly regarded AD/DA's.

 

http://www.lavryengineering.com/lavry-white-papers/

 

It is possible that 96K is already on the wrong side of an ideal sampling rate when you take the practical/technical considerations into account. 

 

Nick

Link to comment
Share on other sites

Lavry's "The Optimal Sample Rate For Quality Audio" seems to suggest that 96k is unnecessary except that extra frequency response is desirable for two main reasons:

- To accommodate extra sensitive hearing

- So that the cumulative 20khz bandlimiting of each device in the chain does not end up affecting the audible range

 

He then suggests the optimal sampling rate should be 60k

To address both the Senator and the Lavry white paper:
Yes, there are many hardware related issues that could affect the quality of the A/D conversion. Poor filtering like Lavry suggests is one of them. Set aside the fact that technical issues can occur at every stage before and after A/D/A conversion (the TrueMatch elimination of preamps being demonstrative of that significance). Today's A/D converters sample at rates much higher than 48k (megasample rates, even). The conversion to 48k and 96k is a mathematical calculation derived from the same high rate sampling. The calculation includes a virtually perfect digital filter to avoid aliasing. Unless there are technical issues with the 48k calculation that don't exist in 96k calculation, any technical issues that exist affect 48k and 96k equally in today's converters.

I don't believe that humans can hear higher than 20khz, but again, the majority of equipment is bandlimited to 20khz.

Link to comment
Share on other sites

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.
Note: Your post will require moderator approval before it will be visible.

Guest
Reply to this topic...

×   Pasted as rich text.   Paste as plain text instead

  Only 75 emoji are allowed.

×   Your link has been automatically embedded.   Display as a link instead

×   Your previous content has been restored.   Clear editor

×   You cannot paste images directly. Upload or insert images from URL.


×
×
  • Create New...