Still not grasping this part of Sample Rates...

Have an urge to learn, or a calling to teach? Want to share some useful Youtube videos? Do it here!
RobC
Posts: 1832
Joined: 10 Mar 2018

11 Dec 2022

Selig gave some good links a while ago, but I still have trouble with a part of this article:

https://sonicscoop.com/the-science-of-s ... n-it-isnt/

Does a higher sample rate still run too fast on a modern system? Will there be still unwanted artifacts, when there's a ~20 kHz filter before the audio gets converted to analog? With software synths, can there really be such calculation inaccuracies happening?

If yes, then how doesn't the same happen with oversampling?

Or is it all about that it's easier to convert a normal 44.1 kHz sample rate audio to analog, than a much higher one?

(Don't get me wrong, I'd just stick to 34 kHz sample rate if I could, cause I can't hear a thing above 17 kHz.) The only reason I run R12 above 44.1 kHz, is because the sequencer has better reaction times due to how the CV system works internally. The rest, I'd just solve with oversampling, if I could. (Yeah, I know, RRP + Reaper works. But I prefer the Reason sequencer.)

Point is, if audio, which is above that ~60 kHz optimal sample rate, gets converted to analog, even if filtered, would mess up, say, a clean 100 Hz sine wave, right?

User avatar
Quarmat
Competition Winner
Posts: 458
Joined: 11 Feb 2021
Location: Europe

11 Dec 2022

When I read posts like yours I sense the heaviness of the amount of things I yet have to learn.

RobC
Posts: 1832
Joined: 10 Mar 2018

11 Dec 2022

Quarmat wrote:
11 Dec 2022
When I read posts like yours I sense the heaviness of the amount of things I yet have to learn.
I thought at first that you'd write that you sense the heaviness of my stupidity. x D

But this isn't really necessary knowledge ~ rather a workflow, efficiency, (and developer) kind of thing.

But of course I still hope to find the answer. Might be a small thing to know, but definitely saves me from doing stupid things.

avasopht
Competition Winner
Posts: 3931
Joined: 16 Jan 2015

11 Dec 2022

In audio you have the time domain (massive rise at 0.00001 seconds), and the frequency domain (between 0.00001s and 0.00002s the phase/frequency spectrum is calculated as XYZ).

When you up/downsample, there will be tradeoffs.

Some methods will preserve the frequencies (but not phase) within each processing window, but will introduce ringing.

You have overshoot and undershoot as well.

What does that mean to your 100hz sine wave? If it starts at 0.00001 seconds, who knows? Maybe the process makes it start at 0.000005 seconds instead. Maybe its phase has changed.

It won't make much difference to the timbre of a guitar but might impact its attack. And it will affect the waveform (in the time domain). Maybe the change introduces higher peaks somewhere (because the phase of the active frequencies all align where they didn't before).

In the best-case scenario, only phasing is impacted. But for the lowest frequencies, that may be more noticeable. And it is rarely the best case scenario.

To gain an intuition to down/up sampling, open up an image in Photoshop or GIMP, then increase the image size (that's up sampling), and then reduce it back to the original size (down sampling). For fun, compare applying effects to the image at its original size, and applying effects to the up sampled version and then down sampling back to the original size. Your image editor will probably ask how you want to downsample (bicubic, linear, nearest neighbour ... ... ... audio processing is exactly the same as they use the same signal processing algorithms).

I think Izotope made a Youtube video about this, demonstrating how their system reduces the amount of up/down sampling, while showing you the potential perils of every plugin up/down sampling and being fed into each other.

RobC
Posts: 1832
Joined: 10 Mar 2018

14 Dec 2022

avasopht wrote:
11 Dec 2022
In audio you have the time domain (massive rise at 0.00001 seconds), and the frequency domain (between 0.00001s and 0.00002s the phase/frequency spectrum is calculated as XYZ).

When you up/downsample, there will be tradeoffs.

Some methods will preserve the frequencies (but not phase) within each processing window, but will introduce ringing.

You have overshoot and undershoot as well.

What does that mean to your 100hz sine wave? If it starts at 0.00001 seconds, who knows? Maybe the process makes it start at 0.000005 seconds instead. Maybe its phase has changed.

It won't make much difference to the timbre of a guitar but might impact its attack. And it will affect the waveform (in the time domain). Maybe the change introduces higher peaks somewhere (because the phase of the active frequencies all align where they didn't before).

In the best-case scenario, only phasing is impacted. But for the lowest frequencies, that may be more noticeable. And it is rarely the best case scenario.

To gain an intuition to down/up sampling, open up an image in Photoshop or GIMP, then increase the image size (that's up sampling), and then reduce it back to the original size (down sampling). For fun, compare applying effects to the image at its original size, and applying effects to the up sampled version and then down sampling back to the original size. Your image editor will probably ask how you want to downsample (bicubic, linear, nearest neighbour ... ... ... audio processing is exactly the same as they use the same signal processing algorithms).

I think Izotope made a Youtube video about this, demonstrating how their system reduces the amount of up/down sampling, while showing you the potential perils of every plugin up/down sampling and being fed into each other.
Okay, now I understand why people recommended to stick to the sample rate of the planned release. That's pretty much always either 44.1 or 48 kHz.

The whole process seems almost like how FIR / IIR filters work, and what they do. In case of downsampling, they always are used.

So yeah, while there might be reasons to use either a higher sample rate, or oversampling; at the end of the day, it might do more harm than good.

While I think most devices can handle 48 kHz sample rate, I guess the most universal is 44.1 kHz, which is better to stick to.

All in all, thank you, this is definitely the kind of info I didn't know about sample rate conversion, and makes me worry less about using higher ones.

RobC
Posts: 1832
Joined: 10 Mar 2018

18 Dec 2022

avasopht wrote:
11 Dec 2022
In audio you have the time domain (massive rise at 0.00001 seconds), and the frequency domain (between 0.00001s and 0.00002s the phase/frequency spectrum is calculated as XYZ).

When you up/downsample, there will be tradeoffs.

Some methods will preserve the frequencies (but not phase) within each processing window, but will introduce ringing.

You have overshoot and undershoot as well.

What does that mean to your 100hz sine wave? If it starts at 0.00001 seconds, who knows? Maybe the process makes it start at 0.000005 seconds instead. Maybe its phase has changed.

It won't make much difference to the timbre of a guitar but might impact its attack. And it will affect the waveform (in the time domain). Maybe the change introduces higher peaks somewhere (because the phase of the active frequencies all align where they didn't before).

In the best-case scenario, only phasing is impacted. But for the lowest frequencies, that may be more noticeable. And it is rarely the best case scenario.

To gain an intuition to down/up sampling, open up an image in Photoshop or GIMP, then increase the image size (that's up sampling), and then reduce it back to the original size (down sampling). For fun, compare applying effects to the image at its original size, and applying effects to the up sampled version and then down sampling back to the original size. Your image editor will probably ask how you want to downsample (bicubic, linear, nearest neighbour ... ... ... audio processing is exactly the same as they use the same signal processing algorithms).

I think Izotope made a Youtube video about this, demonstrating how their system reduces the amount of up/down sampling, while showing you the potential perils of every plugin up/down sampling and being fed into each other.
Meanwhile I found a video where a guy explains it with an impulse - though couldn't finish watching, due to internet connection issues...

Now, even if there would be some sound design DAW that can run at 768 kHz sample rate, I can imagine painful things happening after conversion.
That said, I have to admit, after I was done working usually at 192 kHz, and finalizing the music to 16 bit 44.1 kHz, I never really noticed any damage. It may be more noticeable now that I work with maximum dynamic range, though.

That said, I talked about downsampling a 'loudness war' trash from my past, and downsampling it only 1 time.
If it constantly happens, up and down, I bet that can get messy fast.

One thing is still not clear, though. That one article says, anything running above 60 kHz sample rate, runs too fast, thus I presume would introduce artifacts, even with a low pass filter-on around 20 kHz.
I'm not sure what would be happening there; if the article is dated (?), etc.

I'll probably have to make a sine sweep test at the highest sample rate possible with softwares I have - as well as what my DAC can handle (768 kHz), then listen for artifacts. The better outcome would be silence.
I think the max is usually ~ 300+ kHz.

Note that I talk about extremely destructive sound design; so 1 wave cycle can be trashed very quickly. At 44.1 kHz, but even 192 kHz, the sound becomes lo-fi before I know it.

Stuff like mixing/mastering doesn't have much benefit, if I remember correctly, around more than 96 kHz.

Luckily, as long as we don't get to the point of adding all kinds of envelopes to the synthesized sound, it can tolerate downsampling more.
As such, I'd only work at a super high sample rate, when it's beneficial. (Pitch-distortion for example can be an incredibly cool effect. But also incredibly destructive.)

Jac459
Posts: 677
Joined: 29 Mar 2022
Location: Singapore
Contact:

18 Dec 2022

Hi,

I am no expert on the matter but interested by it and I can't resist to add my 2 cents.

First, for the reasons avasopht gave, converting from 44.1 to 48 is not oversampling but a quite destructive process reducing significantly the sound quality. Oversampling needs to be a multiple like 88.2 for usual cd and 96 for audio generally targeting video.

Second, I don't think soft synths have lesser good sound that hardware from output point of view. Quite the contrary, when you are inside your DAW, you keep always the same Sample rate, there is no AD conversion so clearly the sound is better in quality. In the past, I guess hardware expansive synths were having super high end DSP that was producing sound a PC could not. Now I think this is finished. I would guess an M1 can surpass any synth dsp....

On the pure oversampling aspect, for music production, I guess it makes definitely sense for samplers as it gives you capacity to go down in the octaves. For internal processing I guess what matters the most is the number of bits (dynamic range) mainly.

On the fact that super high resolution introduces artefacts I read a very interesting article demonstrating it. In particular imagine your tweeters, when they have to reproduce these frequencies they can't do well the work for the frequencies in your audible range.
Bitwig and RRP fanboy...

RobC
Posts: 1832
Joined: 10 Mar 2018

19 Dec 2022

Jac459 wrote:
18 Dec 2022
Hi,

I am no expert on the matter but interested by it and I can't resist to add my 2 cents.

First, for the reasons avasopht gave, converting from 44.1 to 48 is not oversampling but a quite destructive process reducing significantly the sound quality. Oversampling needs to be a multiple like 88.2 for usual cd and 96 for audio generally targeting video.

Second, I don't think soft synths have lesser good sound that hardware from output point of view. Quite the contrary, when you are inside your DAW, you keep always the same Sample rate, there is no AD conversion so clearly the sound is better in quality. In the past, I guess hardware expansive synths were having super high end DSP that was producing sound a PC could not. Now I think this is finished. I would guess an M1 can surpass any synth dsp....

On the pure oversampling aspect, for music production, I guess it makes definitely sense for samplers as it gives you capacity to go down in the octaves. For internal processing I guess what matters the most is the number of bits (dynamic range) mainly.

On the fact that super high resolution introduces artefacts I read a very interesting article demonstrating it. In particular imagine your tweeters, when they have to reproduce these frequencies they can't do well the work for the frequencies in your audible range.
Hey!

Yes, that's also useful info again!

Hmm, I guess I gotta make a decision then, and carefully choose the sample rates.

I heard that video needs 48 kHz, and most playback systems support it, too, so that will be my final sample rate, then, after mastering was finished.

That sounds cool with the M stuff by Apple. Would be fun to discover, as I've always been curious - only the previous Apple devices didn't make up for the price before this M tech, and software running natively.

(On a side note: One mastering guy does what I kind of considered: recording the analog output of the DAC. - Not sure if that would help any though.)

I'm glad to see that ~ perhaps you're the first who is willing to admit that with samplers, a higher sample rate does matter. : )

If ultra frequencies are passed on, then true, that would create artifacts. But my DAC for example does use a low pass filter at 24 kHz. I think it's linear phase. I wondered why so high, but that might be to keep the pre and after ringing outside the audible range.

I do have an Ableton Live Lite, too, which should run the RRP and supports to run at 384 sample rate. Which should be useful for synthesizing, and detailed editing during sound design.

Thank you very much, too, for the detailed response!

Jac459
Posts: 677
Joined: 29 Mar 2022
Location: Singapore
Contact:

19 Dec 2022

RobC wrote:
19 Dec 2022
Hey!

Yes, that's also useful info again!

Hmm, I guess I gotta make a decision then, and carefully choose the sample rates.

I heard that video needs 48 kHz, and most playback systems support it, too, so that will be my final sample rate, then, after mastering was finished.

That sounds cool with the M stuff by Apple. Would be fun to discover, as I've always been curious - only the previous Apple devices didn't make up for the price before this M tech, and software running natively.

(On a side note: One mastering guy does what I kind of considered: recording the analog output of the DAC. - Not sure if that would help any though.)

I'm glad to see that ~ perhaps you're the first who is willing to admit that with samplers, a higher sample rate does matter. : )

If ultra frequencies are passed on, then true, that would create artifacts. But my DAC for example does use a low pass filter at 24 kHz. I think it's linear phase. I wondered why so high, but that might be to keep the pre and after ringing outside the audible range.

I do have an Ableton Live Lite, too, which should run the RRP and supports to run at 384 sample rate. Which should be useful for synthesizing, and detailed editing during sound design.

Thank you very much, too, for the detailed response!
Continuing the discussion:

I heard that video needs 48 kHz, and most playback systems support it, too, so that will be my final sample rate, then, after mastering was finished. ==> 48khz seems to be a new trend. Just you should be careful, if you need to provide both formats, of the conversion 44 to 48 or 48 to 44 because a lot of interpolation is going on. If you are coming from 88 or 96, converting back to 44 or 48 should have less issue. 88 to 44 is perfect, no interpolation, 96 to 48, same. 88 to 48 or 96 to 44 are not perfect but still better than 44 vs 48 in terms of interpolation,
Interpolations are mostly generating noise.

That sounds cool with the M stuff by Apple ==> I am the contrary of an apple fan boy but the M is just crushing the competition... hope it will change one day. Until then, there is not much choice at least for laptops...

I'm glad to see that ~ perhaps you're the first who is willing to admit that with samplers, a higher sample rate does matter. : ) ==> I may be wrong but my reasoning is that if you are not in a case of multisampler but just one sample you play with, if you go one octave down you are at 22khz already. I think this is better to take a bit of contingency, again when it will be in the mix there will be less interpolations.

For RRP, I do agree, I think it is better to stay at higher sample rate during synthesis and internally before output.
Bitwig and RRP fanboy...

RobC
Posts: 1832
Joined: 10 Mar 2018

20 Dec 2022

Jac459 wrote:
19 Dec 2022
RobC wrote:
19 Dec 2022
Hey!

Yes, that's also useful info again!

Hmm, I guess I gotta make a decision then, and carefully choose the sample rates.

I heard that video needs 48 kHz, and most playback systems support it, too, so that will be my final sample rate, then, after mastering was finished.

That sounds cool with the M stuff by Apple. Would be fun to discover, as I've always been curious - only the previous Apple devices didn't make up for the price before this M tech, and software running natively.

(On a side note: One mastering guy does what I kind of considered: recording the analog output of the DAC. - Not sure if that would help any though.)

I'm glad to see that ~ perhaps you're the first who is willing to admit that with samplers, a higher sample rate does matter. : )

If ultra frequencies are passed on, then true, that would create artifacts. But my DAC for example does use a low pass filter at 24 kHz. I think it's linear phase. I wondered why so high, but that might be to keep the pre and after ringing outside the audible range.

I do have an Ableton Live Lite, too, which should run the RRP and supports to run at 384 sample rate. Which should be useful for synthesizing, and detailed editing during sound design.

Thank you very much, too, for the detailed response!
Continuing the discussion:

I heard that video needs 48 kHz, and most playback systems support it, too, so that will be my final sample rate, then, after mastering was finished. ==> 48khz seems to be a new trend. Just you should be careful, if you need to provide both formats, of the conversion 44 to 48 or 48 to 44 because a lot of interpolation is going on. If you are coming from 88 or 96, converting back to 44 or 48 should have less issue. 88 to 44 is perfect, no interpolation, 96 to 48, same. 88 to 48 or 96 to 44 are not perfect but still better than 44 vs 48 in terms of interpolation,
Interpolations are mostly generating noise.

That sounds cool with the M stuff by Apple ==> I am the contrary of an apple fan boy but the M is just crushing the competition... hope it will change one day. Until then, there is not much choice at least for laptops...

I'm glad to see that ~ perhaps you're the first who is willing to admit that with samplers, a higher sample rate does matter. : ) ==> I may be wrong but my reasoning is that if you are not in a case of multisampler but just one sample you play with, if you go one octave down you are at 22khz already. I think this is better to take a bit of contingency, again when it will be in the mix there will be less interpolations.

For RRP, I do agree, I think it is better to stay at higher sample rate during synthesis and internally before output.
Many thanks for even more great pointers!

I consider that music and other audio content usually gets uploaded to Youtube, or gets used in video games, so 48 kHz will be the most universal.
Plus I'm used to the 96, 192, 384, 768 numbers anyway. : )

I'm still a desktop PC user. I'm not a nut about Apple either. : ) But maybe Windows will try to keep up, eventually.

Correct, I don't multi-sample. I prefer if the sound is the same on all keys - in case of synthesized sounds at least. The change of pitch adds a flavor of its own, or Mimic's time stretching can be interesting, too. I rather like to shape one sound, then when it's all done, and the pitch and/or time stretching is set, I can multi-sample / export and resample to 48 kHz.
Normally, synthesized sounds aren't all that different for multiple notes or different hits, unless the oscillators' phase run freely, like with Subtractor. Plus, 1 sample spread across the keyboard, will be perfectly in sync, which opens even more possibilities and control.

With your info, I'm a bit more confident about trying even Reaper, since it can at least oversample up to 768 kHz; and run at 384 kHz at least.

User avatar
selig
RE Developer
Posts: 11685
Joined: 15 Jan 2015
Location: The NorthWoods, CT, USA

20 Dec 2022

Jac459 wrote:
18 Dec 2022
First, for the reasons avasopht gave, converting from 44.1 to 48 is not oversampling but a quite destructive process reducing significantly the sound quality. Oversampling needs to be a multiple like 88.2 for usual cd and 96 for audio generally targeting video.
That’s not at all how sample rate conversion works, or are you speaking from experience? It’s not destructive because there is an intermediate step where the signal is up-sampled to a very high sample rate and then down sampled back down. Because of the common frequency used, you can convert “cleanly” from any rate to any other rate this way. So it’s no simpler for one conversion than another.
IIRC, VERY early sample rate converters didn’t do this, so there used to be an issue converting between odd rates. But this hasn’t been used for many years now.
Of course, there’s no sonic advantage to increasing the sample rate, but there is also no harm done either.
This is how a modern DAW like Reason (and LUNA, maybe others) work as “sample rate agnostic” machines, meaning you can throw in a bunch of different sample rate audio files and they automatically convert to the current rate. So if conversion significantly reduced audio quality none of these DAWs would/could ‘work’…
:)
Selig Audio, LLC

avasopht
Competition Winner
Posts: 3931
Joined: 16 Jan 2015

20 Dec 2022

While samplerate conversion is not destructive, you can get preringing from linear phase filters like those used in FabFilter (which also has the option of minimum phase filters).

This is what I meant (which shouldn't be confused with a loss of audio quality, as Selig explained).



So, while I wouldn't say anything in opposition to Selig, I do think it's important to know these processes aren't 100% free and can introduce subtle differences (and this video demonstrates what this can do with multiple mics on a drum kit).

That being said, I would expect DAWs to favour minimum phase filters over linear phase for oversampling.
Last edited by avasopht on 20 Dec 2022, edited 1 time in total.

Jac459
Posts: 677
Joined: 29 Mar 2022
Location: Singapore
Contact:

20 Dec 2022

selig wrote:
20 Dec 2022
That’s not at all how sample rate conversion works, or are you speaking from experience? It’s not destructive because there is an intermediate step where the signal is up-sampled to a very high sample rate and then down sampled back down. Because of the common frequency used, you can convert “cleanly” from any rate to any other rate this way. So it’s no simpler for one conversion than another.
IIRC, VERY early sample rate converters didn’t do this, so there used to be an issue converting between odd rates. But this hasn’t been used for many years now.
Of course, there’s no sonic advantage to increasing the sample rate, but there is also no harm done either.
This is how a modern DAW like Reason (and LUNA, maybe others) work as “sample rate agnostic” machines, meaning you can throw in a bunch of different sample rate audio files and they automatically convert to the current rate. So if conversion significantly reduced audio quality none of these DAWs would/could ‘work’…
:)
I kind of disagree here. Let me explain myself more in detail and please correct me if you still think I am wrong. My knowledge indeed come from university and could be outdated.

When you are upsampling a source, you are in fact in reality, adding data point to a curve that has normally only measured data points. With 44.1khz, you measure the sound 44,100 times per second, most often on a scale of 16bits. If you upsample to 88.2khz. What you do is to add between each point, a calculated point. At a very basic level, this point could be the one in the middle of the line between 2 measured points, but of course, mathematics come to the rescue here thanks to interpolation and try to deduct the real position of the point thanks to a more global context.
But yet, it is a calculated point. Not a measured point.
It is easy to try by yourself upsampling algorithms by trying an 11khz source and try to upsample to 22 or 44 kHz. Clearly you should hear very well the difference between an upsampled and native 44.1khz material.

So going back to my simplified statement, you are right to say that converting from 44.1 to 48 is done smartly using upsampling and then interpolation, but yet, you enter fully in the real of calculated interpolated points and you are loosing in quality as it is not measured points. Same if you are moving from say 96 to 44. But then, the loss of quality should be extremely tiny. If you move from 88 to 44. You won't have loss at all in the sense that your downsampled 44 will have the same quality than a native 44.
That's what I meant in my previous post.
You still disagree?
Bitwig and RRP fanboy...

Jac459
Posts: 677
Joined: 29 Mar 2022
Location: Singapore
Contact:

20 Dec 2022

avasopht wrote:
20 Dec 2022
While samplerate conversion is not destructive, you can get preringing from linear phase filters like those used in FabFilter (which also has the option of minimum phase filters).

This is what I meant (which shouldn't be confused with a loss of audio quality, as Selig explained).



So, while I wouldn't say anything in opposition to Selig, I do think it's important to know these processes aren't 100% free and can introduce subtle differences (and this video demonstrates what this can do with multiple mics on a drum kit).

That being said, I would expect DAWs to favour minimum phase filters over linear phase for oversampling.
Interested to understand why samplerate conversion is not destructive, that's not my understanding of sound processing. Do I miss something? I read about linear vs minimum phase but I don't see the link.
Bitwig and RRP fanboy...

avasopht
Competition Winner
Posts: 3931
Joined: 16 Jan 2015

20 Dec 2022

Jac459 wrote:
20 Dec 2022
I kind of disagree here. Let me explain myself more in detail and please correct me if you still think I am wrong. My knowledge indeed come from university and could be outdated.

When you are upsampling a source, you are in fact in reality, adding data point to a curve that has normally only measured data points. With 44.1khz, you measure the sound 44,100 times per second, most often on a scale of 16bits. If you upsample to 88.2khz. What you do is to add between each point, a calculated point. At a very basic level, this point could be the one in the middle of the line between 2 measured points, but of course, mathematics come to the rescue here thanks to interpolation and try to deduct the real position of the point thanks to a more global context.
But yet, it is a calculated point. Not a measured point.
It is easy to try by yourself upsampling algorithms by trying an 11khz source and try to upsample to 22 or 44 kHz. Clearly you should hear very well the difference between an upsampled and native 44.1khz material.
Recall that we have a wave/frequency spectrum duality in audio (very similar to the wave/particle duality in physics).

While we currently represent audio as PCM waves, that is only a method of representation.

We could just as easily (well, not easily) represent the audio in the frequency spectrum (using density over time).

In fact, I wouldn't even call our PCM wave a representation of the wave, but a view of the wave from the sample rate's point of view. They are just samples, after all, and not the wave itself.

At a different sample rate, you will see a difference PCM representation, but the wave itself has not changed - nor has its sound. The PCM waveform values are merely showing you samples of the same wave ... just at a different sample rate.

The points you see in a PCM waveform are just samples of the wave, not the wave itself. And so, seeing a different series of PCM values at different sample rates does not mean the wave itself has changed at all.

And these are linear processes, so they can in theory be used to perfectly reverse the process (unless you've completely removed frequencies).

However, as mentioned in my previous comment, artifacts can be introduced during the process depending on the method used.

There are upsampling and downsampling filters for doubling and halving the sample rate that might not be subject to the same range of artefacts, but that's something pitchblende or Robotic Bean would be better positioned to elaborate on.

avasopht
Competition Winner
Posts: 3931
Joined: 16 Jan 2015

20 Dec 2022

Jac459 wrote:
20 Dec 2022
Interested to understand why samplerate conversion is not destructive, that's not my understanding of sound processing. Do I miss something? I read about linear vs minimum phase but I don't see the link.
Well, it's more to do with the process being linear (not to be confused with linear phase), which means you can apply the operations in different orders and yield the same result.

For instance ... the effects series of EQ1, reverb 1, echo 1, EQ 2, reverb 2, echo 2 would yield the exact same result as the series echo 2, EQ 1, reverb 1, EQ 2, reverb 2, echo 1 (providing there are no non-linear effects, which you can find in some reverb and echo units).

Also, linear processes do not add frequencies, and oversampling does not add frequencies (could you imagine if it did?).

That doesn't directly explain why it's not destructive ... ... unless you focus only on the frequency spectrum, and you'll see it remains perfectly intact!

The PCM wave values are indirect pointers to the true wave (frequency spectrum over time), so apart from the inaudible frequencies filtered during downsampling, your frequency spectrum remains unaltered and you should be able to restore the original if you wanted through inverse filters).
Last edited by avasopht on 20 Dec 2022, edited 1 time in total.

Jac459
Posts: 677
Joined: 29 Mar 2022
Location: Singapore
Contact:

20 Dec 2022

avasopht wrote:
20 Dec 2022
Jac459 wrote:
20 Dec 2022
I kind of disagree here. Let me explain myself more in detail and please correct me if you still think I am wrong. My knowledge indeed come from university and could be outdated.

When you are upsampling a source, you are in fact in reality, adding data point to a curve that has normally only measured data points. With 44.1khz, you measure the sound 44,100 times per second, most often on a scale of 16bits. If you upsample to 88.2khz. What you do is to add between each point, a calculated point. At a very basic level, this point could be the one in the middle of the line between 2 measured points, but of course, mathematics come to the rescue here thanks to interpolation and try to deduct the real position of the point thanks to a more global context.
But yet, it is a calculated point. Not a measured point.
It is easy to try by yourself upsampling algorithms by trying an 11khz source and try to upsample to 22 or 44 kHz. Clearly you should hear very well the difference between an upsampled and native 44.1khz material.
Recall that we have a wave/frequency spectrum duality in audio (very similar to the wave/particle duality in physics).

While we currently represent audio as PCM waves, that is only a method of representation.

We could just as easily (well, not easily) represent the audio in the frequency spectrum (using density over time).

In fact, I wouldn't even call our PCM wave a representation of the wave, but a view of the wave from the sample rate's point of view. They are just samples, after all, and not the wave itself.

At a different sample rate, you will see a difference PCM representation, but the wave itself has not changed - nor has its sound. The PCM waveform values are merely showing you samples of the same wave ... just at a different sample rate.

The points you see in a PCM waveform are just samples of the wave, not the wave itself. And so, seeing a different series of PCM values at different sample rates does not mean the wave itself has changed at all.

And these are linear processes, so they can in theory be used to perfectly reverse the process (unless you've completely removed frequencies).

However, as mentioned in my previous comment, artifacts can be introduced during the process depending on the method used.

There are upsampling and downsampling filters for doubling and halving the sample rate that might not be subject to the same range of artefacts, but that's something pitchblende or Robotic Bean would be better positioned to elaborate on.
That's why pcm is not the best way to encode sound, dsd is a better one. Or we could even talk about the controversial mqa.
Yet, in audio processing, most of the time, we are in pcm realm. You are right to say that pcm is only one way to measure a sound. But it is still a measure. And a measure is still better than an interpolated/calculated point. If it wasn't the case, we could restore 11khz recording to 44 and get an awesome quality... unfortunately we can't.
Bitwig and RRP fanboy...

Jac459
Posts: 677
Joined: 29 Mar 2022
Location: Singapore
Contact:

20 Dec 2022

Actually Avasopht, I think we are having a quiproquo because I read again your post of the 11th where you say exactly what I am trying to say.
So to summarise very simply. I am just reacting to Selig point saying than changing sample rate from 44.1 to 48 or 48 to 44.1 is not damaging because we go through intermediary upsampling. I disagree because even if I am well aware on the fact that some interpolation algorithms will do some damage control, yet, your pcm representation won't be measured anymore but calculated.
Bitwig and RRP fanboy...

User avatar
selig
RE Developer
Posts: 11685
Joined: 15 Jan 2015
Location: The NorthWoods, CT, USA

20 Dec 2022

@jac459
If I change one sample out of millions, the waveform is no longer accurate - but can you hear it?
I was responding to this specific claim about sample rate conversion being:
“a quite destructive process reducing significantly the sound quality”.
What is the difference between altering one sample and altering many samples? There is no way to accurately qualify issues with sample rate conversion, but when you say it’s is QUITE DESTRUCTIVE and SIGNIFICANTLY REDUCING the sound quality, I have to assume it’s something you WILL hear and hear BIG TIME. But that’s not been my experience at all.
So is sample rate conversion 100% transparent? No, but neither does ANY sample rate currently available capture ALL frequencies!
But in both cases it’s not necessary at all, because ‘humans’.
For the frequencies, we don’t need infinite bandwidth because we humans don’t hear it. For the sample rate conversion all we need to do it keep any artifacts well below the threshold of human detection. It’s the human factor that is important to consider in all of these cases.
The other thing about sample rate conversion, is you should only need to do it once, maybe twice if you’re not being careful. What I mean is sure, if you convert back and forth hundreds of times you’ll hear the artifacts because they will be cumulative. But we don’t do that, again one or two conversions (if any) is the most that typically happens in most workflows. Even if you switch sample rates back and forth in Reason, the original audio file is always used to create the new sample rate audio file so it’s only ever converted once.

So yes there are artifacts, no we can’t hear them in today’s well designed convertor algorithms in my experience. :)
Selig Audio, LLC

avasopht
Competition Winner
Posts: 3931
Joined: 16 Jan 2015

20 Dec 2022

Jac459 wrote:
20 Dec 2022
That's why pcm is not the best way to encode sound, dsd is a better one. Or we could even talk about the controversial mqa.
Yet, in audio processing, most of the time, we are in pcm realm. You are right to say that pcm is only one way to measure a sound. But it is still a measure. And a measure is still better than an interpolated/calculated point. If it wasn't the case, we could restore 11khz recording to 44 and get an awesome quality... unfortunately we can't.
We don't use 11 khz because it can't represent frequencies above 5.5 khz.

If, however your audio is definitely below 5.5 khz, it's fine to use as explained by Nyquist.

An interpolated/calculated point is neither better nor worse. What matters is whether the sample rate conversion has lost information or not.

Note: interpolation is a solved problem where there is only one correct answer. So the calculation can be 100% precise.

Think of it like this, a 100 hz sinewave is just a 100 hz. If you were to pick any point in time you can ask: what is the amplitude and phase at 100 hz, and there is only 1 answer (either 0 amplitude and n/a phase, or some amplitude and some phase).

When you represent it as PCM, so long as you have a sample rate above 200 hz, you can always answer this question using the PCM values.

When you interpolate/calculate the PCM value in a different sample rate, if you are able to yield the same answer from the converted data, absolutely nothing has been lost (even though the sample values were interpolated/calculated).

Jac459
Posts: 677
Joined: 29 Mar 2022
Location: Singapore
Contact:

20 Dec 2022

selig wrote:
20 Dec 2022
@jac459
If I change one sample out of millions, the waveform is no longer accurate - but can you hear it?
I was responding to this specific claim about sample rate conversion being:
“a quite destructive process reducing significantly the sound quality”.
What is the difference between altering one sample and altering many samples? There is no way to accurately qualify issues with sample rate conversion, but when you say it’s is QUITE DESTRUCTIVE and SIGNIFICANTLY REDUCING the sound quality, I have to assume it’s something you WILL hear and hear BIG TIME. But that’s not been my experience at all.
So is sample rate conversion 100% transparent? No, but neither does ANY sample rate currently available capture ALL frequencies!
But in both cases it’s not necessary at all, because ‘humans’.
For the frequencies, we don’t need infinite bandwidth because we humans don’t hear it. For the sample rate conversion all we need to do it keep any artifacts well below the threshold of human detection. It’s the human factor that is important to consider in all of these cases.
The other thing about sample rate conversion, is you should only need to do it once, maybe twice if you’re not being careful. What I mean is sure, if you convert back and forth hundreds of times you’ll hear the artifacts because they will be cumulative. But we don’t do that, again one or two conversions (if any) is the most that typically happens in most workflows. Even if you switch sample rates back and forth in Reason, the original audio file is always used to create the new sample rate audio file so it’s only ever converted once.

So yes there are artifacts, no we can’t hear them in today’s well designed convertor algorithms in my experience. :)
We fully agree then... 😀.

"Significantly reducing" was indeed a bit too strong but here it is a question of point of view.
Audiophiles will chase any risk of sound quality reduction even buying 10kUSD cables or stuff like that. From this point of view, as a music producer you may want to pay special attention.
But yes, you are right, most humans, including me, won't hear a difference...
Bitwig and RRP fanboy...

Jac459
Posts: 677
Joined: 29 Mar 2022
Location: Singapore
Contact:

20 Dec 2022

avasopht wrote:
20 Dec 2022
Jac459 wrote:
20 Dec 2022
That's why pcm is not the best way to encode sound, dsd is a better one. Or we could even talk about the controversial mqa.
Yet, in audio processing, most of the time, we are in pcm realm. You are right to say that pcm is only one way to measure a sound. But it is still a measure. And a measure is still better than an interpolated/calculated point. If it wasn't the case, we could restore 11khz recording to 44 and get an awesome quality... unfortunately we can't.
We don't use 11 khz because it can't represent frequencies above 5.5 khz.

If, however your audio is definitely below 5.5 khz, it's fine to use as explained by Nyquist.

An interpolated/calculated point is neither better nor worse. What matters is whether the sample rate conversion has lost information or not.

Note: interpolation is a solved problem where there is only one correct answer. So the calculation can be 100% precise.

Think of it like this, a 100 hz sinewave is just a 100 hz. If you were to pick any point in time you can ask: what is the amplitude and phase at 100 hz, and there is only 1 answer (either 0 amplitude and n/a phase, or some amplitude and some phase).

When you represent it as PCM, so long as you have a sample rate above 200 hz, you can always answer this question using the PCM values.

When you interpolate/calculate the PCM value in a different sample rate, if you are able to yield the same answer from the converted data, absolutely nothing has been lost (even though the sample values were interpolated/calculated).
I am well aware of Nyquist. It seems you don't get my point. I still think we are stuck in a quiproquo.

I fully disagree on this point though: "interpolation is a solved problem where there is only one correct answer."
Interpolation is a solved problem, but it is absolutely better to measure... what you are recording 44k times by second is not a sin wave but the sum of thousands of sin waves....
Bitwig and RRP fanboy...

avasopht
Competition Winner
Posts: 3931
Joined: 16 Jan 2015

20 Dec 2022

Jac459 wrote:
20 Dec 2022
Interpolation is a solved problem, but it is absolutely better to measure... what you are recording 44k times by second is not a sin wave but the sum of thousands of sin waves....
This is also true of thousands of sine waves.

That's the beauty of it all, and even why processing PCM waves even works (when you think about it more deeply).

The other way to understand it is this ...

One wave produces one frequency spectrum.

That one frequency spectrum can tell you at any samplerate what the value is at any given time. Hence why there is always only one precise answer, and can be calculated accurately from the wave.

avasopht
Competition Winner
Posts: 3931
Joined: 16 Jan 2015

20 Dec 2022

That all being said, probably best to ignore all the descriptions I've given.

What Selig said is the most important (emphasis mine) ...
selig wrote:
20 Dec 2022
If I change one sample out of millions, the waveform is no longer accurate - but can you hear it?
I was responding to this specific claim about sample rate conversion being:
“a quite destructive process reducing significantly the sound quality”.
What is the difference between altering one sample and altering many samples? There is no way to accurately qualify issues with sample rate conversion, but when you say it’s is QUITE DESTRUCTIVE and SIGNIFICANTLY REDUCING the sound quality, I have to assume it’s something you WILL hear and hear BIG TIME. But that’s not been my experience at all.
So is sample rate conversion 100% transparent? No, but neither does ANY sample rate currently available capture ALL frequencies!
But in both cases it’s not necessary at all, because ‘humans’.
For the frequencies, we don’t need infinite bandwidth because we humans don’t hear it. For the sample rate conversion all we need to do it keep any artifacts well below the threshold of human detection. It’s the human factor that is important to consider in all of these cases.
The other thing about sample rate conversion, is you should only need to do it once, maybe twice if you’re not being careful. What I mean is sure, if you convert back and forth hundreds of times you’ll hear the artifacts because they will be cumulative. But we don’t do that, again one or two conversions (if any) is the most that typically happens in most workflows. Even if you switch sample rates back and forth in Reason, the original audio file is always used to create the new sample rate audio file so it’s only ever converted once.

So yes there are artifacts, no we can’t hear them in today’s well designed convertor algorithms in my experience. :)

Jac459
Posts: 677
Joined: 29 Mar 2022
Location: Singapore
Contact:

20 Dec 2022

I thing you didn't read my last answers mate :-).
Bitwig and RRP fanboy...

Post Reply
  • Information
  • Who is online

    Users browsing this forum: Google [Bot] and 8 guests