> it's probably worth avoiding the resampling of 44.1 to 48 kHz
Ehhm, yeah, duh? You don't resample unless there is a clear need, and even then you don't upsample and only downsample, and you tell anyone that tries to convince you otherwise to go away and find the original (analog) source, so you can do a proper transfer.
I'm working on a game. My game stores audio files as 44.1kHz .ogg files. If my game is the only thing playing audio, then great, the system sound mixer can configure the DAC to work in 44.1kHz mode.
But if other software is trying to play 48kHz sound files at the same time? Either my game has to resample from 44.1kHz to 48kHz before sending it to the system, or the system sound mixer needs to resample it to 48kHz, or the system sound mixer needs to resample the other software from 48kHz to 44.1kHz.
If 44.1kHz is otherwise sufficient but you have a downstream workflow that is incompatible, there are arguments for doing this. It can be done with no loss in quality.
From an information theory perspective, this is like putting a smaller pipe right through the middle of a bigger one. The channel capacity is the only variable that is changing and we are increasing it.
That seems a rather shallow - and probably incorrect - reading of the source. This is an efficiency and trust trade off as noted:
> given sufficient computing resources, we can resample 44.1 kHz to 48 kHz perfectly. No loss, no inaccuracies.
and then further
> Your smartphone probably can resample 44.1 kHz to 48 kHz in such a way that the errors are undetectable even in theory, because they are smaller than the noise floor. Proper audio equipment can certainly do so.
That is you don't need the original source to do a proper transfer. The author is simply noting
> Although this conversion can be done in such a way as to produce no audible errors, it's hard to be sure it actually is.
That is that re-sampling is not a bad idea in this case because it's going to have any sort of error if done properly, it's just that the Author notes you cannot trust any random given re-sampler to do so.
Therefore if you do need to resample, you can do so without the analog source, as long as you have a re-sampler you can trust, or do it yourself.
A very common clear need is incorporating 44.1khz audio sourcesinto video. 48khz is 48khz because 48khz divided by 24fps, 25fps, or 30fps is an integer (and 44.1khz is not).
Also, for decades upsampling on ingest and downsampling on egress has been standard practice for DSP because it reduces audible artifacts from truncation and other rounding techniques.
Finally, most recorded sound does not have an original analog source because of the access digital recording has created…youtube for example.
Lots of Live/Audigy era Creative sound cards would resample everything to 48kHz, with probably one of the worst quality resamplers available, to the chagrin of all bitperfect fanatics... still probably one of their best selling sound cards.
> All audio signal is _perfectly_ represented in a digital form
What? No. All bandwidth limited signal is. Which means periodic. Causal signals like audio can be approximated, with tradeoffs. Such as pre-ringing (look at sinc(x), used to reconstruct sampled signal — how much energy is in the limb preceding the x=0.)
Is the approximation achieved by filtering the 44.1kHz DAC good enough? Yes, yes it is. But the math is way more involved (i.e. beyond me) than simply "Niquist".
This popular myth that limited frequencies we can hear and limited frequencies in Fourier transform sense is the same thing is quite irritating.
> We do [cubic curve fitting] all the time in image processing, and it works very well. It would probably work well for audio as well, although it's not used -- not in the same form, anyway -- in these applications.
Is there a reason the solution that "works very well" for images isn't/can't be applied to audio?
I'm not sure I understand the "just generate it" perspective. If you want to generate a much higher sampling rate signal that has a common multiple of your input and output sampling rate, "just generating it" is going to involve some kind of interpolation, no? Because you're trying to make data that isn't there.
If you want to change the number of slices of pizza, you can't simply just make 160x more pizza out of thin air.
Personally I'd just do a cubic resample if absolutely required (ideally you don't resample ofc); it's fast and straightforward.
Edit: serves me right for posting, I gotta get off this site.
I wonder if this problem could be "solved" by having some kind of "dual mode" DACs that can accept two streams of audio at different sample rates, likely 44.1khz and 48khz, which are converted to analog in parallel and then mixed back together at the analog output.
Then at the operating system level rather than mixing everything to a single audio stream at a single sample rate you group each stream that is at or a multiple of either 44.1khz or 48khz and then finally sends both streams to this "dual dac", thus eliminating the need to resample any 44.1khz or 48khz stream, or even vastly simplifying the resample of any sample rate that is a multiple of this.
As a real world example, on Windows, unless you take exclusive access of the audio output device, everything is already resampled to 48khz in the mixer. Well, technically it gets resampled to the default configured device sample rate, but I haven't seen anything other than 48khz in at least a decade if ever. Practically this is a non-issue, though I could understand wanting bit-perfect reproduction of a 44.1 khz source.
I'm kinda shocked that there's no discussion of sinc interpolation and adapting it's theoretical need for infinite signals to some finite kernel length.
For a sampled signal, if you know the sampling satisfied Nyquist (i.e., there was no frequency content above fs/2) then the original signal can be reproduced exactly at any point in time using sinc interpolation. Unfortunately that theoretically requires an infinite length sample, but the kernel can be bounded based on accuracy requirements or other limiting factors (such as the noise which was mentioned). Other interpolation techniques should be viewed as approximations to sinc.
Sinc interpolation is available on most oscilloscopes and is useful when the sample rate is sufficient but not greatly higher than the signal of interest.
If you're taking something from 44.1 to 48, only 91.875% of the data is real, so 8.125% of the resulting upsampled data is invented. Some of it will correlate with the original, real sound. If you use upsampling functions tuned to features of the audio - style, whether it's music, voice, bird recordings, NYC traffic, known auditorium, etc, you can probably bring the accuracy up by several percent. If the original data already has the optimizations, it'll be closer to 92%.
If it's really good AI upsampling, you might get qualitatively "better" sounding audio than the original but still technically deviates from the original baseline by ~8%. Conversely, there'll be technically "correct" upsampling results with higher overall alignment with the original that can sound awful.
There's still a lot to audio processing that's more art than science.
> In reality, the amount of precision that can actually be "heard" by the human ear probably lies between 18 and 21 bits; we don't actually know, because it's impossible to test.
This sounds contradictory - what would be the precision that can be heard in a test then?
22 comments
[ 3.6 ms ] story [ 46.3 ms ] threadEhhm, yeah, duh? You don't resample unless there is a clear need, and even then you don't upsample and only downsample, and you tell anyone that tries to convince you otherwise to go away and find the original (analog) source, so you can do a proper transfer.
I'm working on a game. My game stores audio files as 44.1kHz .ogg files. If my game is the only thing playing audio, then great, the system sound mixer can configure the DAC to work in 44.1kHz mode.
But if other software is trying to play 48kHz sound files at the same time? Either my game has to resample from 44.1kHz to 48kHz before sending it to the system, or the system sound mixer needs to resample it to 48kHz, or the system sound mixer needs to resample the other software from 48kHz to 44.1kHz.
Unless I'm missing something?
From an information theory perspective, this is like putting a smaller pipe right through the middle of a bigger one. The channel capacity is the only variable that is changing and we are increasing it.
> given sufficient computing resources, we can resample 44.1 kHz to 48 kHz perfectly. No loss, no inaccuracies.
and then further
> Your smartphone probably can resample 44.1 kHz to 48 kHz in such a way that the errors are undetectable even in theory, because they are smaller than the noise floor. Proper audio equipment can certainly do so.
That is you don't need the original source to do a proper transfer. The author is simply noting
> Although this conversion can be done in such a way as to produce no audible errors, it's hard to be sure it actually is.
That is that re-sampling is not a bad idea in this case because it's going to have any sort of error if done properly, it's just that the Author notes you cannot trust any random given re-sampler to do so.
Therefore if you do need to resample, you can do so without the analog source, as long as you have a re-sampler you can trust, or do it yourself.
Also, for decades upsampling on ingest and downsampling on egress has been standard practice for DSP because it reduces audible artifacts from truncation and other rounding techniques.
Finally, most recorded sound does not have an original analog source because of the access digital recording has created…youtube for example.
I.e. no one cares.
I am ashamed to admit this took me a long time to properly understand. For further reading I'd recommend:
https://people.xiph.org/~xiphmont/demo/neil-young.html https://www.youtube.com/watch?v=cIQ9IXSUzuM
What? No. All bandwidth limited signal is. Which means periodic. Causal signals like audio can be approximated, with tradeoffs. Such as pre-ringing (look at sinc(x), used to reconstruct sampled signal — how much energy is in the limb preceding the x=0.)
Is the approximation achieved by filtering the 44.1kHz DAC good enough? Yes, yes it is. But the math is way more involved (i.e. beyond me) than simply "Niquist".
This popular myth that limited frequencies we can hear and limited frequencies in Fourier transform sense is the same thing is quite irritating.
Is there a reason the solution that "works very well" for images isn't/can't be applied to audio?
If you want to change the number of slices of pizza, you can't simply just make 160x more pizza out of thin air.
Personally I'd just do a cubic resample if absolutely required (ideally you don't resample ofc); it's fast and straightforward.
Edit: serves me right for posting, I gotta get off this site.
Then at the operating system level rather than mixing everything to a single audio stream at a single sample rate you group each stream that is at or a multiple of either 44.1khz or 48khz and then finally sends both streams to this "dual dac", thus eliminating the need to resample any 44.1khz or 48khz stream, or even vastly simplifying the resample of any sample rate that is a multiple of this.
For a sampled signal, if you know the sampling satisfied Nyquist (i.e., there was no frequency content above fs/2) then the original signal can be reproduced exactly at any point in time using sinc interpolation. Unfortunately that theoretically requires an infinite length sample, but the kernel can be bounded based on accuracy requirements or other limiting factors (such as the noise which was mentioned). Other interpolation techniques should be viewed as approximations to sinc.
Sinc interpolation is available on most oscilloscopes and is useful when the sample rate is sufficient but not greatly higher than the signal of interest.
If it's really good AI upsampling, you might get qualitatively "better" sounding audio than the original but still technically deviates from the original baseline by ~8%. Conversely, there'll be technically "correct" upsampling results with higher overall alignment with the original that can sound awful. There's still a lot to audio processing that's more art than science.
This sounds contradictory - what would be the precision that can be heard in a test then?
Makes me think of GPS where the signal is below the noise floor. Which still blows my mind, real RF black magic.