I admit I do not know a lot about upsamling or PCM to DSD conversion, but I also made some research and many say that one cannot create information that was not there.
When I heard that from an expert from an audio shop I was suddenly thinking about computer graphics. In computer graphics it is common knowledge that one can create additional information that visually improves the picture. Nobody questions upsamling or anti aliasing there as the effects are easy to see with our eyes. With music itâs more difficult to see the difference but is improving music not literally the same than improving computer graphics? I think there are similarities and therefore for me itâs proven that upsampling and pcm to dsd conversion can improve the sound. All that remains is the question if the algorithms used are already as sophisticated as in computer graphicsâŚ
mjw
(Here I am with a brain the size of a planet and they ask me to pick up a piece of paper. Call that job satisfaction? I don't.)
2
I think terminolgy is really important here. Whether the source is a CD (44.1 kHz 16 bit PCM) or DSD, neither are lossy. All the information needed to recreate the original recording is present. So, nothing lost needs to be recreated.
Moreover, for any given samples, only one waveform will pass through them all. Upsampling uses mathematics to interpolate values between two known data points.
Returning to your image analogy, the reality is that there are no pixels, but data points representing a particular hue. Therefore it is possible to use interpolation to add addition data points that may improve image quality of lossy formats such as JPEG and MPEG-4.
However, the crucial point with audio is that you can recreate the original waveform with the samples since these formats are not lossy (except, for example MP3.)
Hi Marc, can you give an example of the additional information that is created in case of pictures? Is it just about interpolating pixels to increase resolution or enhancing contrast or color balance? You can consider that to be additional information, but it comes entirely from the image, so nothing external is added.
Itâs not just that. The latest innovation is DLSS3 that generates additional frames and DLSS2 that renders the image in a lower but faster resolution and uses then AI to upscale the picture to screen resolution. There is so much more innovation done in that area than in improving sound but Iâm SURE that AI imprivenent of sound in real time will be the next big thingâŚ
Best is if you ask that an AI. Iâm not a mathematician and this is very very complexâŚ
Upsampling enables better recovery of said information.
The notion that adding information creates a better sounding result in the way adding frames in video improves the result isnât a good analogy. The reason is it improves a subjective measurement (frame rate) which gamers pursue but the reality is it often doesnât look great. People have reported blurriness or even motion sickness when employed.
And what about AI upscaling? A visual imrovement can be an analogy for a sound quality improvement. Beyond that I think that AI will very well in reconstruction of the original waveformâŚ
As said before in this thread.. nothing is lost in the original (lossless) pcm or dsd sound file, so nothing needs to be âreconstructedâ into an âoriginalâ waveform because the original is already present from the beginning.
IMO your analogy with visual improvements/reconstruction in (pixelated?) images does not completely fly here , because that assumes that the original picture does not have all the information to begin with.
I agree with @Henry_McLeod and @Andy_L. There are visual improvements to be made in video where AI can help. Audio on the other hand works differently. Setting aside transducer imperfections, digital audio can be captured and rendered with an accuracy exceeding our hearing capabilities. And unlike in video, up-sampling audio doesnât improve sound in any way; itâs done for different reasons during D/A conversion.
I would think AI could help with restoring poor audio recordings (e.g. eliminating artifacts like clicks, pops, maybe even small dropouts), but those are not audiophile applications.
The answer to that would be twofold. Listen, but do not consider the result bit perfect. I actually think AI enhancement might be useful in some applications like cleaning up or enhancing poor recordings.
Ok. Now I asked AI for support. FLAC is NOT LOSSLESS. Yes the compression is lossless but the sampling of the wave creates the loss. That is what needs to be reconstructed when making upsampling.
Thatâs probably another hallucination of AI. FLAC is lossless and thereâs nothing it can do about analog wave sampling. Regarding sampling itself, as long as it satisfies the Nyquist condition, itâs also lossless as far as audibility is concerned.
No, the samples are not the original waveform, just a recipe to recreate the original waveform according to a mathematical formula (which in itself is sort of contradictory with itself, so it is imperfect for the real world purpose to begin with).
Then the question is about how to reconstruct the analog waveform the sparsely spaced digital samples are representing the original analog waveform. This is similar to drawing exercises for kids where you have numbered dots, and you need to connect those with lines and arcs to recreate the picture those are representing.
Also note that the ADCâs and the tools used to create those samples are far from perfect. So there are errors embedded to the source data. So even if your digital delivery package is lossless, the analog-to-digital conversion is not lossless. Best reconstruction algorithms also take this into account and attempt to correct at least the worst errors.
As I mentioned above, up-sampling is used during the D/A process (for this particular case, a better term would be oversampling https://en.wikipedia.org/wiki/Oversampling ) in order to be able to reduce the bit depth from 16 or 24 bits to something like 1-6. This makes the conversion to analog much easier, since the number of levels is a lot smaller.
You can think about the A/D conversion as a reversed D/A process: the analog signal is sampled at a much higher frequency than needed (i.e. it is oversampled), quantized with a small number of bits (again like 1-6), down-sampled to 44.1-192kHz and then re-quantized with 16-24 bits ( to get what we call PCM). If you use only one bit for the first quantization and skip the down-sampling, you get DSD. Some people think thatâs better, although from a technical standpoint and considering the limits of human hearing, it can be demonstrated thatâs not the case.
2 Likes
mjw
(Here I am with a brain the size of a planet and they ask me to pick up a piece of paper. Call that job satisfaction? I don't.)
16
Yes, it could be useful for those wanting to create a vinyl rip: removing pops and clicks, and splitting the recording into tracks.
There are sufficient samples to faithfully construct the original waveform.
This is a little misleading, since only one waveform can pass through the âdotsâ. Thereâs no straight line or arcs between the dots.
I think that it depends on your DAC, the music boxes,⌠and your ears whether you hear the difference between DSD and FLAC or not.
Let me explain it with an analogy: If you drive a car on a rural country road in e.g. India then you will most probably not notice when you drive over a 10 cent coin, but if you drive on a perfectly flat road you might notice it.
mjw
(Here I am with a brain the size of a planet and they ask me to pick up a piece of paper. Call that job satisfaction? I don't.)
19
That"s because DSD often uses a different master. Compare different versions of Red Book, and they are often different.
If format changes the sound, itâs not doing its job.
Only if the original waveform didnât include frequencies exceeding half of the sampling rate. Which is most of the time not the case for 44.1k/48k sampling rate. So the source gets bandlimited by analog and digital filters. Which are imperfect and in any case leave their fingerprint in the data.
To faithfully reproduce any input signal, you would need infinite sampling rate and infinite word length. And you would still have the same challenge on what kind of practical implementation can most accurately reconstruct that. As R2R ladders (PCM word length) have their practical and physical limitations, just as any other method.
So these days, none of the ADCâs or DACâs use PCM at the conversion stage, but are instead oversampled delta-sigma converters with some variations. And there things like the conversion stage design details, modulator design details and the digital filter (needed for PCM compatibility) design details make a huge difference.
DSD is much more native for modern converters than PCM. And it doesnât require âbrickwallâ bandwidth limiting which in itself causes various side effects.
Except that most DAC chips perform only oversampling to 352.8/384k rate through largely imperfect digital filters. And rest of the oversampling through sample-and-hold (zero-order-hold) which is just copying same sample N times. This is exactly drawing straight horizontal lines on the waveform.
Now the point is which of the imperfect digital filters you choose?
Personally, I donât do any of that sample-and-hold stuff, I run proper digital filters up to final rate like 50 MHz, before passing the data to the modulator.
, because that assumes that the original picture does not have all the information to begin with.