Some commonsense

[Jud]

1 hour ago, mansr said:

PCM is lossless in the sense that it captures perfectly any signal within the well-defined bounds determined by sample rate and bit depth. This is in stark contrast to perceptual coding systems that discard a little bit here and a little bit there according to their psychoacoustic models of what is or isn't audible.

 

Since there are vanishingly few ADCs/workstations that don't decimate a sigma-delta modulated form to some (other, if you consider the original SDM form to be PCM) form of PCM, and this is not mathematically lossless, we are left in nearly all instances with a non-perfect capture, mathematically speaking.  Not saying anything about the impact on human perception of the signal, as I imagine with good filtering there's very little or no perceptual effect.

[Jud]

9 hours ago, Rt66indierock said:

Paul, for there to be high resolution albums to purchase a high-resolution format must be specified in the recording contract.

 

A possibly significant exception is analog masters. Half of all music streamed last year was "back catalog," so the companies know there's a market.

[Jud]

17 hours ago, mansr said:

The term "lossless" has a widely accepted meaning.

 

Yes, there are any arbitrary number of mathematically lossless conversions that can be made between AIFF, ALAC, FLAC, and WAV at a given sample rate, for instance. Conversion between sample rates isn't a mathematically lossless operation, though as you've pointed out it can be done so as to be "perceptually lossless."

[Jud]

16 hours ago, mansr said:

I took a 48 kHz music file and resampled it to 96 kHz and back 16 times using SoX at the highest quality setting. The spectrogram of the resulting difference looks like this:

 

As expected, there is slight change in the transition band of the filter. Below this, the level of the difference is about -150 dB RMS, -140 dB peak. In other words, the "loss" is confined to a few LSBs for all content below a certain frequency. Using higher precision arithmetic would reduce this further.

 

Music and difference files attached.

file1.flac 4.76 MB · 0 downloads file2.flac 4.75 MB · 0 downloads diff.flac 923.96 kB · 0 downloads

 

Would be interesting to see one of the Native DSD files recorded live to DSD256 compared to the same file decimated to Redbook resolution with a filtering chain that might typically be used for mass market product.

[Jud]

2 minutes ago, mansr said:

No, that is not what I'm saying. Sample rate conversion can be done with arbitrarily high precision. Pick any non-zero number, and the error can be made smaller. The band limiting filter merely needs to be made sufficiently close to the ideal sinc filter. It might require 1024-bit precision and 10 billion filter taps, but it can be done.

 

In practice, the SRC precision need only be high enough to fully utilise the resolution of the target format, typically 24 bits. Taking into account realistic DAC noise levels, the requirements can be relaxed further still.

 

Except we were talking about a definition of "lossless." I would guess you could make a damn fine mp3 as well.

[Jud]

Just now, mansr said:

Compared how?

 

As you did in your prior comment, comparing the initial file to the sample rate converted one. I'd like to see something quite typical of the usual filtering chain used. Since we're talking about common sense, I'd like to get an idea of what's common.

[Jud]

3 hours ago, mansr said:

Getting back to the specific term "lossless," it is not normally applied to raw data formats such as PCM. Rather, the lossless/lossy distinction is used when characterising compression algorithms. FLAC is lossless, whereas MP3 is not.

 

Yes, exactly.

[Jud]

4 hours ago, mansr said:

How would you carry out this comparison. You can't, for a multitude of reasons, do a sample by sample subtraction between a DSD file and a downsampled version of the same.

 

I'm not sure what's common, but it sure ain't sense.

 

How odd. All I'm asking is for you to perform an operation similar to the one you posted about in comment #65 in this thread, and I get a lot of handwaving about how impossible it is.

 

 

[Jud]

49 minutes ago, mansr said:

Let me try again. In the test you're referencing, I started with a 48 kHz, 24-bit file, resampled it to 96 kHz and back to 48 kHz a number of times, the end result being a 48 kHz 24-bit file, same as the original. The difference is then calculated by simply subtracting corresponding samples in the original and the processed files.

 

You are asking me to start with a DSD256 file and downsample it to some normal PCM rate. Then what? The sample rates differ, so subtracting samples is not possible. Convert the downsampled file back to DSD? Then we'd be comparing sigma-delta modulators more than anything else, and besides, DSD being a 1-bit format means arithmetic (such as subtracting one sample from another) isn't possible.

 

Yes, downsample to 16/44.1, using filtering that would be typical in the industry for mass market music.  Then straight to the spectrogram comparison, as I imagine all the bouncing back and forth was to prove a point that I didn't need proved the first time, much less again.  What I'd like to get is some idea of how much if at all the spectrogram of a typical Redbook product differs from the SDM'd signal that was the initial stage in the digital recording process.

[Jud]

2 hours ago, mansr said:

I don't have the (expensive) DAW software likely to be used by commercial studios. Would you accept that SoX resampling is probably of similar quality?

 

It depends on the parameters used. Can you adjust the parameters to result in something like the "half-band filters" sometimes mentioned on the forum?

 

2 hours ago, mansr said:

 

The purpose of the back and forth bouncing was to show that the resampling, even after many iterations, causes only a barely noticeable change in the passband. What makes you think it would be any different with a DSD source?

 

Only the quality of the filtering used - see above.

[Jud]

1 hour ago, PeterSt said:

Maybe the idea that DSD would be Walhalla ?

(not my idea of it)

 

No, not nearly. Only that this is somewhat similar to the initial SDM form of the signal in the workstation, ADC, whatever, before decimation to the final product.

[Jud]

2 hours ago, mansr said:

Sure, I can do that.

 

Most studios probably run the ADCs at 48/24 or 96/24, using the chip's internal decimation filter. Are you looking for an illustration of how these perform compared to high-quality software algorithms? 

 

Yes. I'd like to know what the level of measurable difference, if any, would be for a typical recording chain between the original signal and a Redbook result (Redbook because I'm not interested at the moment in mp3 or other lossy compressed formats).

[Jud]

1 hour ago, mansr said:

What do you mean by original signal? Typical recordings are multi-tracked at 44.1/48 kHz with a substantial amount of processing and mixing before a CD release eventually pops out. Comparing any individual microphone feed to the final CD wouldn't make any sense. It's not supposed to be the same.

 

Original SDM signal in the ADC/workstation immediately upon initial digitization before decimation. My notion is the live DSD256 recording would serve as a very rough approximation.

[Jud]

1 hour ago, mansr said:

Most ADCs use a multi-bit sigma-delta A/D converter. The data it produces is then digitally processed on-chip to produce PCM or DSD output. The 1-bit modulator providing the DSD output may well cause at least as much "damage" as the decimation filters used for PCM.

 

Certainly a possibility, though one would hope the folks doing the recording for the labels sold on NativeDSD would be sufficiently finicky to use a reasonable modulator.

 

If you have a closer proxy to suggest from which you can derive a spectrogram (potentially playable on a consumer DAC - I have some thoughts about other things I might like to do to compare the files), of course I'd be interested.

[Jud]

19 minutes ago, mansr said:

The modulator is whatever the chip designers at TI, AKM, or wherever concocted.

 

NativeDSD provides recording equipment information fairly often, so it might be possible to determine.

 

19 minutes ago, mansr said:

 

I'm still struggling to understand what exactly you're looking to achieve.

 

You can stop struggling. 🙂 I'm not looking to "achieve" anything, just to try to get some sense of what types of changes - vanishingly low or otherwise, doesn't matter - result from decimation in the typical recording chain.

[Jud]

19 minutes ago, jabbr said:

if I was recorded and mastered at a certain nitrate,

 

Mmm, nitrates - hot dogs and bacon!

[Jud]

45 minutes ago, mansr said:

Well, obviously any frequencies higher than half the new sample rate are lost. This should not be news to you. Besides that, imperfections in the decimation filters will have a few minor effects:

• Very slight passband ripple.
• Some drop-off a little before fs/2.
• Aliasing of residual high-frequency content due to finite stopband attenuation.

The details can be found in the ADC data sheet.

 

That's fine. Up for performing the exercise if a live-to-DSD256 file is supplied?

[Jud]

9 hours ago, mansr said:

You have yet to specify "the exercise."

 

Well, I've given it a try several times.  Perhaps I was bad at explaining, perhaps you were less than enthused about understanding, perhaps neither.  In any case, I'll drop it.

[Jud]

8 hours ago, mansr said:

I understand you want to compare something between a DSD256 file and a downsampled version of it. You also mentioned spectrograms.

 

Starting with a DSD256 file from a NativeDSD sampler (track 5 from this album) I happened to have handy, here's what I've done:

1. Downsample to 44.1 kHz.
2. Upsample to DSD256 rate, keeping 24-bit sample precision.
3. Subtract the original DSD samples (as ±1 values).
4. Plot spectrogram of difference up to 44.1 kHz.

 

Above 21 kHz, we see a little high-frequency content from the recording that was lost in the 44.1 kHz conversion. Below that, nothing. The level of the difference in the 0-20 kHz range is -170 dB RMS, -155 dB peak. This is too small to be representable in 24-bit precision.

 

Is this anything like what you were looking for?

 

That's fine. Thank you.

[Jud]

2 hours ago, tmtomh said:

 

That's quite eye-opening.

 

The devil is always in the details.

[Jud]

22 hours ago, esldude said:

So peak levels in that are maybe -80 db FS in the 20-25 khz range.  So if your 0 dbFS level plays back at 120 db SPL, these ultrasonics would be 40 dbSPL.  Considering hearing around 20 khz if you have it has a threshold of about 100 dbSPL, I'm going out on a limb here and saying that isn't going to be heard.  This is before we consider masking by louder lower frequencies too.  

 

Is my impression correct that the differences there were less than with the 2L samples?