Skip to main content
Topic: Naim sample (Read 4046 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

Naim sample

Reply #1
Let me say some words about how this file was done.
I did downsample the original 192kHz file with SoX parameters "-b 92 -a -v 44100 dither -a -f low-shibata"
We have some tiny amount of aliasing with that but SoX luckily only adds this aliasing above the passband. Starting the lowpass at ~20kHz (-b 92) we have only very low amplitude left at 22kHz. The amount of alias is really minimal then. Much of it drowns in the dithernoise.
You may wonder why we should alias at all but using a not so steep filter leads to less ringing. Allowing aliasing once more reduces it.
The setting above has clearly less ringing showing in impulse responses as others defaults like iZotope for example.

Since math can't be fooled pruducing even less ringing with linear phase can only be archieved with a lowpass kicking in earlier (below 20kHz) so butchering frequency response or allowing more aliasing, no matter what software you use.

Now the upsampling back to 192kHz for a fair comparison ruling out hardware.
Normaly as learned above people think steep filter settings should be avoided because of ringing. Here it doesn't matter anymore. We already filtered out pretty much any content near 22kHz so a steep filter simply has no content left to let it ring. I used a SoX -b99 setting therefore. We can also prevent problems from using 2 times a similar filter.
You can create a delta file from the original naim sample if you like. It will only show how good it works.
One word to the choice of dither. I like the work of Naoki Shibata and did some dithernoise shootout for my own pleasure really liking the Shibata approach. Since i don't know enough how strong shaped dithernoise with strong HF energy can interact with hardware i use the less shaped version.
Is troll-adiposity coming from feederism?
With 24bit music you can listen to silence much louder!

Naim sample

Reply #2
Let me say some words about how this file was done.
I did downsample the original 192kHz file with SoX parameters "-b 92 -a -v 44100 dither -a -f low-shibata"
We have some tiny amount of aliasing with that but SoX luckily only adds this aliasing above the passband.


Response in the stopband is > 100 dB down from the mean passband in an analysis with 96 KHz bandwith. Suffice it to say, no way is this audible or troublesome.


Quote
Starting the lowpass at ~20kHz (-b 92) we have only very low amplitude left at 22kHz. The amount of alias is really minimal then. Much of it drowns in the dithernoise.


Here's a plot of the frequencies of interest:

[attachment=8241:sox_resa...r_wombat.png]

Quote
Since math can't be fooled pruducing even less ringing with linear phase can only be archieved with a lowpass kicking in earlier (below 20kHz) so butchering frequency response or allowing more aliasing, no matter what software you use.


The wording above "Butchering" is IMO unfortunately emotionally charged.  It presumes accurate knowledge of issues for which AFAIK the facts are not actually known.

Enlargement of above;

[attachment=8242:sox_resa...enlarged.gif]

We can see that response @ 20 KHz is down only a tiny fraction of a dB.  This frequency this is the extent of what is considered to be normal hearing. so the threshold of detection of changes can safely presumed to be a dozen dB or more. The modern world is full of perceptual coders that brick wall the signals they process as low as 16 KHz. By brick wall, it is observed that losses > 60+ dB @16-18 KHz are thought by many skillful and insightful coder developers to be entirely acceptable for high sound quality.

Naim sample

Reply #3
We can see that response @ 20 KHz is down only a tiny fraction of a dB.  This frequency this is the extent of what is considered to be normal hearing. so the threshold of detection of changes can safely presumed to be a dozen dB or more. The modern world is full of perceptual coders that brick wall the signals they process as low as 16 KHz. By brick wall, it is observed that losses > 60+ dB @16-18 KHz are thought by many skillful and insightful coder developers to be entirely acceptable for high sound quality.

I don't know where you want to go. Debating for cd format and 44.1kHz a linear frequency response up to 16khz is enough? I like the idea of keeping 20kHz perfectly.
Is troll-adiposity coming from feederism?
With 24bit music you can listen to silence much louder!

Naim sample

Reply #4
We can see that response @ 20 KHz is down only a tiny fraction of a dB.  This frequency this is the extent of what is considered to be normal hearing. so the threshold of detection of changes can safely presumed to be a dozen dB or more. The modern world is full of perceptual coders that brick wall the signals they process as low as 16 KHz. By brick wall, it is observed that losses > 60+ dB @16-18 KHz are thought by many skillful and insightful coder developers to be entirely acceptable for high sound quality.

I don't know where you want to go. Debating for cd format and 44.1kHz a linear frequency response up to 16khz is enough? I like the idea of keeping 20kHz perfectly.


I didn't think that Science (i.e. what works) was up for vote in a popularity contest or personal likes or dislikes.

According to the evidence I've provided your own Sox parameters "Corrupt" frequency response at 20 KHz with a loss about a tenth of a dB or maybe 3 tenths. How can you square that with your desire for "...keeping 20 kHz perfectly"?  ;-)

I think a good starting point might be keeping response within a tenth of a dB up to 10 KHz, allow maybe 1 dB down at 12 kHz, 3 dB down at 16 kHz, 10-15  dB down at 20 kHz, and pretty much brick wall after that. An approximate 2 KHz transition band.

But that needs to be tested, no?

Naim sample

Reply #5
According to the evidence I've provided your own Sox parameters "Corrupt" frequency response at 20 KHz with a loss about a tenth of a dB or maybe 3 tenths. How can you square that with your desire for "...keeping 20 kHz perfectly"?  ;-)
But that needs to be tested, no?

-b 92 m is 20286Hz for SoX since the filter kicks in earlier you can use 93% or 94% but a tenth of a dB, come on!
Until lately here we got  recommendations about using SRC with 99% of BW and alike. Now that some AES paper showed some small possibility that a filter with 500Hz wide so  ~97% is barely audible under still debatable conditions everything is new?

I use this setting since years and i bet if some listening tests offered on the net were made with this simple SoX settings their outcome could have been more valid. There were experts doing badly resampled stuff. Just create a delta file to check, even loudness is kept.

It will be hard doing real abx testing of these things. At least this naim sample was not abxble against the original 192khz file from the person who started the question about it.
Is troll-adiposity coming from feederism?
With 24bit music you can listen to silence much louder!

 
SimplePortal 1.0.0 RC1 © 2008-2019