Dithering is a Mathematical Process - NOT a psychoacoustic process.

[Hayabusa]

What rules? Curious... My system shaped the noise to move it out of band and control the amplitude. What sort of psychoacoustic rules are used to optimize dither in audio systems? Do they correlate to the math? Noise decorrelation is pretty well defined and has been for decades; I don't know anything (relatively) about psychoacoustics.

from https://en.wikipedia.org/wiki/Noise_shaping

Noise shaping is a technique typically used in digital audio, image, and video processing, usually in combination with dithering, as part of the process of quantization or bit-depth reduction of a digital signal. Its purpose is to increase the apparent signal-to-noise ratio of the resultant signal. It does this by altering the spectral shape of the error that is introduced by dithering and quantization; such that the noise power is at a lower level in frequency bands at which noise is considered to be less desirable and at a correspondingly higher level in bands where it is considered to be more desirable. A popular noise shaping algorithm used in image processing is known as ‘Floyd Steinberg dithering’; and many noise shaping algorithms used in audio processing are based on an ‘Absolute threshold of hearing’ model.

 

[solderdude]

I suppose Radsone's marketing of their "advanced dithering" is just BS?
They're quite literally advertising two "Radsone-tuned" dithering modes.

Paul can claim 'pkane-tuned' dithering mode

 

[mansr]

What rules? Curious... My system shaped the noise to move it out of band and control the amplitude. What sort of psychoacoustic rules are used to optimize dither in audio systems? Do they correlate to the math? Noise decorrelation is pretty well defined and has been for decades; I don't know anything (relatively) about psychoacoustics.

The audibility of dither noise below 20 kHz can be minimised by shaping it along the sensitivity curves of the ear. This is most useful when using 44.1/48 kHz sample rate since there is no ultrasonic region where noise can be dumped.

 

[DonH56]

from https://en.wikipedia.org/wiki/Noise_shaping

Noise shaping is a technique typically used in digital audio, image, and video processing, usually in combination with dithering, as part of the process of quantization or bit-depth reduction of a digital signal. Its purpose is to increase the apparent signal-to-noise ratio of the resultant signal. It does this by altering the spectral shape of the error that is introduced by dithering and quantization; such that the noise power is at a lower level in frequency bands at which noise is considered to be less desirable and at a correspondingly higher level in bands where it is considered to be more desirable. A popular noise shaping algorithm used in image processing is known as ‘Floyd Steinberg dithering’; and many noise shaping algorithms used in audio processing are based on an ‘Absolute threshold of hearing’ model.

Yes, nice link, I read that earlier, since I have not considered dither for some time. Part of my problem is that I was introduced to noise decorrelation for radar systems and other RF communications (voice/data links, EW, etc.) and not audio. I was made aware of it in audio around the same time but paid little attention to it until some years later. My applications were not audio so psychoacoustics played no part whatsoever in how I researched and utilized dither in my designs. I actually worked with it later at a company that produced high-speed imagers as there were some research efforts to see if adding dither inside the imager or in the processor after was better for performance and SWAP (size, weight, and power).

At any rate, clearly I am not understanding the debate, and will quit derailing the thread! - Don

 

[bennetng]

With 44/48k sample rate there is not much room to shape the noise with only 1 or 2 (for symmetry) LSBs. For example Audacity uses more than 2 LSBs for the "shaped" dither.

 

[DonH56]

The audibility of dither noise below 20 kHz can be minimised by shaping it along the sensitivity curves of the ear. This is most useful when using 44.1/48 kHz sample rate since there is no ultrasonic region where noise can be dumped.

Thank you. A scheme I used in the primordial past had the same problem (needed to apply in-band decorrelation with minimal damage to everything else) and was solved by a fairly complicated circuit (designed by myself and a really good RF engineer I was working with at the time). Should have patented it but the company had a "don't let the secrets out" mentality and the end program was black. The solution was specific to I-Q systems, however, so not applicable to audio.

 

[j_j]

Sigh. This takes me back to my days at Nicolet...

/very old man

I had one of your audio spectrum analyzers connected to an AA501 for about two decades. That was a very useful setup, let me tell you.

 

[j_j]

As far as I know, there are different dithering algorithms that generate different noise shaping responses, and the developers choose them by psychoacoustics tests.

There are many options, including shaped dither, and dither including noise shaping (where you shape the quantization noise spectrum as well).

I used the most basic, simple method going for a simple reason, it shows conclusively that without dither, INFORMATION IS LOST.

 

[j_j]

The audibility of dither noise below 20 kHz can be minimised by shaping it along the sensitivity curves of the ear. This is most useful when using 44.1/48 kHz sample rate since there is no ultrasonic region where noise can be dumped.

Indeed, if you use 96K sampling rate and 25kHz bandwidth, you can dump a veritable (*&(*&*(load of noise above 25K. No speculation on my part, that

 

[j_j]

Yes, nice link, I read that earlier, since I have not considered dither for some time. Part of my problem is that I was introduced to noise decorrelation for radar systems and other RF communications (voice/data links, EW, etc.) and not audio.

If you were doing broadband signals, you're probably best off using white dither, of course. What you want to do is improve the SNR of the part of the spectrum you need. But your basic point holds, psychoacoustics IS NOT the reason for dither. Dither preserves information.

 

[j_j]

To help with the squint factor, here is a 4k version of the first plot.

 

[danadam]

As far as I know, there are different dithering algorithms that generate different noise shaping responses, and the developers choose them by psychoacoustics tests.

Here is from Benchmark:
https://benchmarkmedia.com/blogs/application_notes/12142585-word-length-reduction-of-digital-audio

Here is SoX

 

[mansr]

Indeed, if you use 96K sampling rate and 25kHz bandwidth, you can dump a veritable (*&(*&*(load of noise above 25K. No speculation on my part, that

Like this:

Flat TPDF dither (blue) for reference.

 

[j_j]

Like this:
View attachment 48002

Flat TPDF dither (blue) for reference.

Let me guess - LPC fit and delta-sigma approach?

 

[mansr]

Let me guess - LPC fit and delta-sigma approach?

32nd order FIR filter plugged into the SoX dither module.

 

[j_j]

32nd order FIR filter plugged into the SoX dither module.

That'll do it!

I was thinking closed-loop noise shaping, but no harm

 

[xr100]

Here's a swept sine at -144dB, dithered and noise-shaped to 16-bits:

I've shown this sort of thing and created demo files before, to "convince" digital "non-believers"... but it doesn't work on them!

And don't get them started on sampling rate ("timing error... missing between the samples..." )

 

[j_j]

And don't get them started on sampling rate ("timing error... missing between the samples..." )

Yeah, that nonsense is still around, and still widely promoted, including in places where people should have known better.

 

[xr100]

from https://en.wikipedia.org/wiki/Noise_shaping

What a mess of an article! :-(

However, to quote from it--the key basic point in understanding noise shaping:

"Noise shaping works by putting the quantization error in a feedback loop."

IOW it's not a question of simply "EQ'ing" dither and then adding an already spectrally modified dither to the signal; and noise shaping can be used without dither.

 

[j_j]

What a mess of an article! :-(

However, to quote from it--the key basic point in understanding noise shaping:

"Noise shaping works by putting the quantization error in a feedback loop."

IOW it's not a question of simply "EQ'ing" dither and then adding an already spectrally modified dither to the signal; and noise shaping can be used without dither.

Indeed. No idea which SOX does, of course.