DISTORT - audio distortion simulation app

[pkane2001]

Introducing DISTORT - free software designed to help determine the audibility thresholds of various audio distortions with your own equipment, at your own pace, with your own music. Some basic instructions and installer download are available on the Distort web page:

https://distortaudio.org

The basic idea is to shape the desired harmonic, noise floor, jitter, and other distortions and then to apply them to any piece of recorded music or a test signal. You can play the distorted file directly, using the play button, or save it as a 32-bit WAV and play it using your preferred player software or analyze using your preferred digital audio analysis app.

My hope is that DISTORT can help answer some burning audio questions and help to address some persistent audio myths:

 

• Is SINAD of 110dB audible?
• Is THD below 1% audible?
• Do even harmonics sound better than odd?
• Do higher-order harmonics sound less pleasant?
• Does SET distortion really sound good?
• At what level does the noise floor or the mains frequency become audible?
• What kind of music is improved by lower THD? Even harmonics? Lower noise floor?
• What does jitter sound like? When does it become audible?
• etc., etc., etc.

 

This thread is for the discussion of features, functions, capabilities, bug reporting, and improvement suggestions related to DISTORT.

[pkane2001]

This question was raised in another thread by @bluesman:

 

So, let's discuss how one would go about confirming what distort does.

 

The simplest way to verify is to test it with various test signals to confirm that it does what's advertised and nothing more. This is a common approach by engineers studying an unknown device. Input a known signal and measure what the device does at the output.

 

Mathematically, internal operation is conceptually simple: DISTORT applies an arbitrary non-linear transfer function to the input signal. This is a mathematical simulation of what happens in the analog domain. Assuming that you select the right function, it will simulate the exact HD and IMD for that device. This is not only easy to confirm, but easy to understand:

 

Output = F(input)

 

Where F() is a non-linear transfer function of your choice. The transformation takes place in the time domain.

 

Once a non-linear function is applied, DISTORT does some FFT analysis on the output to show harmonic distortion, and compute a few metrics, it also shows the actual waveform of the distorted signal so you can verify it applied the function you selected. These are done after the non-linear function has been applied, so there's no cheating here: the distortions shown on FFT are those generated by the non-linear function and nothing else.  If there's second harmonic in the output, it is because the non-linear transfer function caused it to be created. If there's IMD between multiple tones, it's because of the non-linear xfer function.

 

If you don't trust DISTORT FFT analysis, then write the resulting waveform into a file (DISTORT will save it for you), and load it into any other analysis tool, from REW to APx500 to see that the desired distortions have been applied. This is what engineers do when they measure an unknown device.

 

[opus101]

Is DISTORT modelling a purely static non-linearity? If the answer to this is 'yes' then what real-world electronics produces distortion according to a static non-linearity? I'm curious whether the model used corresponds to real-world devices here. There was much talk on the other thread about 'validation' but that seemed to me to be misguided - the important aspect for me is validation not of the software itself but rather of the model being used. All simulation software has this issue - how close are the models being used to the reality?

[sandyk]

50 minutes ago, opus101 said:

Is DISTORT modelling a purely static non-linearity? If the answer to this is 'yes' then what real-world electronics produces distortion according to a static non-linearity? I'm curious whether the model used corresponds to real-world devices here. There was much talk on the other thread about 'validation' but that seemed to me to be misguided - the important aspect for me is validation not of the software itself but rather of the model being used. All simulation software has this issue - how close are the models being used to the reality?

I would also suggest that unless Simulation Software confirms the results of carefully controlled DBT sessions that it is useless for this purpose in it's present form. Measurements are normally only able to act as guidelines to the expected Audible (or Visual) performance of a device and should never be taken as 100% definitive on their own, as many (most except Paul?) ASR members frequently do .

I am reminded here of the performance of older comparison S/W Audio DiffMaker that rarely nulled to 100%, yet results were taken by many Objectivists , including several that used to post in this forum ,as absolute proof.

Barry Diament also rejected DiffMaker as not being reliable due to this. 

 Not so long ago, we accepted differences of around 1dB as being inaudible, yet these days , even Archimago now acknowledges that some people are able to hear differences as small as 0.1dB, and not necessarily at a specific frequency only. John Dyson's little PM group has also confirmed this on quite a few occasions, and Paul should also be well aware of this, as he is a member of John's small PM group of members.

[Audiophile Neuroscience]

4 hours ago, opus101 said:

There was much talk on the other thread about 'validation' but that seemed to me to be misguided - the important aspect for me is validation not of the software itself but rather of the model being used. All simulation software has this issue - how close are the models being used to the reality?

 

I can only suggest that this is part and parcel of what I would consider validation. If the software uses an invalid model (not saying it does) then the software is not valid.

 

These are all things I would expect independent validation tests would consider.

[opus101]

If it can't model typical circuit behaviour then I'd say it needs an updated model.

 

I dimly recall having been here before with a paper by Geddes justifying his distortion metric. It used a static transfer function to model distortion, noting that not all systems can be so modelled.

[Audiophile Neuroscience]

13 hours ago, pkane2001 said:

Where F() is a non-linear transfer function of your choice.

 

I do NOT profess to be an expert on this subject. My limited understanding however is that Transfer function approach typically is applied to linear time invariant systems. The input signal can be written as an integral "of scaled and shifted delta impulses, and due to linearity and time invariance, the response to this integral/sum equals the integral/sum of scaled and shifted impulse responses"

 

"For non-linear systems you can compute or measure the system's response to an impulse, but this function does not tell you anything about the system's response to other input signals, or even to a scaled impulse. This is why the impulse response and its transform do not have any significance for non-linear systems."

[Audiophile Neuroscience]

8 minutes ago, opus101 said:

If it can't model typical circuit behaviour then I'd say it needs an updated model.

 

I dimly recall having been here before with a paper by Geddes justifying his distortion metric. It used a static transfer function to model distortion, noting that not all systems can be so modelled.

 

 So, one of my naive questions would relate to how accurately the Transfer function does what it purports to do ie represent, model or simulate something. Can you reliably expect to apply the input distortion metric (noise floor, THD, whatever) to complex music and know that the output faithfully represents what a real world component would do to the music?

[Audiophile Neuroscience]

4 hours ago, sandyk said:

I would also suggest that unless Simulation Software confirms the results of carefully controlled DBT sessions that it is useless for this purpose in it's present form. Measurements are normally only able to act as guidelines to the expected Audible (or Visual) performance of a device and should never be taken as 100% definitive on their own, as many (most except Paul?) ASR members frequently do .

I am reminded here of the performance of older comparison S/W Audio DiffMaker that rarely nulled to 100%, yet results were taken by many Objectivists , including several that used to post in this forum ,as absolute proof.

Barry Diament also rejected DiffMaker as not being reliable due to this. 

 Not so long ago, we accepted differences of around 1dB as being inaudible, yet these days , even Archimago now acknowledges that some people are able to hear differences as small as 0.1dB, and not necessarily at a specific frequency only. John Dyson's little PM group has also confirmed this on quite a few occasions, and Paul should also be well aware of this, as he is a member of John's small PM group of members.

 

You are saying that the simulated transformed output signal+distortion needs to be compared with a real world sample of signal+distortion by doing blind listening tests, yes?

[opus101]

The issue (assuming that it does indeed use a static non-linearity model which @pkane2001 has yet to confirm) is that real-world circuits have frequency dependent distortion. This kind of distortion can't be modelled by arbitrary polynomial transfer functions which give frequency-independent results.

 

One route by which frequency dependent distortion enters the picture is due to the use of feedback. Feedback circuits require compensation (frequency response tailoring) in order to have acceptable stability. Therefore the amount of feedback-wrought correction of any static non-linearity is decreasing with frequency because the 'excess gain' available decreases with frequency. This is why opamps in general show rising THD with increasing frequency.

[sandyk]

43 minutes ago, Audiophile Neuroscience said:

All simulation software has this issue - how close are the models being used to the reality?

+1

[sandyk]

32 minutes ago, opus101 said:

Therefore the amount of feedback-wrought correction of any static non-linearity is decreasing with frequency because the 'excess gain' available decreases with frequency. This is why opamps in general show rising THD with increasing frequency.

Surely it also depends on the overall Bandwidth being greatly in excess of the maximum frequency to be corrected, otherwise smearing of HF detail is likely ? This can be more readily seen when NFB is used with Analogue TV  (especially Sync pulse area).This is also likely to be part of the reason why amplifiers with 30MHZ (or 50MHZ) Ft output devices  generally sound better than  the old <4mHz output devices ? 

[opus101]

2 minutes ago, sandyk said:

This is also likely to be part of the reason why

amplifiers with >30MHZ Ft output devices  generally sound better than  the old <4mHz output devices ? 

 

Wider bandwidth devices are going to allow more loop gain therefore better PSRR. Also they have smaller parasitic capacitances (Ccb for example) which leads to better PSRR too.

[pkane2001]

1 hour ago, opus101 said:

The issue (assuming that it does indeed use a static non-linearity model which @pkane2001 has yet to confirm) is that real-world circuits have frequency dependent distortion. This kind of distortion can't be modelled by arbitrary polynomial transfer functions which give frequency-independent results.

 

One route by which frequency dependent distortion enters the picture is due to the use of feedback. Feedback circuits require compensation (frequency response tailoring) in order to have acceptable stability. Therefore the amount of feedback-wrought correction of any static non-linearity is decreasing with frequency because the 'excess gain' available decreases with frequency. This is why opamps in general show rising THD with increasing frequency.

Right. There is a frequency dependent setting for feedback that gives you that effect also. Simply simulates a negative or positive feedback with a frequency filter.

[pkane2001]

2 hours ago, Audiophile Neuroscience said:

 

 So, one of my naive questions would relate to how accurately the Transfer function does what it purports to do ie represent, model or simulate something. Can you reliably expect to apply the input distortion metric (noise floor, THD, whatever) to complex music and know that the output faithfully represents what a real world component would do to the music?

The function, as I said, is configurable. You can specify the desired distortion, and then measure its effect to confirm that it matches the real world. As an example, I included a number of transfer functions from published models by researchers and engineers, such as triode preamp or guitar amp. Since DISTORT doesn’t dictate the model you should use, but gives you the ability to configure one, the onus is on you to try to duplicate any specific device behavior.
 

My goal was to faithfully (read measurably equivalently) reproduce the effect of common distortions on any musical or test signal.

 

[pkane2001]

6 hours ago, sandyk said:

I would also suggest that unless Simulation Software confirms the results of carefully controlled DBT sessions that it is useless for this purpose in it's present form. Measurements are normally only able to act as guidelines to the expected Audible (or Visual) performance of a device and should never be taken as 100% definitive on their own, as many (most except Paul?) ASR members frequently do .

I am reminded here of the performance of older comparison S/W Audio DiffMaker that rarely nulled to 100%, yet results were taken by many Objectivists , including several that used to post in this forum ,as absolute proof.

Barry Diament also rejected DiffMaker as not being reliable due to this. 

 Not so long ago, we accepted differences of around 1dB as being inaudible, yet these days , even Archimago now acknowledges that some people are able to hear differences as small as 0.1dB, and not necessarily at a specific frequency only. John Dyson's little PM group has also confirmed this on quite a few occasions, and Paul should also be well aware of this, as he is a member of John's small PM group of members.

DBT is useful, but not required. Measurements confirm what DISTORT does or doesn’t do. There is no claim in DISTORT regarding audibility— these claims will be made by DISTORT users, if they so chose.

[pkane2001]

2 hours ago, Audiophile Neuroscience said:

 

I can only suggest that this is part and parcel of what I would consider validation. If the software uses an invalid model (not saying it does) then the software is not valid.

 

These are all things I would expect independent validation tests would consider.

 

Can you list, specifically, what it is you'd like to validate? A model can be validated by measuring its effect on the output. This is how engineers determine what a device does, this is the easiest black-box test that you could possibly conduct. This is how DISTORT has been validated.

 

I could give you the equations for, for example, the construction of harmonic distortion, but how would you validate that these equations are correct? Do you know what the "correct" non-linear distortion model looks like to be able to verify it?

[SoundAndMotion]

Whenever differing  definitions are used, miscommunication is risked. We saw that with "intermodulation" in the other thread.

 

I have used "transfer function" (TF) for the past 40 years to mean frequency domain TFs. One usually uses the complex variable "s", switches between freq.- and time-domain with the Laplace transform, and a linear, time-invariant, causal system is assumed/required. 

More recently I have seen a time-domain use of TF, especially among amplifier engineers. This is the way @pkane2001 is using it. It is frequency independent, and transforms every time domain input value, x, into output y=f(x), as Paul has mentioned. The special case of f(x)=x is an infinite bandwidth, gain=1, linear response. Any and all f(x)'s that are not pure gains, i.e. straight lines, are non-linear (gosh!) and produce HD and IMD.

 

Much confusion can be avoided if it is stated whether time-domain or frequency-domain is intended. 

[pkane2001]

2 hours ago, sandyk said:

Surely it also depends on the overall Bandwidth being greatly in excess of the maximum frequency to be corrected, otherwise smearing of HF detail is likely ? This can be more readily seen when NFB is used with Analogue TV  (especially Sync pulse area).This is also likely to be part of the reason why amplifiers with 30MHZ (or 50MHZ) Ft output devices  generally sound better than  the old <4mHz output devices ? 

 

Alex, please keep your posts on the original topic. 

[sandyk]

1 hour ago, pkane2001 said:

 

Alex, please keep your posts on the original topic. 

Obviously you didn't notice that I was replying to Richard, whose reply you didn't mark as off topic !

[opus101]

5 minutes ago, sandyk said:

Obviously you didn't notice that I was replying to Richard, whose reply you didn't mark as off topic !

 

Correction - your response wasn't in any way a reply to me and as @pkane2001 suggests it dragged the discussion off topic.

[pkane2001]

Just now, sandyk said:

Obviously you didn't notice that I was replying to Richard, whose reply you didn't mark as off topic !

 

Your response was unrelated to DISTORT simulation, his was right on point. In fact, while testing DISTORT, an ex-opamp designer raised the same question about feedback on ASR. This is why DISTORT now has band-limited feedback simulation logic in it. Turned out it was really easy to do in a simulation.

[SoundAndMotion]

Hi Paul,

 

You and I both create software to answer questions, but I'm impressed that you go the extra mile to try to make it useful to others besides yourself. DISTORT is an example. Nice!

 

A handful of random stuff:

Why ALL CAPS? Is DISTORT an acronym or just intended to annoy old people. Former?:What's it mean? Latter?:It works!

 

When I see graphs, I have an uncontrollable urge to jump in... zoom, pan, change axis limits, zoom out, find the value at a given point, either by click or zoom-zoom-zoom. I want to set up camp, build a bonfire and roast some data... If you think I'm fit to be tied, I'm fine with that. Ignore me. Otherwise, graph tools would be nice.

Does the scale for the sliders on the right mean anything? Does 1000 mean 100%?

@opus101 and others mention the nature of your nonlinearity. opus asked:

Quote

"Is DISTORT modelling a purely static non-linearity? If the answer to this is 'yes' then what real-world electronics produces distortion according to a static non-linearity? I'm curious whether the model used corresponds to real-world devices here."

I'm going to assume that for many situations, I can assume that the bandwidth of my signal is well within the flat amplitude, zero-phase region of the electronics, and that other non-static things like temperature have stabilized. Then, yes, I would want a purely static non-linearity, and I think all analog electronics would exhibit this behavior (to some extent: selectable in DISTORT). But then I want to ask if your non-linear transfer function is unique for a given set of harmonics: the answer is no. You need the phase of each harmonic. I doubt phase differences of a few degrees matter, but what about 180°? And I'm going to go out on a limb here, but I believe if you have the magnitude and phase of each harmonic, then the transfer function is unique, and therefore the model has no degrees of freedom for error (with the stated assumptions). Is this true? It "feels" true, but I have to work through it a bit more...

 

So, would entering mag and phase of harmonics be interesting in DISTORT?

 

Cheers, SAM

[opus101]

4 minutes ago, SoundAndMotion said:

I'm going to assume that for many situations, I can assume that the bandwidth of my signal is well within the flat amplitude, zero-phase region of the electronics, and that other non-static things like temperature have stabilized. Then, yes, I would want a purely static non-linearity, and I think all analog electronics would exhibit this behavior (to some extent: selectable in DISTORT).

 

Let's assume for a moment that this is true. Then it would follow that all preamp and power amplifier distortions could be compensated for by digitally pre-distorting the DAC's output. At least down to the noise floor of the DAC in question. Seems too easy.

 

[pkane2001]

11 minutes ago, opus101 said:

 

Let's assume for a moment that this is true. Then it would follow that all preamp and power amplifier distortions could be compensated for by digitally pre-distorting the DAC's output. At least down to the noise floor of the DAC in question. Seems too easy.

 

 

Actually, something I've already tested, and it does work to a larger degree than one would think. This particular type of correction is used in DeltaWave to remove measured non-linear frequency, amplitude, and phase distortions from a "distorted", recorded signal. The result is a much cleaner output, with significantly reduced distortions. But that's a discussion for another thread