why don't we use predistortion more in the audio world?

So I started thinking about melding old and new technology today and started wondering why we don't use pre-distortion more in home audio. Has anyone played with this before? If we were to measure a tube amplifier's THD, phase, frequency response, etc. couldn't we create an inverse filter which would improve the performance of the amplifier? Could this work to replace negative feedback in our amplifiers for that "zero negative feedback" sound without any of the drawbacks? Could this be taken a step further to measure an amps performance when connected to a particular set of speakers so that you take into account the complex interactions between the speaker and the amplifier?
 
predistortion requires accurate knowledge of the error you want to correct

even measured at one time you haven't captured all of the longer term variables eg tubes age, thermal history moves bias points, the power supply sags...

the technological advantage is to negative feedback - why it is universally used is that it can continue to correct for unknown, varying errors with no more "intelligence" than a difference measuring stage, excess loop gain and a feedback network

only in the weird world of audiophiles, among the subset that refuses psychoacoustic controls when making listening comparisons, does negative feedback have any poor connotation
 
  • Like
Reactions: 1 user
certainly more difficult but Volterra, Wiener series can capture/describe frequency dependent nonlinearities

a few papers have been written about using them in loudspeaker correction

I suspect Dr Geddes would claim the parts of loudspeaker distortion you could correct with predistortion aren't particularly audible
 
I once read about a multitrack studio tape recorder that had a predistortion circuit in the recording circuitry to reduce the distortion in the final played back signal. I later read an opinion of it on rec.audio.pro, the preferred sound was with the predistortion circuit turned off.

It's remarkable that someone could even manage to do a predistortion circuit back then. Nowadays it would be handled with floating-point DSP code. I've read where it's easy enough to cancel the floor bounce from a speaker. Most other problems are more subtle and harder to measure, much less knowing how to pre-correct them.
 
Predistortion would require some form of adaptive recalibration from time to time, as components age or mains voltage varies. Negative feedback solves the same problem and inherently adjusts to most reasonable circuit changes.

Predistortion is most useful in situations where NFB is difficult to apply, such as RF PAs - where there sometimes really are time delay issues. There the main issue is not exact fidelity, but avoiding spectral regrowth which can interfere with adjacent channels. If you have to satisfy, say, a -50dBc spec then you don't care what IM rubbish etc. you throw into the adjacent channel as long as it stays below -50dBc.
 
BeoLab 5 active speaker uses voice coil heating compenstaion


"
1 DSP overview. Digital Signal Processing, in brief DSP, is the generic headline for a range
of software based innovations, responsible for a level of performance and safety, that
cannot be obtained by a loudspeaker on an analogue platform. The features are all
physically integrated in processors mounted on the “engine module” of BeoLab 5, an
aluminium structure holding both the amplifier modules and power supply.
2 Thermal Compression Compensation. Playing at high volume for long periods of time
causes ordinary loudspeaker designs to decrease output significantly and synchronously
decrease timbral quality. The warmer it gets, the worse it performs. This happens because
the electrical resistance of the voice coils increase along with increase in temperature.
With Thermal Compression Compensation, dedicated software monitors the temperature
gain, and counteracts any response changes by applying the necessary corrective filtering.
If, however, maximum temperature is reached after excessive exposure to extreme signals,
the Thermal Protection System makes sure the signal level is automatically reduced to
prevent driver units and voice coils being damaged.
"
 
I started a thread a few months back on the use of pre-distortion based on neural network-type structures, but was given short shrift!

A neural network, or similar structure, can be trained to create, effectively, a multi-dimensional lookup table that is 'indexed' by a vector (an array comprising multiple values) at its input. In theory, this would allow it to produce an output based on a combination of past and future samples.

At the start of training (using a combination of test tones, noise, real music and synthetic 'difficult' signals perhaps), the signal would be pre-distorted randomly or maybe a straight 1:1, and the amplifier's output into the particular speaker monitored for accuracy. Gradually, the pre-distortion would be adjusted in the direction of better accuracy. Once trained, the network would be frozen and used in 'runtime mode' only.

The amplifier would be genuinely feedback-less at runtime, but an objection someone raised before was that it was, in effect, simply deferred feedback (from the training). However, the training does more than simply feed back an instantaneous amplitude error to the input: the 'error' can be any measurement we wish, including total least squares error or peak error over some time period etc. At runtime, the amplifier can see into the future so, for example, the system can deliberately allow (introduce) some error that could otherwise have eliminated, if it is going to allow the amplifier to 'prepare' better for a much more troublesome signal further ahead- this is the sort of thing that can fall out of the training, automatically.

Factors such as transistor and speaker coil temperature could be fed into the network as more inputs, but neural networks become more practical the more you eliminate 'dimensions', so we might want to run the output transistors at a fixed temperature, for example.

It may be a completely impractical idea, but I think it would be possible in theory.
 
  • Like
Reactions: 1 user
predistortion requires accurate knowledge of the error you want to correct

It requires accurate prediction of the signal that, when added to the input signal, will minimise the output error. Which isn't necessarily the same thing, I would suggest.

Presumably even a partial correction would allow you to reduce the amount of NFB required to achieve the same distortion level, so it sounds like an idea worth investigating if you hate the idea of NFB. People around here are always fiddling about with bias currents etc. in order to do something similar, without accurate knowledge of the error they're trying to correct.

There does seem to be an inconsistency between the importance some people attach to an idea like zero feedback, and their willingness to dismiss any solution that is slightly 'hybrid'. If we were to suggest a system that required a quarterly, computerised, 1 hour re-calibration sequence to be performed, it would be dismissed out of hand as a joke, too inconvenient. If, on the other hand, we were to suggest that the audio storage medium be of a format that could easily be damaged by normal handling and required 'cleaning' upon every play, no one would bat an eyelid.
 
CopperTop said:
It requires accurate prediction of the signal that, when added to the input signal, will minimise the output error. Which isn't necessarily the same thing, I would suggest.
I think you will find that this is mathematically equivalent to what jcx said. You need to get the predistortion right. That means careful calibration. This will never be exact. The result could be lots of low level high order distortion. If you think of valve and SS as being at opposite ends of a line of distortion 'flavour', this would be beyond SS!

CopperTop said:
Presumably even a partial correction would allow you to reduce the amount of NFB required to achieve the same distortion level, so it sounds like an idea worth investigating if you hate the idea of NFB.
Possibly less NFB, but probably higher frequency NFB. Incomplete predistortion will have the same effect as re-entrant distortion from NFB: proportionally more higher order terms. It solves the problem already more easily solved by NFB, while introducing the same problem as NFB.

As I said, people only use predistortion where NFB cannot be used.
 
I think you will find that this is mathematically equivalent to what jcx said.

Not sure about that. If I can measure the error accurately (between actual output and scaled up version of input signal), it still doesn't tell me what I must add to the input in order to correct it... To do that, I effectively have to have accurate knowledge of the characteristics of the output stage. This is the sort of thing that is done when linearising a system in industry using an inverse neural network model of the 'plant' (or so I believe).
 
I started a thread a few months back on the use of pre-distortion based on neural network-type structures, but was given short shrift!

A neural network, or similar structure, can be trained to create, effectively, a multi-dimensional lookup table that is 'indexed' by a vector (an array comprising multiple values) at its input. In theory, this would allow it to produce an output based on a combination of past and future samples.

At the start of training (using a combination of test tones, noise, real music and synthetic 'difficult' signals perhaps), the signal would be pre-distorted randomly or maybe a straight 1:1, and the amplifier's output into the particular speaker monitored for accuracy. Gradually, the pre-distortion would be adjusted in the direction of better accuracy. Once trained, the network would be frozen and used in 'runtime mode' only.

The amplifier would be genuinely feedback-less at runtime, but an objection someone raised before was that it was, in effect, simply deferred feedback (from the training). However, the training does more than simply feed back an instantaneous amplitude error to the input: the 'error' can be any measurement we wish, including total least squares error or peak error over some time period etc. At runtime, the amplifier can see into the future so, for example, the system can deliberately allow (introduce) some error that could otherwise have eliminated, if it is going to allow the amplifier to 'prepare' better for a much more troublesome signal further ahead- this is the sort of thing that can fall out of the training, automatically.

Factors such as transistor and speaker coil temperature could be fed into the network as more inputs, but neural networks become more practical the more you eliminate 'dimensions', so we might want to run the output transistors at a fixed temperature, for example.

It may be a completely impractical idea, but I think it would be possible in theory.

I postulate that current technology available is of a level that could utilize predistortion to improve an amplifiers response further then what is currently possible with negative feedback. It would need regular recalibration but I dont see why you couldnt use test tones initially and then the music signal itself later. Hardware wise, couldnt you use a raspberry pi, a dac and an adc? It is really the software that is beyond my skills.

Exactly how many parameters are able to be improved by pre distortion? Frequency response and phase are obvious but van it be used to improve tramsienr response or imd?
 
CopperTop said:
Not sure about that.
I assumed jcx meant error in the sense of imperfection in the in-out function, not the raw output-'desired output'. If you know what the forward function is then in principle you can calculate the inverse function and use it for predistortion.

DJNUBZ said:
I postulate that current technology available is of a level that could utilize predistortion to improve an amplifiers response further then what is currently possible with negative feedback.
I postulate that the best you are ever likely to achieve will be worse, and more expensive, than properly-used NFB. Audio is an ideal application for NFB, because forward time delays are negligibly small when compared with bandwidth and equipment size is such that the lumped quasi-static (i.e. non wave) approximation of EM is valid.
 
I assumed jcx meant error in the sense of imperfection in the in-out function, not the raw output-'desired output'. If you know what the forward function is then in principle you can calculate the inverse function and use it for predistortion.


I postulate that the best you are ever likely to achieve will be worse, and more expensive, than properly-used NFB. Audio is an ideal application for NFB, because forward time delays are negligibly small when compared with bandwidth and equipment size is such that the lumped quasi-static (i.e. non wave) approximation of EM is valid.

Well that is entirely possible as well, in fact without any testing we are both equally as right. Are there any papers on this field of work that might shed some more light?

Negative feedback is relatively simple, low cost and works but it is about 90 year old technology. Now that digital is at a point where it is "relatively" simple, low costs and works, shouldn't we test this? I imagine that a predistortion system in a desgin where the ADC,DAC and processor are all part of the amplifier could offer benefits that negative feedback isn't capable of doing. This system wouldn't have to be constantly measuring either, it could take samples over time and adjust accordingly to account for part wear over time. Sort of like digital room correction except that it is internal to the amp and only serves to maximize the performance of the amp and nothing else. If I am thinking clearly (I haven't had any coffee yet today) I don't see why you couldn't use such a technology to flatten response and lower THD to vanishingly low levels. You could also correct for the phase issues that seem to be inherent in tube amplifers? There is a company making a class D amplifier with very low distortion that employes hawksford's forward error correction. I don't see how this wouldn't be taking everything to the next level above that.

Could you build a proof of concept using DRC? Instead of using a mic in a room, you take the signal off of the amp?
 
All this digital correction stuff suffers from change in the system from the time calibration is performed. Why not use a simple time accurate analog solution?

A high frequency low amplitude pilot signal gets summed into the amplifier input and the pilot gets measured at the output of the system (an amplifier here, speakers would be too slow). The output pilot level is used to adjust the gain of the system by varying the tail current in a differential stage so that the pilot output level remains constant. The pilot signal gets removed from the final output by either a filter, adding a nulling signal, or just ignoring it, since it is above the audible frequencies. This has the advantage over conventional NFB in that a fixed low gain level (an R attenuator) path is all that is needed. No stability problems here.
 
control theory is a lot broader than audio amplifier design - but often the extensions of the theory were driven by problems encountered controlling machines, chemical processes, many industrial applications where the limitations of the "plant", sensors are major impediments, changes expensive or physically impossible and lots of effort is justified on the control side


in audio amplifiers the "plant" would correspond to audio gain stage, output transistors depending on your viewpoint

but audio output transistors are so much faster than the required "working bandwidth" for audio signals,
amp error sensing (feedback network, input diff pair) can have more S/N than the source recording we want to reproduce, is far more linear than any audio transducer

together these mean we can pretty much do as much as can be done by any method with negative feedback alone



dynamic speakers are physical performance limited "plants", sensing adds expense, economical sensors have accuracy, noise limits - so only limited negative feedback can be applied - and that really only for subs/woofers


for lots of info on dynamic speaker driver distortions, measurements, design cruise the Klippel site:

Home

specifically for a summary view of loudspeaker distortion correction try:

http://cogsys.imm.dtu.dk/nonlincomp/Klippel.pdf

and its often worthwhile poking around sites Google directs you to - to see if related material is in the higher level directories:

http://cogsys.imm.dtu.dk/nonlincomp/
 
Last edited:
I am slightly confused, because the learned replies suggest that there is no question concerning feedback, and that amplifiers are understood completely. I'm happy to accept that, but I thought there was supposed to be a controversy or something..? I could have sworn that I have seen threads with thousands of posts concerning the mysteries of why some amplifiers measure well apparently, but sound mediocre, with suggestions that it may be transient instability, thermal distortion etc. exacerbated by feedback, and beyond conventional analysis.
 
  • Like
Reactions: 1 user
CopperTop said:
and beyond conventional analysis.
People who don't understand something will sometimes claim that it is beyond conventional analysis. That way they can either feel more comfortable in their ignorance by believing that nobody else understands it either, or they can leave the field supposedly open for their unconventional 'explanation'.