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

Successful predistortion requires knowledge of ALL possible distortions that a given amp can produce under ALL possible operating conditions. We do this in some linear RF amplifiers because there is typically a lossy cable between the RF amp and its antenna that constrains the load to something near 50 ohms. Predistortion was not possible in a portable device like a cell phone in 2011 due to a much wider range of possible antenna loads when these tests were done. We used cartesian feedback instead of predistortion in phones. The driving modulation signal characteristics are also well known and understood. The amp is thoroughly tested before the correction algorithm is generated to map these error conditions. This is different for every amp made, so testing and calibration on each unit is required. The enclosed pictures show a primitive prototype system under test over 10 years ago.

An audio amp can see a wide variety of loads, and that load will vary with frequency and amplitude of applied signal. It will also see a wide variety of music ranging from mellow to metal. There are far too many variables in play to build a predistortion algorithm unless some sort of adaptive learning is used. This was being tested for the RF (4G LTE) world when I retired 8 years ago. I have not followed that technology since I left, so I can't say what progress has been made, but DSP's have become WAY more powerful in 10 years.

The spectrum analyzer photo shows three overlapping plots. The yellow trace is the input to the amplifier. It is a 5 MHz wide "block" of LTE signal with virtually nothing on either side. The space on either side of this signal is another RF channel that could belong to a different user. Intrusion into this channel is tightly regulated by government authorities, with ETSI controlling the worldwide standards. The blue signal is a test amplifier being pushed to maximum power output (about 4 watts) without predistortion applied. The unwanted emission is degraded by about 20 dB due to internal IMD products in the amplifier. This amp does not meet legal requirements and cannot be put on the air. The pink trace is the same amp with predistortion applied after several "adaptive learning" cycles. Predistortion results in about 10 dB improvement in IMD, making this amp just barely legal over all expected load and temperature conditions.
 

Attachments

  • ScinteraSetupa.jpg
    ScinteraSetupa.jpg
    128.3 KB · Views: 77
  • TempTestSetupLargea.jpg
    TempTestSetupLargea.jpg
    117.3 KB · Views: 75
  • ScinteraRunninga.jpg
    ScinteraRunninga.jpg
    119 KB · Views: 72
  • Like
Reactions: 1 user
When I read the thread title I´m thinking tube preamp and Nelson Pass´s H2 Generator for the audiophile world.

I often asked myself though why it (obviously) is possible to create something like a "Kemper amp" (guitarist will know this) that is able to simulate (solid state+DSP) numerous guitar amps and thus mostly tube amps
but you never get to see this used for hifi?
I heard the Kemper amp and also know legendary amps like Soldano, Mesa Boogie etc. and the Kemper does a real good job of simulating characteristic distortion.
Kemper amp
 
For an inverting voltage gain stage, what if you bring the output back down to unity with a resistor divider, and then feed that signal into a second copy of the gain stage, built with closely matched components? The first stage should give you the anti-error signal for the second stage. Appropriate resistor values, and an additional RC network between stages could provide the first stage with an approximation of the load seen by the second stage, as presented by the power stage.

(Apologies if this has already been suggested - I didn't read every single post.)
 
Predistortion (the genuine non-linear dynamic variety, not just linear equalization) is sometimes used in equipment with built-in loudspeakers, for example to get slightly less unacceptable sound out of the minuscule loudspeakers built into mobile phones. It should be possible to do the same with active HiFi loudspeakers. It doesn't make much sense for an audio amplifier, as just applying a lot of negative feedback is much simpler and more robust.
 
Heh - given that stage 2 is receiving an inverted audio signal, unlike the case of simple negative feedback, I don't think the stage 1 distortion will negate the stage 2 distortion; I think it will compliment it :rolleyes:

For this concept to work, stage 2 needs a non-inverted audio signal, in which case this begins to resemble either negative feedback or error correction feedback. In either case, the use of a second stage means there's no feedback loop, so I still wonder how the behavior would differ from a single stage with a loop - specifically the distortion spectrum.
 
The ES9038Q2M DAC chip has a THD compensation function as described in the data sheet:

"THD Compensation
Sabre2M THD Compensation removes the non-linearity of the DAC resistors and to a lesser degree the non-linearity of
passive components in the output stage. Taking the I-V characteristic curve of a real resistor you will notice that it as a slight
downward curvature. As more current flows, more power dissipates the resistor heats and the resistance rises.
Non-linearity of the DAC output resistors can lead to output distortion in two ways:
• Amplitude modulation of the output current from the DAC
• Gain modulation of the output stage as the output impedance of the DAC swings with the audio signal
The Sabre2M includes models for its output resistors and can compensate for their characteristic curve by finely adjusting the
DAC codes for large and small signal amplitudes."

I have the chip but this function is off by default. It is activated by software, which puts it out of reach for me at the moment. It's interesting that it does exist, though.
 
On the other hand, negative feedback often seems to be misused and misunderstood, so it's no big deal.

Designing audio amplifiers with >10 MHz bandwidth for ultra-low internal distortion (unrealistic nonsense), only for it to be broken by speakers that inject giant phase shifts aka "time delays" in the sub-kilohertz range back into the feedback loop. Lol.


Speaker distortion from its assortment of non-linear parts often ends up getting amplified, not attenuated. Case in point: "break-up modes" that are clearly worse with some amplifiers than with others. If it's a speaker issue, then it shouldn't matter, but it does. What a headache! This is what happens when people try to compartmentalise a "system" into isolated "components" that allegedly aren't meant to interact with each other.

FIR filters are another kind of feed-forward signal modification that is "blind" to the results. It's just a matter of getting it right.
 
People who don't understand something will sometimes claim that it is beyond conventional analysis.
True. But Doug Self in one of his books details a bunch of distortions, and provides explanations for the nature of the source some of these distortions. Some distortion sources (like crossover distortion) are described as "complex'; not beyond conventional analysis.
 
The ES9038Q2M DAC chip has a THD compensation function as described in the data sheet:

"THD Compensation
Sabre2M THD Compensation removes the non-linearity of the DAC resistors and to a lesser degree the non-linearity of
passive components in the output stage. Taking the I-V characteristic curve of a real resistor you will notice that it as a slight
downward curvature.

Pardon my ignorance, but doesn't prudent selection of resistors avoid this altogether? IOW if your resistor is being modulated by heat, isn't it undersized or just the wrong resistor? We're talking about line level circuitry.

I built a power amplifier where I put oversized resistors with low temperature coefficient in the feedbak loop. Did it improve the amplifier's performance? I don't know. It does work great.
 
I can't help but think that frequency-dependent EQ is a waste of time and is bound to sound unnatural no matter how the settings are derived. The alternative is to go with an impulse response correction that aims to give you the headphone sound at your listening position - but we know that doesn't work if you move your head by a millimetre. Seems to me that the best compromise will be what most people say: cut down the reflections with acoustic treatments and only apply time- and frequency-diminishing impulse response correction at the lowest frequencies.

Frequency dependent EQ is maybe more common than you realize. It's how they get extended bass out of tiny systems. And I have been experimenting with it for a while with nice hi fi components, and if done correctly it's not only not unnatural but it makes some of the obvious shortcomings of a speaker virtually vanish.

You don't have to equalize it to withing 1 dB to make big improvements. Plus too much EQ (especially bass) will require too much amplifier power (and speaker cone travel) which makes the whole exercise pointless. A good rule of thumb is to equalize half the deficit to start and also strictly control low frequency cutoff with high order filters to avoid overdriving the woofers.


I'm not talking about equalizers like the units that were popular in the 1980s. Those are almost universally terrible. I'm talking about hard wired equalizers without potentiometers etc. It's nothing new (Bose did it way back in the day) but it works better than you might think.
 
Really? It's the foundation of the vinyl format, AKA the RIAA curve. 40dB of pre-post EQ at the extremes.

We just had this discussion the other day.

Properly EQ'd speakers can bend the laws of physics. I get so much sound out of my small 8" two ways. My buddy was just over and he was playing some bass heavy dance music on my system. We had it turned way up and he said it was just like being at the disco. He's a semi-pro DJ.

No way could they do that without EQ and a steep high pass filter. Let's say we would most definitely have exceeded xmax.
 
Well that's a different story. There was a discussion here about heat modulation of "components" on a chip and the claim that it produces a measurable artifact.
In the 1990's, when I was still at university, some guy from Philips gave a talk about their bitstream converters. He said that the first prototypes had a distortion way above spec and that they had traced it down to a polysilicon resistor with such a short thermal time constant that its temperature just followed the signal squared, even at 20 kHz. They solved it by using a physically much larger polysilicon resistor/array of polysilicon resistors.
 
Re-reading some of the older comments.. Thoughts:

-The trouble with NFB is that it often plays to criteria that have a low (or even negative) correlation to subjective audio quality, such as unweighted THD (which ignores the Fletcher-Munsen curves, masking, etc.)

-Linearising the sweeping gain curve of single-ended MOSFET stages or tubes, seems like it could be a fertile playground for correction algorithms to do the job better than analogue feedback.

-I've already received feedback (LOL) that the end result would essentially be the same, but I disagree. Analogue NFB is infinitely recursive, so even if the distortion character of a MOSFET is that it produces a small amount of H2 and nothing else, by design NFB will force it to generate H3, H4, H5... ad infinitum (-H1 and -H2 are fed back to the input, -H1 cancels part of the gain while -H2 is amplitude modulated by H1 to create H3).

-On the other hand, predistortion could have a low order, such as 2-3, reducing the strongest harmonics, but without generating a long series of new ones in higher registers that are unmasked or where the ear is more sensitive.

-I'm not sure about implementation though. Periodic calibration sounds painful. Adaptive learning techniques seem promising though.

-Another question is how dynamic are the distortions? E.g. thermally modulated gain as the MOSFET die heats and cools? Voltage drifts? This all seems fairly predictable and maybe a matter of having look-up tables with enough dimensions.
 
  • Like
Reactions: 1 user