I am still rolling variant Splif's... but I must admit they usually go up in smoke...
Altmann:
So it's interesting about the bass and the ambience and vocals, with lack of fatigue.
Bass is perhaps released of the servo like grip, so we now hear the speaker being driven, rather than controlled.
If the GNFB of the speaker wires affects the bass - I wonder what it's doing to the mid and treble..
Better imaging and a more airy sound, and no fatigue - this is interesting as it implies the GNFB is 'fighting' the speaker signal - any correction to the voltage on those wires is probably going to affect phase of the treble, even just slightly.
When I put my OPT outside of the GNFB loop, I also found the same effect: Better bass, more effortless and open sound, and better vocals. The transistor version of using seperate MOSFETs for the feedback and speaker is probably a purer way of decoupling the speaker from the smplifier, but perhaps the OPT does a similar thing.
Time to hunt for an old Quadrophonic receiver and modify it? 🙂
I wonder if this scheme gives similar effect to just putting a resistor after a (global Fdbk ) amplifier to lower the damping factor?
I was at a hifi store that loaned out a resistor switching box for series connecting resistors to loudspeakers, then selling sets of resistors according the resistor selected. It not only changes the damping but can change the frequency response as well. The effects can be dramatic... as expected it could.
Bob Carver once did a demonstration at an American HiFi show, claiming he could make a semi-con amplifier sound just like his 14 x KT88 amplifier, changes to be made overnight. Next day, attendees (under less than optimum conditions, but still) were unable to tell them apart. All he had changed is described in post #344 and #345.
All good fortune,
Chris
All good fortune,
Chris
You probably need a separate bias circuit for each output pair to control current through them. Or select the transistors/Mosfets for bias thresholds. Some emitter/source degeneration resistors could help too.
I think he was referencing a different type of splif 😀
I think this is very much related, because avery GNFB does in fact have a resistor (inside an inductor) - the Zobel network, which isolates the amplifier to the resistor value (at very high frequencies), to prevent the tail wagging the dog.
A resistor has a similar, wideband isolating effect.
The OPT of a tube amplifier is a similar barrier, the GNFB of which is naturally limited by the 'phase delay' of the OPT re. stability.
The Altmann idea, which I appear to have re-invented LOL, takes this to the limit - of completely isolating the amplifier.
So I wonder if Carver's resistor simply allowed the amplifier to behave better?
What is damping factor? Damping implies control, the servo control of an amplifier over a speaker.
At 20Hz, this is easy! At 5kHz - perhaps it's a bit of a mess 😀
i suspect if the output and input of an amplifier (playing real music) are compared on a scope (one with a 'subract trace' perhaps), both driving a speaker directly, and then via a 10R resistor - the waveform at the amplifier output would be a closer match when the resistor is there.
I recreated Baxandall's distortion tests and there is a giant wrinky pink elephant in the room which Jan Didden mentioned in his article. It mainly applies to transistors that are driven to the point of almost turning off, IE producing 10% distortion before the addition of feedback.
When you actually look at the waveforms it's obvious that feedback is pushing the transistor (diode) further toward turning off, increasing the abruptness of the transfer curve, which we already know increases high order harmonics without the need to resort to unwieldy math.
Bruno's comment about applying feedback to SET amps is on the money as this is that exact scenario. So as a rule of thumb, stages with compressive behavior should not be used with weak feedback if they swing below 1/3rd of their quiescent current or their compressive limit. Without feedback they will naturally resist turning off and generating high order harmonics, but when you add feedback it pushes them harder into that limit generating a sharper error voltage curve.
Those who arrive at this purely through math may overlook the fact that what they are calculating represents a transistor driven almost into clipping. If you use sane design practices, you don't get this problem. Years ago on this forum we neophytes were told we should set the quiescent of transistors so that we don't have more than 50% modulation at max output - this avoids the problem almost entirely. 0% modulation would be ideal obviously, but the vast majority of the problem occurs above 50% modulation.
Think about it this way. If your transistor is driven to 10% peak compression without feedback, then feedback will push it 10% further. Then the transistor is pushed to 20% compression and so high order harmonics skyrocket. These numbers aren't strictly accurate but the concept is correct.
It is not an issue with EF output stages because the transfer function doesn't go anywhere near 10% compression even though the transistors are turning on and off. Beta droop could do it if driven hard, but then SOA would also be suspect.
Furthermore, the issue of re-entrant distortion is actually far, far simpler than past discussion tends to suggest. Here is a much simpler and intuitive way to understand what is happening.
When your amp receives an input signal, the input transistor converts it's input voltage into a collector current. The collector current is converted to a voltage at the output of the amp which eventually becomes a voltage that is subtracted from the input voltage of the transistor. Notice that we have a current that is being converted to a voltage. This is the same thing a resistor does, the only difference being the current and voltage appear at different nodes making this a transresistance. Therefore the transresistance of the feedback loop around the input transistor can be modeled as a degeneration resistor at the transistor's emitter. Since loop gain falls above a certain frequency there would also be a capacitor across that resistor.
The "successive approximation" of a 1st order feedback loop then is no different than the "successive approximation" of this capacitor. For a 2nd order loop like in TPC, the capacitor just becomes an impedance with a higher slope, but at the unity loop gain point it is dominated by the 1st order slope in order to be stable. What this means is that there is no more "recursion" or "circulation" in this feedback loop than would occur in a transistor loaded by a capacitor. If the gain drops to the point where the unity gain point becomes a 2nd order slope, we do get a sort of recursion in the form of ringing and instability (an inductor in series with the capacitor). There you go - if your amp is not ringing or oscillating then you have no recursion to worry about. This also corroborates the fact that even though minimum phase systems technically have group delay, that delay is always too low to cause recursion in a stable amplifier, and the physical delay is far lower still. The main distortion effect of recursion due to a barely stable amp would be successive demodulation of peaks after stimulation of the ringing frequency. So a barely stable feedback loop leads to demodulation at the resonant frequency to the degree that one of the internal stages gets overdriven.
So say we have a loop gain of 100. Our BJT input transistor biased at 1mA has 26ohm transresistance, so for 100x gain our effective emitter resistor must be 26ohm*100=2.6k. With such high degeneration we would expect the current swing to be tiny and this is what we observe when we close the feedback loop; feedback is reducing the current swing of the input transistor and it's distortion by acting like a 2.6k degeneration resistor.
Since the degeneration is so much larger than the BJT's transresistance, it dominates the transfer curve until the transistor's compression becomes quite severe. By the time the transistor's dynamic resistance rises to the value of the resistor, the transistor's current is 0.026/2.6k=10uA! The transistor itself is peak compressing at near 100% but since it's transresistance is equal to the degeneration, the system compression is only around 50% at the crest of the waveform. The higher order harmonics however have skyrocketed because feedback/degeneration does not allow the transistor to limit it's own output swing.
This is what would happen if feedback did not also decrease current swing at the same time. However current swing actually has a stronger effect than feedback. A transistor's harmonic distribution is a function of it's current modulation. So what we observe is that in an input stage, feedback decreases high order harmonics rather than increasing them.
So there you go. This phenomenon boils down to:
1: Feedback in some cases can increase the sharpness of the error waveform (high order harmonics) by not allowing transistors to limit their own output swing when they are nearly turning off.
2: Feedback increases high order harmonics only by pushing a stage further into compression. This largely occurs over 0-20db of feedback where the most swing change occurs whereas the difference between 40db and 60db of feedback is minor for the harmonic distribution.
2: In the input stage where feedback decreases current swing, feedback reduces high order harmonics.
3: In the VAS where feedback increases current swing, feedback increases high order harmonics, but you would have to reduce the feedback to below 10db to have a meaningful effect - resulting in all harmonics being high even if their distribution is lower in frequency. The better way to deal with this is to increase the VAS quiescent so it's modulation is lower.
4: At quiescent modulation lower than 50% this phenomenon is virtually nonexistant.
5: Whether feedback increases high order harmonics or decreases them depends entirely on topology. When closing the feedback loop causes a transistor to experience less current swing, feedback is directly reducing high order harmonics, not increasing them. When closing the feedback loop causes a stage to be pushed further into compression, high order harmonics increase and the error waveform becomes sharp.
6: Feedback does not have any special effect on high order harmonics that passive components wouldn't also have, except in the case of 2nd order feedback slopes which just behave like a more severe version of a passive component. If feedback is over 20db, 2nd order slopes only reduce harmonics, even if the reduction of high order harmonics is less than the reduction of low order harmonics.
These apply to transconductance compression, but can be adapted to compression in other forms of gain that occur in a feedback loop.
Another way to look at this is that gain staging is more important than level of feedback if your goal is a good harmonic distribution. A transistor's harmonic distribution is directly related to it's modulation and feedback in itself has no special magical effects that could not be modeled as a degeneration resistor/capacitor or higher order impedance.
The same effects occur with voltage swing and have voltage analogues, but current swing is usually the worst offender.
When you actually look at the waveforms it's obvious that feedback is pushing the transistor (diode) further toward turning off, increasing the abruptness of the transfer curve, which we already know increases high order harmonics without the need to resort to unwieldy math.
Bruno's comment about applying feedback to SET amps is on the money as this is that exact scenario. So as a rule of thumb, stages with compressive behavior should not be used with weak feedback if they swing below 1/3rd of their quiescent current or their compressive limit. Without feedback they will naturally resist turning off and generating high order harmonics, but when you add feedback it pushes them harder into that limit generating a sharper error voltage curve.
Those who arrive at this purely through math may overlook the fact that what they are calculating represents a transistor driven almost into clipping. If you use sane design practices, you don't get this problem. Years ago on this forum we neophytes were told we should set the quiescent of transistors so that we don't have more than 50% modulation at max output - this avoids the problem almost entirely. 0% modulation would be ideal obviously, but the vast majority of the problem occurs above 50% modulation.
Think about it this way. If your transistor is driven to 10% peak compression without feedback, then feedback will push it 10% further. Then the transistor is pushed to 20% compression and so high order harmonics skyrocket. These numbers aren't strictly accurate but the concept is correct.
It is not an issue with EF output stages because the transfer function doesn't go anywhere near 10% compression even though the transistors are turning on and off. Beta droop could do it if driven hard, but then SOA would also be suspect.
Furthermore, the issue of re-entrant distortion is actually far, far simpler than past discussion tends to suggest. Here is a much simpler and intuitive way to understand what is happening.
When your amp receives an input signal, the input transistor converts it's input voltage into a collector current. The collector current is converted to a voltage at the output of the amp which eventually becomes a voltage that is subtracted from the input voltage of the transistor. Notice that we have a current that is being converted to a voltage. This is the same thing a resistor does, the only difference being the current and voltage appear at different nodes making this a transresistance. Therefore the transresistance of the feedback loop around the input transistor can be modeled as a degeneration resistor at the transistor's emitter. Since loop gain falls above a certain frequency there would also be a capacitor across that resistor.
The "successive approximation" of a 1st order feedback loop then is no different than the "successive approximation" of this capacitor. For a 2nd order loop like in TPC, the capacitor just becomes an impedance with a higher slope, but at the unity loop gain point it is dominated by the 1st order slope in order to be stable. What this means is that there is no more "recursion" or "circulation" in this feedback loop than would occur in a transistor loaded by a capacitor. If the gain drops to the point where the unity gain point becomes a 2nd order slope, we do get a sort of recursion in the form of ringing and instability (an inductor in series with the capacitor). There you go - if your amp is not ringing or oscillating then you have no recursion to worry about. This also corroborates the fact that even though minimum phase systems technically have group delay, that delay is always too low to cause recursion in a stable amplifier, and the physical delay is far lower still. The main distortion effect of recursion due to a barely stable amp would be successive demodulation of peaks after stimulation of the ringing frequency. So a barely stable feedback loop leads to demodulation at the resonant frequency to the degree that one of the internal stages gets overdriven.
So say we have a loop gain of 100. Our BJT input transistor biased at 1mA has 26ohm transresistance, so for 100x gain our effective emitter resistor must be 26ohm*100=2.6k. With such high degeneration we would expect the current swing to be tiny and this is what we observe when we close the feedback loop; feedback is reducing the current swing of the input transistor and it's distortion by acting like a 2.6k degeneration resistor.
Since the degeneration is so much larger than the BJT's transresistance, it dominates the transfer curve until the transistor's compression becomes quite severe. By the time the transistor's dynamic resistance rises to the value of the resistor, the transistor's current is 0.026/2.6k=10uA! The transistor itself is peak compressing at near 100% but since it's transresistance is equal to the degeneration, the system compression is only around 50% at the crest of the waveform. The higher order harmonics however have skyrocketed because feedback/degeneration does not allow the transistor to limit it's own output swing.
This is what would happen if feedback did not also decrease current swing at the same time. However current swing actually has a stronger effect than feedback. A transistor's harmonic distribution is a function of it's current modulation. So what we observe is that in an input stage, feedback decreases high order harmonics rather than increasing them.
So there you go. This phenomenon boils down to:
1: Feedback in some cases can increase the sharpness of the error waveform (high order harmonics) by not allowing transistors to limit their own output swing when they are nearly turning off.
2: Feedback increases high order harmonics only by pushing a stage further into compression. This largely occurs over 0-20db of feedback where the most swing change occurs whereas the difference between 40db and 60db of feedback is minor for the harmonic distribution.
2: In the input stage where feedback decreases current swing, feedback reduces high order harmonics.
3: In the VAS where feedback increases current swing, feedback increases high order harmonics, but you would have to reduce the feedback to below 10db to have a meaningful effect - resulting in all harmonics being high even if their distribution is lower in frequency. The better way to deal with this is to increase the VAS quiescent so it's modulation is lower.
4: At quiescent modulation lower than 50% this phenomenon is virtually nonexistant.
5: Whether feedback increases high order harmonics or decreases them depends entirely on topology. When closing the feedback loop causes a transistor to experience less current swing, feedback is directly reducing high order harmonics, not increasing them. When closing the feedback loop causes a stage to be pushed further into compression, high order harmonics increase and the error waveform becomes sharp.
6: Feedback does not have any special effect on high order harmonics that passive components wouldn't also have, except in the case of 2nd order feedback slopes which just behave like a more severe version of a passive component. If feedback is over 20db, 2nd order slopes only reduce harmonics, even if the reduction of high order harmonics is less than the reduction of low order harmonics.
These apply to transconductance compression, but can be adapted to compression in other forms of gain that occur in a feedback loop.
Another way to look at this is that gain staging is more important than level of feedback if your goal is a good harmonic distribution. A transistor's harmonic distribution is directly related to it's modulation and feedback in itself has no special magical effects that could not be modeled as a degeneration resistor/capacitor or higher order impedance.
The same effects occur with voltage swing and have voltage analogues, but current swing is usually the worst offender.
The series output resistor or the Splif scheme may be removing/reducing some speaker currents that don't jive with the audio signal, like resonances, cone breakup, crossover deficiencies .... It could be interesting to put a differential scope probe across the series resistor or between the Splif speaker channel and the Splif resistor load channel. Then try some square wave signal or a frequency scan to see what shows up there. Headphones across too maybe to hear if the fatigue stuff is there when playing music.
What seems missing in this analysis is the strong contribution of IM products that occurs in single ended amplification or single sided distortion. In SE tube amplifiers for example a square law non-linearity creates DC shifting internally that doesn't transfer this DC to the output through the transformer. However all IM products created as difference frequencies above the cutoff do appear. So it is that SE amplifiers can create "artificial" low frequencies, as giving an impression of added bass or weight to all manner of presentations throughout the frequency range.
Although this can be recognized as a positive attribute it requires such attention to the all manner of gain/efficiency in the overloading in the totality of the connections as making it effectively impossible to replicate.
Although this can be recognized as a positive attribute it requires such attention to the all manner of gain/efficiency in the overloading in the totality of the connections as making it effectively impossible to replicate.
Conventional triode amps with no feedback prefer audiophiles who listen to recorded for them audiophile recordings, with one instrument or a voice soloing each time on background of soft rare accompaniment. Increasing with signal level distortions, widening of their spectrum simulate increasing dynamics. But they are not the best for orchestral music, intermodulations kill the sound. That's why I design amps with nested feedback loops, so overall distortions are very low, but no stage goes out of steam like in case of a single global feedback loop when loudness increases, shifting bias due to non-linear asymmetric currents through capacitors. Everything is balanced and controlled to get minimal distortions at decaying sounds to render "air", reverberation right, but distortions at higher levels are hiding behind perceptions that ignore distortions of physical media including own eardrums at high loudness, due to selection of load lines.
Feedback is not bad, if to use it properly to get proper end results.
Feedback is not bad, if to use it properly to get proper end results.
Though the extent of feedback seems often the case going far beyond what is necessary to diminish harmonics or IM from a sonic perspective. That in going so much farther seems can create sonic artifacts as perhaps unidentifiable by conventional measurement, suspiciously presenting of alternative influences perhaps because of excessive feedback going too far.
"Feedback is not bad, if to use it properly to get proper end results." Sure, but what is "proper"?
"Feedback is not bad, if to use it properly to get proper end results." Sure, but what is "proper"?
I would think any sonic artifacts are measurable but maybe not perceived with the same weight? I have the same question, what is excessive? Designs from the 50's and 60's seem to routinely have 20dB or more of global negative feedback, did they get it wrong?
Last edited:
Many of the early US and UK transistor amps used to go up in smoke, but the Japanese transistors didn't. But look at their THD figures, they do not use much feedback.
I think the problem is one of coupling the speaker to the amplifier, too much feedback turns it into a control freak, and this gets sloppier as the frequency rises, so much so that a Zobel network is needed.
The Zobel is the clue that something's not right - it's the duct tape, the kludge. I think the DartZeel sounds good because it's not fighting the speaker. Tube stuff doesn't tend to do that either, as the OPT prevents excessive GNFB.
Opamps sound good - and they use LOTS of GNFB - but they don't drive speakers (Well, chip-amps excepted). The Altmann dummy output feedback system probably sounds good too - they are the transistor solution for Low THD with good sound perhaps, although the Audionics CC2 / REdesigns manage it - but those are built with the goal of stability - which specifically aims to prevent the speaker from upsetting the NFB loop of the amplifier, so again we can see it's isolated resonably well.
An article in April 1954 Wireless World helped me understand a little bit more of what I think Wavebourn is getting at:
Distortion in Negative Feedback Amplifiers - Thomas Roddam (pages 169-172)
Distortion in Negative Feedback Amplifiers - Thomas Roddam (pages 169-172)
Doesn't explain why my 6922 preamp blows away an opa2134 as the preamp section.
* I needed to be convinced before
* I needed to be convinced before
Great to read back into the past, thanks for posting that, I love the wisdom and care in these old magazines!
Wavebourn is perhaps referring to the internat gymnastics of some multistage amplifiers that are straightened out with a single global feedback loop. Looking at the voltage across the feedback resistor can be instructive here.
By nesting the feedback loops, or choosing very linear stages, less correction is needed and so the voltage swings inside are tamed and the whole amplifier tends to behave in a manner that we'd like to imagine 🙂
Additionally, a complex waveform (orchestra) gets more and more complex as it passes each non-linear stage, so a single global loop has a serious problem at this point, it's just going to be multiplying a mess of harmonics. The nested feedback however tames a manageable amount in a very fast way (NFB works best with fast slew rates).
That's my understanding of the situation, matching what Wavebourn is describing I think.
The particular shape of the transfer curve affects how the amplifier sounds. The position on that transfer curve you choose for signal zero also affects it. So if you want a particular result, you achieve it by shaping the transfer curve(s).
This assumes that the distortion is audible to begin with. In order to achieve any particular result in terms of sound, the difference must be audible. But since in the big picture, perfection in the recording chain does not exist, there is not much harm from subtle distortions if they serve some purpose. I accept this even though I always focus on eliminating distortion.
This assumes that the distortion is audible to begin with. In order to achieve any particular result in terms of sound, the difference must be audible. But since in the big picture, perfection in the recording chain does not exist, there is not much harm from subtle distortions if they serve some purpose. I accept this even though I always focus on eliminating distortion.
You are correct. 😉
Also, using nested loops we make each stage bandwidth wider, need less of a global feedback, so the amp is fast, wide bandwidth and flat inside of the audio band, and even may be no need for compensation for the stability with the global feedback loop.