Ultrasonic crossover modulation for lower distortion

I think I've discovered I possibly novel (though impractical) approach to reducing the effects of crossover distortion.

I was simulating a standard CFP output stage using BJTs and trying out various approachs to reduce switching distortion (which all led to nought). The idea I had is to somehow add a speed-up for the switch-off of the main output devices (unlike the darlington / emitter-follower OS a simple speed-up capacitor cannot work with a CFP OS, I had hoped to add some active circuitry to facilitate this).

Eventually I reasoned "if you can't eliminate the switching spikes, why not mess with them instead". So I added an AC current source at ultrasonic frequency direct into the output node.

This has the effect of modulating the cross-over point - rather like Douglas Self's "XD" cross-over displacement approach, but rather than a DC displacement an AC one in the ultrasonic range is injected.

So the cross-over spikes become much less correlated with the audio signal and much more with the ultrasonic tone being injected. Or thought of another way the crossover point is smeared out across a larger range of voltages (on average).

Lots of testing with two-tone intermodulation suggests this can give about 10dB increase in linearity of the CFP output stage (from about -75dB to -85dB for the worst case spurs) [ test tones of 17kHz and 19kHz just below clipping ].

The injected current was about 2 -- 3A peak for best effect.

Since a current is injected into the low impedance point of the output node there is little ultrasonic voltage imposed on the speaker load (without global feedback a CFP output is still very low impedance due to the large amount of local feedback)

Nonetheless I added an LC trap to prevent a tweeter from being exposed to this injected signal. To keep the LC values simple I used 250/π kHz (about 79kHz).

This is very much an academic exercise. For instance the power levels to inject the current are substantial, and creating a pure current source at high level isn't trivial (outside of simulators!).

The technique can address switching distortion but not other non-linearities in the CFP stage. Still 10dB improvement isn't bad (of course there is an increase of ultrasonic artifacts)

I also discovered that adding a sizable capacitor between the two bases helped reduce the difference tone (2kHz) distortion product, presumably by reducing 2nd order products caused by asymmetry.


So the output spectrum (after the LC trap) without any injected ultrasonic current:
without_injection.png



And with:
with_injection.png

Confining attention to just the audio band the spurs are now all below -60dB (the two 20V peak input signals are +23dB each, so the worst spur is -84dB from each tone.


An open question at this point is how this interacts with global feedback, the next simulations I'll undertake will address this.


I used intermodulation to probe the high frequency distortion better (high slew rates produce more switching distortion, but harmonics of 20kHz are not interesting in a real situtation (being inadible).


Clearly the practicalities are substatial and make this a curiousity only I think, but perhaps it might inspire other ideas or experimentation.

Here's the OS and its simulated load:

CFP_output_injection.png
 
I always like discussions regarding output stage bias and the by-products of the different arrangements. If there were a simple perfect solution then there wouldn't be so many varying implementations, and this forum wouldn't exist.:p The OPS is the most important regarding distortion and I notice that details of the OPS are often overlooked when people are trying to upgrade some amp from somewhat OK to something much better.
The variation of Gm as the current crosses over is the main source of crossover distortion. Maintaining a constant gm is the solution but this is not so easy when the devices only conduct half of the time. Adding approperiate emitter resistance can help even out this difference at the expense of overall Gm, but that small reduction is less important as Bipolars and mosfets typically do have high Gm. Never allowing the output devices to turn off is a better solution, as in class A bias or injecting a signal current.:) Class AB is a compromise, regarding power dissipation vs output power and there is an optimum bias for this arrangement but it is not perfect.
The variation of Gm at low conduction is even more pronounced with Darlington devices and mosfets (particularly vertical types), even though these devices generally have much higher Gm overall. While Latfets usually bias around the zero temp coefficient, verticals bias higher to overcome the low Gm region. Injecting all those crossover distortion components into the gain loop of the amplifier is sub-optimal.:dodgy::tilt: In audio, higher transconductance doesn't necessarily mean higher power or lower distortion. The linearity of Gm is more important. This is one reason many prefer to use separate driver/outputs instead of integrated Darlingtons. CFP uses local feedback current to mitigate the change in Gm but is still just a compromise. Bob covers this extensively in his book.

Injection of a signal current to keep the devices from complete shut off is an interesting solution. I guess much of it would be taken up by the low Gm operating region. Since Gm is effected by temperature, would that have an effect on the amount of injected signal?
Another solution I have played with is the Amplified Diode circuit. This was developed for Darlington devices and works very well with mosfets as they have a similar Gm profile. I wonder if the idea of injecting the HF AC current into the error loop could have a similar effect?
I have made a few different circuits using this approach but pretty much stuck to the perimeters of the concept. HERE is an example of the Amplified Diode used with vertical FETs showing the crossover improvement. Although in this example the bias is zero or class B, they would normally be class AB. Effectively, the error corrected drive signal includes the "speed up" component to replace the missing Gm at current crossover. And HERE is an example of the Amplified Diode used with Darlington bipolars.:)
As a side note, although some do not like the extra complexity I think the AD circuit is a great compromise because in addition to the local EC network, the load Z placed on the VAS becomes rather independent on the output speaker.;) Optimization of this circuit is in widening the bandwidth of stable operation since crossover is more or less a square wave. IOW, the faster the EC, the lower the bias can be.:cool:
Not so straight forward though regarding components, layout, ect.:spin:
 
Last edited:
Injection of a signal current to keep the devices from complete shut off is an interesting solution
I don't think that's what happening, I'm still seeing the switching pulses, they are just being stirred up in the time-domain, pushing their energy up to higher frequencies and averaging them out at lower frequencies - kind of like dithering or noise-shaping in a vague sense. CFP switch pulses are on the scale of < 1µs in duration typically with modern output devices with high gain-bandwidth product.


I also reduced the CFP bias resistors to 50 ohms from the normal 100 ohms to help reduce switch-off time - the driver bias currents rival the main output devices bias in fact, but since they are in the 15 to 25mA range its not too onerous (EF outputs need more like 100mA anyway)
 
It must be a good idea: I had the same many years ago :D

I studied it quite thoroughly, and I even made a number of (working) prototypes:

A dive into the past


All the best inventions were created by several people independently I believe! I think you are injecting a voltage into the signal and then having to remove it rather than a current into the output though? That limits dynamic range I think?


I only added the LC trap for safety should the OS fail and the current injector keep going - the ultrasonic output voltage is low with a CFP output stage's impedance.
 
have worked on a similar idea about 15 years ago where a small hf oscillator fed via a gain controlled amplifier
the hf output then taking from the amplifier then rectified and sent to the gain controlled amp the effect being to auto adjust the bias
The hf oscillator idea being pinched from the thought of AC bias on magnetic tape
Trev
 

Elvee

Member
2006-09-08 2:04 pm
All the best inventions were created by several people independently I believe! I think you are injecting a voltage into the signal and then having to remove it rather than a current into the output though? That limits dynamic range I think?
IIRC, I explored all the possible options, including self-oscillating amplifier, external oscillator injected in the OP stage and external power oscillator driving the OP directly.
The dynamic range is never a problem, because the base and modulating frequencies are so different, and a modulating voltage > ~twice the dead zone is sufficient to polish off most of the quirks
 
A large amplitude ultrasonic signal is added to the audio, at a frequency high enough that the magnetic domains are larger than the signal zero crossing spacing. Each domain gets flipped several times, as it passes away from the head the ultimate state (which is binary) is probabalistically determined by the audio signal (which is effectively a bias on the ultrasonic signal). The net sum of many domains in parallel statistically approximates a fairly linear version of the audio signal. The high hiss of magnetic tape is due to the individual domains.
 

benb

Member
2010-04-24 1:52 am
Amazing, I too thought of this many years ago, inspired by magnetic tape bias.

:cheers:
It's an inspirational idea. I've also thought of using ultrasonic AC bias for audio transformers, both mic/signal level as well as power output. One "problem" is, audio engineers often "like" the mic transformer sound, and use it to effect.
just wait until you are older, then your tinitus will provide the dither.....
That presumes hearing is digital!