[img
http://oi58.tinypic.com/e0lrgn.jpg
Waveform at upper left represents the input signal. It is a 1 kHz sinusoidal wave, which gets muted at the middle of plotting. Waveform at the middle output signal. Output signal is evidently clipped, but not with a "fixed" threshold.
Schematic on the right illustrates the process: There are two inputs, which would be driven by a phase inverter stage of some kind. After the gain stage a half wave of the signal is clipped by a diode, in push-pull these are naturally opposing waves. A differential amp (which could as well be a transformer) combines these two signals. The "clean" wave corrects its opposing "clipped" wave and asymmetric hard clipping with dominant even order harmonics converts into softer symmetric clipping with dominant odd order harmonics (and less high order).
Instead of using a "fixed" reference for the diodes, like ground or a DC source, I use a low impedance amplifier to generate a dynamically modulated reference voltage.
Following stage is just a phase inverter to extract inverted signal for the active full wave rectifier. It rectifies the output signal and filters it. The filtered signal is an envelope that corresponds to average of the output signal (more or less). This brings us to waveforms at bottom left....
Blue is this envelope. It ramps up when the sinusoidal input signal appears. It starts to decay when input signal is muted. (This is actually one drawback of the portrayed scheme; with real power supplies recovery from sag is practically instantenious with the first "pulse" from the rectifier.)
Green signal is a "DC" reference (sort of), with which another differential amp compares the envelope signal. This signal could be well filtered DC, in this case there's simply a poorly filtered half wave rectifier tapped into same line.
Red is the output signal. It is basically the difference between the "DC" component and the envelope component inverted. The clipping diodes will clip at this voltage plus their forward voltage.