See Nyquist's stability criterion: at the frequency where the loop gain reaches 0dB, the entire loop phase shift must be lower than 180°. And a substantial margin of 45° is very desirable, so in practice it should not exceed 135°.
Your problem is that the output stage is slower than the opamp, so it introduces too much phase shift into the feedback loop. Also this particular opamp has a wimpy output stage, unable to source much current or drive the difficult load that a class-AB output stage presents. Also, its slew rate is low, so expect problems at crossover.
A quick fix would be to:
- Add a cap between the output of the opamp and the "-" input to provide local feedback at HF
- Isolate the opamp from the output stage by putting one resistor on the output (after the cap) and another resistor on the "-" input.
Like in this schematic (check C3):
(note this output stage is different from yours, it is actually a simple current feedback amplifier with gain).
If you process DC signals, it will work. Now, for lower distortion on AC signals, I'd use a real audio amplifier output stage, with discrete transistors instead of darlingtons, and an opamp with a bit more output current ability than TL072.
If it fits the specs, it is much simpler to use an integrated power op amp like LM1875 or LM3886, these are foolproof, cheap, work very well, and you get output short circuit protection too!
This one isn't bad for its age and simplicity, however...
You say it is "current feedback", well yeah, but the benefits of current feedback (ie speed) are only there when the output slews down! and Tr3 can conduct hard, pushing base current into Tr1. However if you attempt to slew in the other direction then Tr4 runs out of bias current and turns off. Tr3 turns off too.
And then you'll have to wait about... forever for the stored charge in Tr1's base to flow out through R7 which only pulls out a ludicrously low 300 µA or so.
Expect a generous amount of rail sticking on clipping, cross-conduction when coming out of clipping, and this can only run in class A since the amp is unable to turn off Tr1 fast enough to go through the crossover without humongous amounts of distortion.
Also, how is the bias set?
Guys let's be honest. It was good for its time, but now it is obsolete junk (unless you only do sawtooth with the fast edges going down...) ; LM3886 would beat this in pretty much any measurable way (and probably on sound too).
Run a simulation if you wanna have fun. I simmed a MOSFET variant a while ago, was quite fast but I had to use a 30mA pulldown on the gates... and it still had cross-conduction on clipping. Would've had to add drivers, then it's just as complex as another amp. Didn't even bother to build it. There are much better schematics.