Obligatory show layout and schematic. But the problem with gross audio frequency return currents is that they just go wherever they want unlike well behaved high frequency return currents that just stay under the trace.
Is a possible solution to make sure I only reference a virtual ground that I create off something like a shunt voltage reference? I'm using a bipolar power supply currently though. My negative rail is a little noisy in very high frequencies because it's generated by a charge pump but it's got a low noise LDO on it, and the system performance seems the same when I swap it for my bench supply.
Same idea applies to power planes, if you have those and if the sensitive stuff has insufficient PSRR and/or you have signals referenced to the power plane (including voltage dividers).
I've tried slots and splitting the ground and power planes in a range of ways. I've currently got power on the outer 2 layers and ground on the middle 2, routing signals on the outsides. I found continuous planes slots a little more effective than any of my attempts at star style grounds, but I'm not very familiar with how stars are actually meant to be routed. I've got a bunch of power (3.3, 5, -12, 12) so I am a bit confused about how to route the power correctly. I am biasing the JFETs with resisters relative to power and ground.
Have other people resorted to fill isolation in this case? I've seen a few audio products that do (with vactrols / optos for signals) so I assume it's not totally crazy but I'm wondering if the cost / complexity trade off is worth it. Is it the easiest solution given I've tried at least 10 board revisions...