Adding RDS to the signal might be a big work, but I don't need it. I can import excel files to the Rigol Ultrawave-software, so maybe I could create something. The DG-1022 can handle only 4096 points of data in the waveform, but maybe this is sufficient if keep the signal simple.
For example if I create signal where there is 1kHz for left channel and 2kHz for right channel. I think it is enough if this 4096 point data has one cycle of 1kHz signal and 2 cycles of 2kHz signal. The modulated stereo signal would be on bandwidth between 23kHz and 56kHz. So I would need 56 cycles for this 56kHz signal. I guess this 4096 points of data is more than sufficient for this (73 samples per cycle).
The math would be something like this. All signals defined between -1 .. 1.
L(t) = 1kHz signal
R(t) = 2kHz signal
P(t) = 19kHz pilot signal
M(t) = 38kHz carrier
The final signal would be L(t)+R(t)+0.1*P(t)+X*(L(t)-R(t))*M(t).
Could this work? The only thing I don't know is the constant X.