Depends on application.
Audio "woo" aside, there are real reasons to optimize winding self-capacitance and interwinding capacitance. These go hand-in-hand with the leakage inductance as well, the combination of which sets the cutoff bandwidth of the transformer, Fc = 1 / (2*pi*sqrt(LL*Cp)), and the ratio of which sets the characteristic impedance of the transformer, Zo = sqrt(L/C).
The relevance of the impedance is not a power-matching one -- it helps, but it's not about power transfer because an ideal (LC, no R) transformer does not dissipate power! It does, however, facilitate power transfer between a source and load of similar impedance.
The impact is not so much power transfer, but bandwidth. The further the source/load impedance is from Zo, the less the bandwidth is.
Regarding inter-winding (isolation) capacitance, this is similarly unavoidable, but always undesirable. It doesn't interfere with your signal if you keep the signal fully differential (consider, if you model the total capacitance as half and half connected between either end of the windings, then the capacitance cancels out perfectly, differentially), but that's often inconvenient, and normal-mode applications are common. You can at least improve it, by adding a common mode choke which increases the common-mode impedance at high frequencies, one which is also well enough damped not to cause problems at the resonant frequency.
If you need low capacitance, you must choose a high Zo (gives high LL, low Cp) and high cutoff frequency (low LL and Cp). Having to prioritize capacitance may mean making compromises with the circuit design, or its capabilities. For example, high LL discourages flyback DC-DC converters, so you might have to work with a resonant converter instead.
Another example: medical isolation transformers, with very low capacitance (<100pF?) while operating at mains frequency. The Zo is made very large compared to the load resistance (which, say for a 120V 10A isolation transformer, is below 12 ohms to maintain good regulation). This is acceptable, as long as Fc is kept high (which is what is done: it's in the ~kHz range for most mains transformers, tens or hundreds of times the operating frequency -- so Zo can be similarly mismatched with acceptable results).
Tim