This has been addressed many times over many years, both in theory and in practicality. Articles abound in both the ARRL and Jones handbooks, also in QST, Radio, and surplus modification magazines.
What it really boils down to is reliability under continuous service.
To handle hi-fi, aka 'broadcast' audio, the response and distortion specs come into play and are the most important aspect.
To actually get a transformer to pass everything from 50Hz to 15KHz at power is no mean trick. The only way to do it is to use high permeability iron, generous copper, and intelligent winding techniques. Circuit designs also are set up for maximum performance and impedances are specified to the Ohm.
Here you have more size, weight, and expense.
For communications grade performance, the target is limited response (say 300Hz to 3500Hz) and package size. This assumes intermittent voice duty, DC in the secondary, and generally matched impedances.
To wit, look at the modulation iron from an ART-13. Very small, easily held in one hand, but legendary in the ability to handle a lot more power than designed for due to limited frequency response and winding technique.
Another example is a lot of the "poly-pedance" types with multiple tapped windings. They can be made to work in many circuits, and the power handling ability is directly proportional to the audio bandwidth running thru it.
I have seen transformers factory rated at 175 Watts doing a fine job in ham transmitters of 1kW input with a narrowband audio channel and properly set spark gaps added.
Unfortunately, the days of ordering a transformer for a few dollars and 'guinea-pigging' it in a rig just to see where it will yield is over.
Research as to what worked OK in the past would be a good idea, and patient looking for a good vintage part is still viable.
Bottom line: You can mod between 50 and 80% overload if just using restricted voice bandwidth.
73DG