Since the 12,5 kHz and 25 kHz were mentioned in the question, I think it might be a good idea to have an answer mentioning the Carson's rule.
The rule is used for approximate calculation of FM bandwidth, which is the kHz number referred to in 12,5 and 25.
$$ B=2(f_{dev}+f_{mod_{max}})$$
So that means that our bandwidth doesn't just depend on our deviation, but also on the highest modulating frequency.
Usually, you'll have some common combinations of maximum modulating frequency and deviation. For example, in 5 kHz systems, it's not uncommon to have a 3 kHz set as maximum audio frequency. From my interpretation of the quote in the question, I'd say that this are the 3 kHz it's talking about.
If we do the calculation, we get following:
$B=2(5 \mbox{ }kHz +3 \mbox{ } kHz)=16 \mbox{kHz}$, giving us some space for filter slope and the power outside of the Carson's bandwidth, if we're using the 25 kHz step. If we do calculation for so-called narrow-band, we'd get something like this:$B=2(2,5 \mbox{ }kHz + 2,5 \mbox{ }kHz)=10 \mbox{ }kHz$, giving us extra 2,5 kHz space for imperfections.
Here we actually have some other considerations as well. Both transmitter and receiver need to have same audio bandwidth and same deviation in order for the system to work properly. If your deviation is too high, you run the risk of going outside of channel bandwidth and sounding over-modulated. If your deviation is too low, then you'll sound very quiet on reception. Why? Because your spectral lines that represent the tone will be too close together.