I would definitely say that the MTP can be as accurate as DFT for the dynamical matrix, and it's a great example of the power of MTP.

You will probably only need a relatively simple training protocol, since the material will remain crystalline and not turn amorphous or anything weird, and you will only need it to be accurate around the equilibrium. Therefore the crystal protocol (crystalTrainingRandomDisplacements) will most likely be sufficient on its own, making this is a rather straightforward task (but of course, the flip side of that is that the forcefield you generate will not be generic and in particular probably not suitable for MD simulations).

Now, you will of course actually spend more time fitting the MTP than running a single DFT dynamical matrix calculation, but the benefit will be that if you later want to change some parameters and e.g. look at strained crystals etc, the phonon part will be extremely fast.