I have a solution. It's not elegant or pretty, and it doesn't work like magic, but it's tolerable. I'm quite happy to admit that there isn't an elegant or pretty solution, and there's no magic equation to put all my numbers in that will come up with the correct answer.
It's actually all come down to pragmatism and the extent to which I'll tolerate inaccuracies. For each section of the piece, I've worked out an 'ideal' distribution of voices to speakers, and several possible distributions. I total the difference between the two, and the distribution with least differences is the one that I take forward. This probably doesn't make any sense without the context but I hope that it gives an impression of the compromise that I've reached.
Of course, it's time consuming to work out but that's life. I'd rather hack away at a system that I know will produce results that I'm happy with, than come up with a quick fix which almost-but-not-quite does what I want it to.

Having taken nearly 24 hours to find this solution, I am now taking to my bed.

C had a look at the blog today and remarked: 'I'm not reading all that. It's boring.'
My language is also 'poncey'.
Sighs. Rolls eyes. Presses 'Publish Post'.