I'm certainly no signal theory guru but I have worked for a while on a
modal/additive synthesis engine (which the FTM roughly arrives at).
While exploiting the parallel nature of the synthesis model is fairly
straightforward, it still takes massive computational resources to
produce musically useful results, calling for expansion beyond a
single core, or even beyond the CPU. The subsampled polyphase filters
mentioned in the second paper are a neat trick that provides sizeable
efficiency gains, but this comes at the cost of signal quality and
implementation simplicity; it also introduces model latency.
I doubt that a free implementation of wide applicability will arrive