Hallo!
Yes, it would be great if some improvements got made to libxtract as a 'side effect' of the Pd GSoC! Something that might be interesting would be a set of MIR-inspired abstractions that use the libxtract/aubio bindings + the Pd machine learning objects ([knn], [ann_mlp], [ann_som]) to do common tasks - e.g. speech recognition, audio segmentation with labelling etc. A kind of Pd MIR toolkit.
Yes that would be a nice project: first coding the vamp plugin and afterwards a set of abstraction (MIR toolkit) with some common tasks. Then these abstractions would also act as documentation so that pd-people can reproduce the way they work. And if some lower-level features are missing they could be added to libxtract or other VAMP plugins.
Thinking about it, it probably would be a good idea if the vamp host was developed separately from the PluginHost. Particularly for the purposes of GSoC, I think projects should stay a manageable size. The important thing is that the two project teams (VampPlugins, PluginHost) communicate with each other so that we retain the possibility to merge the vamp host code into the generalised plugin host at some later stage if it seems appropriate.
Hm ... I am not sure if this is such a good idea. I guess that LADSPA/VST/... plugins are quite different to VAMP plugins, which output control data. Therefore I am not sure if it actually makes sense to combine those plugin hosts ... (of course they could have a similar syntax/interface, so that it's easy to use for users)
LG Georg