Thanks to the excellent advice from Emmanuel and nice GUI design from Massimo, you can perform functional multiband/multiclass SVM from QGIS.
OTB comes with a detailed Classification Tutorial and there are samples in Monteverdi code to train the classifier using pixels in a polygon to generate SVM Models. Some vector reprojection code also helped.
I tried it out on a sample LandSat TM dataset supplied with ENVI. It has 6 bands, I digitized by hand 3 training areas - vegetation, light soil/rock and dark soil/rock. The current implementation only allows a single training area per label, this is a limitation of my patchy understanding of the OTB vector data attribute parsing system. Ideally any number of training polygons should be used if they have an attribute corresponding to the labels. Not a bad implementation for 3 days of work over a holiday, while running around buying houses.