Modeling pronunciation variation with context-dependent articulatory feature decision trees