Stochastic calculus, non-linear filtering, and the internal model principle: implications for articulatory speech recognition.