The HMMs that we used are left to right, 3 states models, as depicted in figure 4.1. Context independent states are only used to model phone to phone transitions. We still kept context independent phone models for modelling the less co-articulated middle section of the phonetic units.
To obtain a trainable system, we reduced the number of parameters to a limited number of generalized transitions by tying the transition distributions. These generalized transition models, unlike classical generalized diphone models, were not necessarily attached to a particular phone.