With limited training data, infrequent triphone models for speech recognition will not be observed in suficient number. In this report, a speech production approach is used to predict the characteristics of unseen triphones by using a transformation technique in the parametric representation of a formant speech synthesiser. Two techniques are currently tested. In one approach, unseen triphones are created by concatenating monophones and diphones and interpolating the parameter trajectories across the connection points. The second technique combines information from two similar triphones; one with correct context and one with correct midphone identity. Preliminary experiments are performed in the task of rescoring recognition candidates in an N-best list.