Non-linear speech transition visualization

Modelling context effects and segmental transitions in speech recognition systems is very important. Explicitly modelling segmental transitions in a RNN framework would circumvent these problems. We present an interesting application of Principal Curves, an algorithm to extract a non-linear summary of p-dimensional data firstly published by Hastie and Stuetzle (1989). The algorithm can be used to visualize non-linear transient characteristics in speech. We show that between-phone characteristics found within diphones can be used as discriminant information to distinguish ambiguous phones. The technique used is explained and illustrated on the examples /bah/, /dah/ and /gah/.