This paper investigates a new pronunciation verification (PV) approach obtained from the subspace based Gaussian mixture model (SGMM) based pronunciation model. A single SGMM model is trained from disabled speakers' utterances and reference speakers' utterances.The PV scores are computed directly from distances between disabled and reference speaker projection vectors. Both session level and utterance level PV scenarios are presented and evaluated. The PV performance is compared with respect to an approach based on the lattice posterior probabilities.

Index Terms: confidence measure, speech therapy