A study of applying subspace based pronunciation modeling in verifying pronunciation accuracy

This paper investigates a new approach for detecting phoneme level mispronunciations from utterances obtained from impaired children with neuromuscular disorders. This new pronunciation verification (PV) approach is obtained from the subspace based Gaussian mixture model (SGMM) based pronunciation model, where a set of state level projection vectors is applied for representing phonetic variability. SGMM models are trained from disabled speakers' utterances and PV scores are computed directly from distances between disabled and reference speaker projection vectors. An experimental study was performed to evaluate the performance of the SGMM based approach with respect to an approach based on the lattice posterior probabilities. A reduction in equal error rate (EER) of approximately 15% was obtained when the SGMM based scores were combined with lattice posterior probabilities.