Verifying Session Level Pronunciation Accuracy in a Speech Therapy Application

This paper investigates the problem of verifying the pronunciations of phonemes from continuous utterances collected from impaired children speakers engaged in a speech therapy session. A new pronunciation verification (PV) approach based on the subspace Gaussian mixture model (SGMM) is presented. A single SGMM is trained from test utterances collected from impaired and unimpaired speakers. PV measures are derived from model space based distances estimated from impaired and unimpaired speaker dependent SGMM parameters. The PV performance is presented in terms of the method’s ability to detect and report phone level pronunciation errors for an entire speech therapy session and for individual phoneme occurrences. Comparisons are made with respect to an approach using phone level confidence measures estimated from lattice posterior probabilities.