Pronunciation Modelling for Children’s Speech

Authors: Christopher Gebauer, Lars Rumberg, Jörn Ostermann

Abstract:

The accuracy of automatic speech processing systems for children’s speech lags heavily behind the accuracy of systems for adult speech. One of the reasons is a high pronunciation variability in children’s speech. Modeling this variability can be effective to increase performance. We investigate whether MAUS, a system developed for phonemic segmentation and trained on adult speech, which explicitly models deviations from canonical pronunciations, can be applied to children’s speech. We compare it to a recently presented system trained on children’s speech. We evaluate whether the systems can capture pronunciation variability as well as the performance on phonemic segmentation.