An investigation of adaptation techniques for building acoustic models for hearing-impaired children in a CAPT application
Yingke Zhu, Brian Kan-Wing Mak · 2016
This paper describes our efforts in building an Android-based computer-assisted pronunciation training application for the local hearing-impaired (HI) children whose mother tongue is Cantonese. Since Cantonese HI children represent only a minority population in the world, the greatest challenge to the undertaking is the lack of their speech data and the difficulty in collecting sufficient speech data from them for acoustic modeling. We took the approach of building HI children acoustic model from normal-hearing (NH) adults model by adaptation using limited amount of adaptation data. Various feature-based and model-based adaptation methods were investigated. They include linear input networks (LIN) and its variants, Kullback-Leibler divergence (KLD) regularization, and learning hidden unit contributions (LHUC). We report results on phoneme recognition error rate (PER) as well as initial consonant recognition error rate (ICER) because the application currently focuses on the articulation of the initial consonants. The best results show that a combination of KLD and LIN-Nblock may reduce PER and ICER by a relative 11% and 16% respectively.