Deep neural network adaptation for children's and adults' speech recognition

Abstract

English. This paper introduces a novel application of the hybrid deep neural network (DNN)-hidden Markov model (HMM) approach for automatic speech recognition (ASR) to target groups of speakers of a specific age/gender. The group-specific training of DNN is investigated and shown to be inefficient when the amount of training data is limited. To overcome this problem, the recent approach that consists in adapting a general DNN to domain/language specific data is extended to target age/gender groups in the context of hybrid DNN-HMM systems, reducing consistently the phone error rate by 15-20% relative for the three different speaker groups.