Mismatched training data enhancement for automatic recognition of children's speech using DNN-HMM

The increasing profusion of commercial automatic speech recognition technology applications has been driven by big-data techniques, using high quality labelled speech datasets. Children's speech has greater time and frequency domain variability than typical adult speech, lacks good large scale training data, and presents difficulties relating to capture quality. Each of these factors reduces the performance of systems that automatically recognise children's speech. In this paper, children's speech recognition is investigated using a hybrid acoustic modelling approach based on deep neural networks and Gaussian mixture models with hidden Markov model back ends. We explore the incorporation of mismatched training data to achieve a better acoustic model and improve performance in the face of limited training data, as well as training data augmentation using noise. We also explore two arrangements for vocal tract length normalisation and a gender-based data selection technique suitable for training a children's speech recogniser.