Age-Invariant Training for End-to-End Child Speech Recognition Using Adversarial Multi-Task Learning

Automatic speech recognition for children’s speech is a challenging task mainly due to scarcity of publicly available child speech corpora and wide inter- and intra-speaker variability in terms of acoustic and linguistic characteristics of children’s speech. We propose a framework for age-invariant training of the acoustic model of end-to-end speech recognition systems based on adversarial multi-task learning. We use age information additionally to just differentiating between the child and adult domains and thus force the acoustic model to learn age invariant features. Our results on publicly available data sets show that this leads to better leveraging of existing data during training. We further show that usage of adversarial multitask learning should not necessarily be regarded as a substitute for traditional feature space adaptation methods, but that both should be used together for best performance.