End-to-end acoustic modelling for phone recognition of young readers

Automatic recognition systems for child speech are lagging behind those dedicated to adult in the race of performance. This phenomenon is due high acoustic and linguistic variability present caused by their body development, as well lack available data. Young readers’ additionally displays peculiarities, such slow reading rate presence mistakes, that hardens task. work attempts tackle main challenges phone modelling young with limited data improve understanding strengths weaknesses a wide selection model architectures this domain. We find transfer learning techniques highly efficient on end-to-end adult-to-child adaptation small amount Through learning, Transformer complemented Connectionist Temporal Classification (CTC) objective function, reaches error 28.1%, outperforming state-of-the-art DNN–HMM 6.6% relative, other more than 8.5% relative. An analysis models’ performance two specific tasks (isolated words sentences) provided, showing influence utterance length attention-based CTC-based models. The Transformer+CTC an ability better detect mistakes made children, which can be attributed CTC function effectively constraining attention mechanisms monotonic.