Efficient detection of specific language impairment in children using ResNet classifier

Abstract

Specific language impairment (SLI), also known as developmental dysphasia, is a developmental language disorder that affects approximately 7% of the preschool population. It has long-term impact on children’s academic performance and educational progress. Moreover, SLI negatively affects children’s communication skills and social interactions. It is therefore essential to diagnose the condition at the early stage and conduct appropriate speech therapy. However, SLI often remains undetected due to the great variability of the symptoms and lack of optimal diagnostic indicators. In this paper, we propose an efficient approach to automatic SLI detection based on log-power spectra of speech samples. The LANNA children speech corpus containing utterances of healthy controls and SLI-diagnosed children was used during the algorithm development. The utterances were used to calculate the normalized log-power spectrograms. Deep neural network algorithm based on ResNet architecture was used to perform the classification task. The accuracy rate of proposed SLI detection method exceeds 99% in the speakerindependent scenario. Presented algorithm outperforms most of the state-of-the-art algorithms in terms of accuracy rate and requires low computational power due to its simplicity.