Comparison of Read and Spontaneous Speech in Children using Deep Learning

The speech development in children is a complex process. Mental comfort also impacts the way a child communicates. Generally, the confidence level in the speech of a child may vary based on the situation. To analyze the impact of the state of mind in the way child communicates, two types of speech data are considered: read and spontaneous speech data. This study focuses on classifying these two speech types in children using deep learning models like Convolution Neural Network (CNN), VGG-16, and Inception-V3 to assess accuracy for 8-year- old boy and girl. The model analysis mel-spectrograms from the Non-Native Children English Speech (NNCES) Corpus as its input data. These spectrograms visually depict speech signals, highlighting important frequency and temporal characteristics needed to differentiate between read and spontaneous speech in children. For 8-year-old children, the CNN achieved 93% accuracy for boys and 74% for girls, while VGG-16 achieved 92% and 71% for boys and girls, respectively, while Inception-V3 achieved 80% accuracy for boys and 70% for girls illustrating the significant difference between the read and spontaneous speech data.