Improved Vocal Tract Length Perturbation for Improving Child Speech Emotion Recognition

Due to the variability of children's acoustic features during growth and the lack of a publicly available corpus of children's speech emotion data, training a robust children's Speech Emotion Recognition(SER) is more challenging than that of adults. We develop in this paper a new SER learning solution for children, which generates a child-like speech spectrum from an adult's speech spectrum based on vocal tract length normalization (VTLN) and vocal tract length perturbation (VTLP). This improved adult spectrum is then used as augmented data to improve the child speech emotion recognition system. To validate the effectiveness of this data augmentation method, we test it in two mainstream SER task networks based on Compact Convolutional Transformers(CCT) and Convolutional Neural Networks(CNN), respectively, and the proposed data augmentation method improves the accuracy and UAR by 0.32% and 0.22%, respectively, on the Fau Aibo children's emotion data corpus compared to the VTLP baseline system when trained with CCT. When trained with CNN, the accuracy and UAR are improved by 0.39% and 0.26%, respectively, compared to the VTLP baseline system.