An End-to-End Chinese Speech Recognition Model for Children Based on Transformer Neural Networks

Nowadays, most of the speech recognition research focuses on adults, and there are fewer studies on child Chinese speech recognition technology, and the special characteristics of child Chinese pronunciation and language expression also bring great challenges to the research of child Chinese speech recognition. Therefore, in this paper, from the perspective of deep learning-based acoustic modeling of Chinese children, we optimize the existing acoustic model and further study the new acoustic modeling method for children to sharpen the accuracy and decoding proficiency of Chinese children's speech recognition. We present an end-to-end child Chinese speech recognition model based on the Transformer-CTC-M. The Transformer - CTC - M captures International and Domestic context, giving it the ability to learn more advanced features from small amounts of data to strengthen model performance under capacity-limited conditions. The Transformer - CTC - M model can be fully utilized by multiple related tasks to leverage the data, refining the model's capacity for generalization under low-resource conditions. The findings indicate that the proposed model surpasses the mainstream models in recognition performance, and the word error rate of the proposed model reaches 21.28% in the self-constructed database, which indicates a remarkable improvement in comparison to all benchmark models.