Research on Speech Adaptation of Neural Network Acoustic Model in Power Grid Call System

To further improve the speech recognition capability in power grid call systems, this paper proposes an sentiment analysis model KBERT-TextCNN-B to enhance the text sentiment classification model. Firstly, crawl text data related to power services from platforms such as Weibo and Baidu Tieba and establish an sentiment dictionary. Use the BERT model for word vector processing and take the output of the last K Transformer layers as input to the Text CNN model, in order to improve the problem of BERT easily ignoring local information. Simultaneously calculating the emotional values of each statement, Bayesian theory is used to correct the classification probability to improve model performance. Finally, CNN, Naive Bayes, and BERT TextCNN models were selected for experimental comparison on five indicators. The experimental results showed that this model can accurately analyze the emotional information of users in the power service industry and improve its accuracy in providing user services.