RelNet: Relevance-Based Punctuation Prediction Network in Speech
Haifeng Zhao, Xing Chen, Xin Wang, Shilei Huang · 2021 17th International Conference on Computational Intelligence and Security (CIS) · 2021
In the automatic speech recognition system, the text content of the voice transcription is usually different in different application scenarios. Punctuation prediction networks cannot adapt well to many application scenarios at the same time. When the testing set comes from different scenarios, the accuracy of the punctuation prediction system will drop sharply. Therefore, to alleviate the above problem, this paper proposes a relevance calculation module. The experimental results show that adding a relevance calculation module can indeed enhance the generalization of the model. The relevance calculation module enables the model to learn more general and effective characteristics for punctuation prediction to alleviate the problem of model accuracy deterioration. Experiments are conducted on Tsinghua news data and Sohu News data. The final model can get up to 3.59% absolute improvement of the F1 value than the baseline on the testing set.