To solve the poor quality of Very high frequency (VHF) speech communication in the navigation field, a VHF speech enhancement model based on an improved transformer (VHFSE) is proposed in this paper. The long-term and short-term noise are the reasons for the poor quality of VHF voice communication. VHFSE can reduce these two aspects of noise. We select the Two-stage Transformer based Neural Network (TSTNN) as the baseline. The Transformer structure pays attention to global information and parallel computing, which can reduce the long-term noise. In order to strengthen the ability of the model to reduce short-term noise, we add CNN module to the transformer according to the ability of revolutionary neural networks (CNN) to extract local information. Meanwhile, to improve the real-time performance , this study employs the lightweight convolution module (Depthwise Separable Convolution) to efficiency of VHF speech communication. Experimental results show that the proposed model VHFSE obtains the highest PESQ and STOI values than other compared modules. Besides, we apply the self-built dataset in our proposed model. The spectrum diagram shows that our model has the best enhancement effect on navigation VHF speech.
Vessel Traffic Service (VTS) significantly improves the navigation efficiency of ports. This paper proposes a model called Joint Extraction of Triples from the VHF Speech (JER‐VHF) to ensure the efficiency of the VTS. Numerous texts are extracted from the Very High Frequency (VHF) speech communication contents and these texts are organized into a dataset named VHFDT. The proposed model's transforming task transforms the voice communication contents of this dataset into a triple representation. VHFDT has a large number of overlapping triples. Therefore, this paper proposes a combined model with three categories to model the entity relations in VHF sentences, including pre‐training Chinese language model for initializing embedding from VHFDT, BiLSTM for rich features, and Multi‐head Attention for focusing on triples. In experimental part, this study uses Precision(P), Recall(R), and F1 to evaluate the accuracy and effectiveness of the proposed method and baseline models. According to experimental results, the proposed model efficiently extracts the key information from complex language environment and achieves better work on relational triple extraction than other baseline models. The model achieved an F1‐score of 83.2% on the overlapping testing data, which is an improvement of 1.8% compared to the second‐best baseline model.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.